IMAGE QUALITY STATISTICS AND THEIR USE IN STEGANALYSIS AND

COMPRESSION

by
smail Avcba
B.S. in E.E., Uluda University, 1992
M.S. in E.E., Uluda University, 1994

Submitted to the Institute for Graduate Studies in

Science and Engineering in partial fulfillment of
the requirements for the degree of
Doctor
of
Philosophy

Boazii University
2001

ii

IMAGE QUALITY STATISTICS AND THEIR USE IN STEGANALYSIS AND

COMPRESSION

APPROVED BY:

Prof. Blent Sankur.

(Thesis Supervisor)
Assoc. Prof. Lale Akarun....................
Prof. Emin Anarim..
Assoc. Prof. Nasir Memon..
Assoc. Prof. Ycel Yemez...

DATE OF APPROVAL.

iii

TABLE OF CONTENTS

ACKNOWLEDGMENTS..v
ABSTRACT.vii
ZET......viii
LIST OF FIGURES...ix
LIST OF TABLES.x
LIST OF SYMBOLS / ABBREVATIONS.xii
1. INTRODUCTION..1
1.1. Motivation..1
1.2. Approaches.3
1.3. Contributions..5
1.4. Outline6
2. STATISTICAL EVALUATION OF IMAGE QUALITY MEASURES.8
2.1. Introduction...8
2.2. Image Quality Measures..11
2.2.1. Measures Based on Pixel Differences..12
2.2.1.1. Minkowsky Metrics...12
2.2.1.2. MSE in Lab Space.14
2.2.1.3. Difference over a Neighborhood...15
2.2.1.4 Multiresolution Distance Measure.16
2.2.2. Correlation-Based Measures17
2.2.2.1. Image Correlation Measures...17
2.2.2.2. Moments of the Angles...18
2.2.3. Edge Quality Measures.19
2.2.3.1. Pratt Measure..20
2.2.3.2. Edge Stability Measure...21
2.2.4. Spectral Distance Measures..22
2.2.5. Context Measures.25
2.2.5.1. Rate-Distortion Based Distortion Measure.26
2.2.5.2. f-divergences...27
2.2.5.3. Local Histogram Distances.28

iv

2.2.6. Human Visual System Based Measures...29

2.2.6.1. HVS Modified Spectral Distortion.29
2.2.6.2. A Distance Metric for Database Browsing.30
2.3. Goals and Methods......31
2.3.1 Quality Attributes...31
2.3.2. Test Image Sets and Rates....32
2.3.3. Analysis of Variance33
2.3.4. Visualization of Quality Metrics..36
2.4. Statistical Analysis of Image Quality Measures......37
2.4.1. ANOVA Results...38
2.4.2. Self Organizing Map of Quality Measures...42
2.5. Conclusions.44
3. STEGANALYSIS USING IMAGE QUALITY METRICS...46
3.1. Introduction.46
3.2. Choice of Image Quality Measures.49
3.3. Regression Analysis of the Quality Measures56
3.4. Simulation Results..57
3.5. Conclusions.62
4. LOSSLESS AND NEAR-LOSSLESS IMAGE COMPRESSION
WITH SUCCESSIVE REFINEMENT...64
4.1. Introduction.64
4.2. Problem Formulation...66
4.2.1. Successive Refinement.67
4.2.2. P.m.f. Estimation in the First Pass....68
4.2.3. P.m.f. Estimations in the Second Pass..70
4.2.4. Multi-hypothesis Testing..74
4.3. The Compression Method...75
4.4. Experimental Results...78
4.5. Conclusion...80
5. CONCLUSIONS AND FUTURE PERSPECTIVES..81
REFERENCES.....84
REFERENCES NOT CITED...95

ACKNOWLEDGMENTS

I would like to cordially thank my supervisor, Professor Blent Sankur, for

welcoming me in his research group, for his continuous help and guidance, for so many
fruitful discussions and sharing his ideas. I am still amazed by the diversity and the
spectrum of his scientific working knowledge. I cannot emphasize enough how much I
benefited from his knowledge, wisdom and personality. It was an honor to know and work
with him.
I wish to express my sincere thanks to Professor Nasir Memon for several good
reasons. He has been a good friend in so many ways. He has traveled very long distances
and participated in my thesis jury. He welcomed me to VIP Lab at Polytechnic University
and introduced me to the challenges in visual communications. His supervision, support
and confidence in me, is truly appreciated.
I would also like to thank Professors Emin Anarim, Lale Akarun, Ycel Yemez for
serving on my thesis committee.
I express my gratitude to graduate studies supervisor Professor Yorgo stefanopulos
and Professor Emin Anarm for their help on a very critical moment in my PhD study.
Grants from TUBTAK BDP Program and NSF (9724100) are highly
acknowledged, as part of this research was made possible.
I was very fortunate in having good teachers. Professor Khalid Sayood introduced
me to data compression. I have benefited too much from the information theoretic lectures
tutored by Professor Hakan Deli. I should mention Professor Erdal Panayrc for the
course on pattern recognition.
I would like to express my thanks to the Electronics Engineering Department at
Uluda University for allowing me to do my PhD at Boazii University. Special thanks
goes to the chairman Professor Ali Oktay for his support and the very first encouragement;

vi

and to a former Professor at Uluda University, Haluk Gmkaya, as he helped me in

developing first interests in signal processing.
Thanks are also due: To Professors Haldun Hadimiolu, Onur G. Gleryz and Dr.
Kris Popat. To colleagues and friends at Electronics Engineering Department of Uluda
University. To past and present crews of the labs, BUSIM at Boazii and VIP at Poly.
To Tahsin Torun, brahim Dede, Mahmut ifti, Sekin Ar, Hasan H. Otu, Ali Oztrk,
Veysi ztrk, aban Serin, for being there at the right time when I needed them most.
It is the biggest pleasure to express my gratitude to my beloved parents and sister, as
it is their never-ending loving support and encouragement leading me here.
Last, but certainly not least, I wish to convey my insurmountable gratitude to my
wife, for her sincere love, for patience and encouragement during my doctoral studies.

vii

ABSTRACT

IMAGE QUALITY STATISTICS AND THEIR USE IN

STEGANALYSIS AND COMPRESSION

We categorize comprehensively image quality measures, extend measures defined

for gray scale images to their multispectral case, and propose novel image quality
measures. The statistical behavior of the measures and their sensitivity to various kinds of
distortions, data hiding and coding artifacts are investigated via Analysis of Variance
techniques.

Their similarities or differences have been illustrated by plotting their

Kohonen maps. Measures that give consistent scores across an image class and that are
sensitive to distortions and coding artifacts are pointed out.
We present techniques for steganalysis of images that have been potentially
subjected to watermarking or steganographic algorithms.

Our hypothesis is that

watermarking and steganographic schemes leave statistical evidence that can be exploited
for detection with the aid of image quality features and multivariate regression analysis.
The steganalyzer is built using multivariate regression on the selected quality metrics. In
the absence of the ground-truth, a common reference image is obtained based on blurring.
Simulation results with the chosen feature set and well-known watermarking and
steganographic techniques indicate that our approach is able to reasonably accurately
distinguish between marked and unmarked images.
We also present a technique that provides progressive transmission and near-lossless
compression in one single framework. The proposed technique produces a bitstream that
results in progressive reconstruction of the image just like what one can obtain with a
reversible wavelet codec. In addition, the proposed scheme provides near-lossless
reconstruction with respect to a given bound after each layer of the successively refinable
bitstream is decoded. Experimental results for both lossless and near-lossless cases are
presented, which are competitive with the state-of-the-art compression schemes.

viii

ZET

STATSTKSEL MGE KALTESNN STEGO-ANALZDE

KULLANIMI VE KODLAMA

mge nitelik ltlerini detayl bir ekilde snflandrdk, gri seviye imgeler iin
tanmlanm olan imge nitelik ltlerini ok bantl imgelere genilettik. Bu ltleri
deiik bozulumlar, imge kodlama ve damgalama uygulamalarnda ortaya kan
bozulumlar iin istatistiksel olarak karlatrdk.

Bu ltler arasndaki benzerlik ve

farkllklar Kohonen haritalar kullanlarak grselletirildi.

Deiinti analizlerine

dayanarak imge kodlama, damgalama ve dier bozulumlara tutarl ve duyarl tepki veren
ltler saptand.
Damgalanm imgelerde damga varlnn sezimi ile ilgili yntemler sunulmaktadr.
Belirli bir damgalama ynteminin imgede istatistiksel ve yapsal izler brakaca
varsaymndan yola karak, bu izlerin uygun zniteliklerin seimi ve oklu balanm
analizi ile damga varlnn seziminde kullanlabilecei gsterilmitir.

yi bilinen

damgalama yntemleri ve zengin bir imge kmesi zerinde elde edilen grece yksek
doru sezim yzdeleri, nerilen yntemlerin baarl olduunu kantlamaktadr.
Yeni bir yitimsiz kodlama yntemi ile, yitimlerin istenen snrlar iinde kalmasnn
saland, dolaysyla snrl yitimli olarak adlandrlan bir kodlama yntemi gelitirildi.
Yntemimizin bir dier nemli zellii, herhangi bir

geiten sonra elde edilen bit

dizgisinden geri atlan imgede pikseller zerindeki en byk hatann snrlandrlm

olmasdr.

Bu zellii aamal kodlayclar snfnda n plana kan dalgack tabanl

kodlayclar salayamamaktadrlar.

Yitimsiz ve yitimsize yakn kodlamaya ilikin

deneysel sonular, nerilen yntemin baarmnn bilinen en iyi kodlayclar kadar iyi hatta
baz imgeler iin daha da iyi olduunu gstermektedir.

ix

LIST OF FIGURES

Figure 2.1. Box plots of quality measure scores. A) good measure, b) moderate
measure, c) poor measure. The F-scores as well as the significance
level p are given....35
Figure 2.2. SOM of distortion measures for JPEG and SPIHT...43
Figure 3.1. Schematic descriptions of (a) watermarking or stegoing, (b) filtering
an un-marked image, (c) filtering a marked image..52
Figure 3.2. Scatter plots of the three image quality measures (M3: czenakowski
measure, M5: image fidelity, M6: normalized cross-correlation)....53
Figure 3.3. Schematic description of (a) training, and (b) testing.58
Figure 4.1. Ordering of the causal prediction neighbors of the current pixel i , N=6...70
Figure 4.2. The context pixels, denoted by and ! , used in the covariance
estimation of the current pixel . The number of context pixels is K=40..70
Figure 4.3. Causal, ! , and non-causal, , neighbors of the current pixel, ,
used for probability mass estimation in the second and higher passes.....71
Figure 4.4. Schematic description of the overall compression scheme...75
Figure 4.5. Details of the encoder block used in figure 4.4. Here is the length
of the interval (Lm , Rm ] .76
Figure 4.6. The decoder is a replica of the encoder....77

LIST OF TABLES

Table 2.1. List of symbols and equation numbers of the quality metrics11
Table 2.2. ANOVA results (f-scores) for the JPEG and SPIHT compression
distortions as well as additive noise and blur artifacts. For each
distortion type the variation due to image set is also established...39
Table 2.3. One-way ANOVA results for each image class and two-way ANOVA
results for the distortions on the combined and image set independence..40
Table 2.4. ANOVA scores for the bit rate variability (combined JPEG and SPIHT
scores) and coder variation41
Table 3.1. One-way ANOVA tests for watermarking, steganography and pooled
watermarking and steganography...55
Table 3.2. Training and test samples for DIGIMARC and PGS for experiment 1.59
Table 3.3. Training and test samples for COX for experiment 159
Table 3.4. Training and test samples for pooled watermarking algorithms for
experiment 2 (L1: level 1 etc.)...60
Table 3.5. Training and test samples for experiment 3: train on DIGIMARC,
test on PGS and COX....60
Table 3.6. Training and test samples for Stools for experiment 4...63
Table 3.7. Training and test samples for Jsteg for experiment 4.60
Table 3.8. Training and test samples for Steganos for experiment 4..60

xi

Table 3.9. Training and test samples for pooled stego algorithms for experiment 5..61
Table 3.10. Training and test samples for experiment 6: train on
Steganos and Stools, test on Jsteg61
Table 3.11. Training and test samples for pooled watermarking and
steganography algorithms for experiment 761
Table 3.12. Training and test samples for experiment 8: train on Steganos,
Stools and DIGIMARC, test on Jsteg and COX......61
Table 3.13. Performance of the steganalyzer for all the experiments.62
Table 4.1. Comparison of lossless compression results: proposed method
versus CALIC.

.78

Table 4.2. Comparison of 4 different methods of near-lossless compression ( = 1 )...78

Table 4.3. Comparison of 4 different methods of near-lossless compression ( = 3 )..79
Table 4.4. Comparison of 4 different methods of near-lossless compression ( = 7 )..79
Table 4.5. Comparison of bit/pixel efficiency and peak signal to noise ratio
in dB of the proposed algorithm versus the CALIC algorithm.79

xii

LIST OF SYMBOLS / ABBREVIATIONS

Norm of vector a

a, b

Inner product of vectors a and b

Multispectral image

Distorted multispectral image

Ck

kth band of C

C k (i, j )

(i, j)th pixel of the kth band image

C(i, j )

(i, j)th multispectral (with K bands) pixel vector

C kl

lth block of the kth band image

Cm (i, j )

mth scale output of the Derivative of Gaussian operator

C w (i, j )

Pixels in the wxw neighborhood of C (i, j )

DCT

2-D inverse DCT

D( p q )

Relative entropy measure between p and q

d (, )

Distance metric

d rk

Distance in rth multiresolution in the kth band

E(i, j, m )

Edge map at scale m and position (i, j )

Ep

Expectation with respect to p

Cover signal

Stego signal

G m ( x, y )

Derivative of gaussian operator

.
H ()

Entropy funtion

H2

Squared Hellinger norm

H0

Null hypothesis

HA

Alternative hypothesis

H ( )

Band-pass filter in polar coordinates modeling HVS

H (m, n )

Gaussian blur filter

xiii

(Lm , Rm ]

Interval associated with hypothesis m

M (u , v)

Magnitude spectra

pi

p.m.f. of the pixel at position i

p*

optimum p.m.f. w.r.t. a constraint

{p }

Set of neighboring pmfs of current pixel at position i

Q(i, j )

Edge stability map at position (i, j )

R( p )

Average bit rate with respect to p

{ }

Set of real-valued linear prediction coefficients

xm

Indicator function I {x (Lm , Rm ]}

lth largest deviation

Error over all the pixels in the kth bands, C k C k

Variance

(u , v)

Phase spectra

Test set

k (u , v)

DFT of the kth band image

kl (u , v )

DFT of the lth block of the kth band image

Mean

Training set

k (u, v )

DCT of the kth band image

Watermark signal

Set of features in steganalysis of watermarking techniques

Set of features in steganalysis of stego techniques

Set of features in pooled steganalysis

ANOVA

ANalysis Of VAriance

C1

Normalized Cross-Correlation

C2

Image Fidelity

C3

Czekanowski Correlation

i j

j =1

j =1

xiv

C4

Mean Angle Similarity

C5

Mean Angle-Magnitude Similarity

CALIC

Context Adaptive Lossless Image Compression

CIE

International Commission of Illumination

D1

Mean Square Error

D2

Mean Absolute Error

D3

Modified Infinity Norm

D4

L*a*b* Perceptual Error

D5

Neighborhood Error

D6

Multiresolution Error

DCT

Discrete Cosine Transform

DFT

Discrete Cosine Transform

E1

Pratt Edge Measure

E2

Edge Stability Measure

H1

HVS Absolute Norm

H2

HVS L2 Norm

H3

Browsing Similarity

H4

DCTune

HVS

Human Visual System

i.i.d.

independent and identically distributed

IQM

Image Quality Measure

ITU

International Telecommunications Union

JND

Just Noticeable Difference

JPEG

Joint Pictures Experts Group

KS

Kolmogorov-Smirnov

LSB

Least Significant Bit

MAP

Maximum A Posteriori

ML

Maximum Likelihood

MSE

Mean Square Error

p.d.f.

probability distribution function

p.m.f.

probability mass function

PSNR

Peak SNR

RGB

Red Green Blue

xv

S1

Spectral Phase Error

S2

Spectral Phase-Magnitude Error

S3

Block Spectral Magnitude Error

S4

Block Spectral Phase Error

S5

Block Spectral Phase-Magnitude Error

SNR

Signal to Noise Ratio

SOM

Self Organizing Map

SPIHT

Set Partitioning In Hierarchical Trees

SRC

Spearman Rank Correlation

VQEG

Video Quality Experts Group

Z1

Rate Distortion Measure

Z2

Hellinger distance

Z3

Generalized Matusita distance

Z4

Spearman Rank Correlation

1. INTRODUCTION

1.1. Motivation

There has been an explosive growth in multimedia technology and applications in the
past several years. Efficient representation for storage, transmission, retrieval and security
are some of the biggest challenges faced.

The first concern addressed in this thesis is the efficient compression of data. Visual
information is one of the richest but also most bandwidth consuming modes of
communication. However, to meet the requirements of new applications such as mobile
multimedia, interactive databases (encyclopedias, electronic newspaper, travel information,
and so on) powerful data compression techniques are needed to reduce the bit rate, even in
the presence of growing communications channels offering increased bandwidth. Other
applications are in remote sensing, education and entertainment.

Not only reducing the bit rate but functionalities of progressive transmission or
progressive decoding of the bit stream became more important features of compression
schemes. A typical application is data browsing. A user may want to visualize the picture
at a lower quality to save transmission time. Another application is tele-radiology where a
physician can request portions of an image at increased quality (including lossless
reconstruction) while accepting unimportant portions at much lower quality, thereby
reducing the overall bandwidth required for transmitting an image.

Although most applications require high compression ratios, this requirement is in

general in conjunction with the desire for high quality in the resulting content.
Guaranteeing a certain level of quality after compression has become a prime concern for
content providers, as the quality in the resulting content is the most important factor in the
success of an application in the market place.

A second concern in the thesis is the understanding of image quality metrics.

Compression, transmission and sensor inadequacy lead to artifacts and distortions affecting

image quality. Identifying the image quality measures that have highest sensitivity to these
distortions would help systematic design of coding, communication and imaging systems
and of improving or optimizing the picture quality for a desired quality of service.

The third concern is image security and secret communication.

Given the

proliferation of digital images, and given the high degree of redundancy present in a digital
representation of an image (despite compression), there has been an increased interest in
using digital images as cover-objects for the purpose of data hiding. Since unlimited
number of copies of an original can be easily distributed or forged, the protection and
enforcement of intellectual property rights is an another important issue.

A digital

watermark is intended to complement cryptographic processes, and is an imperceptible

signal added to digital content that can be later detected or extracted in order to make some
assertion about the content. Although the main applications for digital watermarking
appear to be copyright protection and digital rights management, watermarks have also
been proposed for secret communication, that is, steganography. However, despite this
obvious and commonly observed connection to steganography, there has been very little
effort aimed at analyzing or evaluating the effectiveness of watermarking techniques for
steganographic applications. Instead, most work has focused on analyzing or evaluating
the watermarking algorithms for their robustness against various kinds of attacks that try to
remove or destroy them.

If digital watermarks are to be used in steganography

applications, detection of their presence by an unauthorized agent defeats their very

purpose.

Even in applications that do not require hidden communication, but only

robustness, we note that it would be desirable to first detect the possible presence of a
watermark before trying to remove or manipulate it.

The steganalysis is especially important, as there is a number application areas

interested in information hiding, such as

Military and intelligence agencies require secret communications.

Terrorists and criminals also place great value on secret communications.

Law enforcement and counter intelligence agencies are interested in

understanding these technologies and their weaknesses, so as to detect and trace
hidden messages.

Schemes for digital elections and digital cash make use of anonymous
communication techniques.

1.2. Approaches

The focus of this thesis is on three challenges of visual communications. One is the
efficient representation of image data for storage and transmission, which is the art and
science of identifying models for different types of structures existing in image data to
obtain compact representations. Second is image quality and the third is multimedia
security.

These are diverse research areas, but also are integral parts of visual

communications as a whole. Findings in one field are readily used in the others. For
example incorporation of a good image model in a compression scheme decreases the bit
rate to represent the image. Still this image model can be used in image data hiding to get
a true idea of the data-hiding capacity. Good image models together with quality metrics
incorporating human visual system are indispensable in the design of both image coding
and watermarking systems as we want visually pleasing, compactly represented and robust
images resistant to various attacks.

Image Quality: There is a wealth of research on subjective and/or objective image

quality measures to reliably predict either perceived quality across different scenes and
distortion types or to predict algorithmic performance computer vision tasks.

Our

approach is different from the companion studies, in that set of objective image quality
measures has been statistically analyzed to identify the ones most sensitive and
discriminative to compression, watermarking, blurring and noise distortions.

The

identified measures are important in the design and performance evaluation of

compression, watermarking and imaging systems.

Given that an imaging system

introduces blurring, by incorporating the most discriminative measure to blurring, among

other constraints, and using it as an objective function in the design process would alleviate
this problem. Similarities, differences and mutual correlations of image quality measures
have been illustrated by plotting Kohonens self organizing maps. In case one measure is
not tractable in the design process, still another mutually correlated and simpler measure
can be used. Also, if a combination of measures is required, the redundant measures can
be easily eliminated.

Watermark and Stego-mark Detection:

Most of the work in watermarking has

focused on analyzing or evaluating the watermarking algorithms for their robustness

against various kinds of attacks that try to remove or destroy them. However, despite the
obvious and commonly observed connection to steganography, there has been very little
effort aimed at analyzing or evaluating the effectiveness of watermarking techniques for
steganographic applications. General detection techniques as applied to steganography
have not been devised and methods beyond visual inspection and specific statistical tests
for individual techniques are not present in the literature.

Our approach addresses exactly an automatic detection of the presence of watermark

or steganographic marks in images.

Hiding information in digital media requires

alterations of the signal properties that introduce some form of degradation, no matter how
small. These degradations can act as signatures that could be used to reveal the existence
of a hidden message. For example, in the context of digital watermarking, the general
underlying idea is to create a watermarked signal that is perceptually identical but
statistically different from the host signal. A decoder uses this statistical difference in
order to detect the watermark. However, the very same statistical difference that is created
can potentially be exploited to determine if a given image is watermarked or not. We show
that addition of a watermark or message leaves unique artifacts, which can be detected
using the most discriminative image quality measures identified from their statistical
analysis and multivariate regression analysis.

We show that selected image quality

measures form a multidimensional feature space whose points cluster well enough to do a
classification of marked and non-marked images.

Lossless Progressive Compression: Lossless or reversible compression refers to

compression approaches in which the reconstructed data exactly matches the original.
Near-lossless compression denotes compression methods, which give quantitative
guarantees on the nature of the loss that is introduced. Near-lossless compression is
potentially useful in remote sensing, medical imaging, space imaging and image archiving
applications, where the huge data size could require lossy compression for efficient storage
or transmission. However, the need to preserve the validity of subsequent image analysis
performed on the data set to derive information of scientific or clinical value puts strict
constraints on the error between compressed image pixel values and their originals. In

such cases, near-lossless compression can be used as it yields significantly higher

compression ratios compared to lossless compression and at the same time, the quantitative
guarantees it provides on the nature of loss introduced by the compression process are
more desirable compared to the uncertainties that are faced when using lossy compression.

Another pillar of this thesis is the proposal of a novel image compression scheme
that provides progressive transmission and near-lossless compression in one single
framework. We formulate the image data compression problem as one of asking the
optimal questions to determine, respectively, the value or the interval of the pixel,
depending on whether one is interested in lossless or near-lossless compression. New
prediction methods based on the nature of the data at a given pass are presented and links
to the existing methods are explored. The trade-off between non-causal prediction and
data precision is discussed within the context of successive refinement. Context selection
for prediction in different passes is addressed. Experimental results for both lossless and
near-lossless cases are presented, which are competitive with the state-of-the-art
compression schemes.

1.3. Contributions

The major contributions of this thesis can be highlighted as follows:

We have presented collectively a set of image quality measures in their

multispectral version and categorized them. Image quality measures have been
statistically analyzed to identify the ones most sensitive and discriminative to
compression, watermarking, blurring and noise distortions. We have also pointed
out the image features that should be taken more seriously into account in the
design of more successful coding, imaging and data hiding systems.

The

correlation between various measures has been depicted via Kohonens SelfOrganizing Map. The placement of measures in the two-dimensional map has
been in agreement with ones intuitive grouping.

By using statistically most significant image quality measures, we have

developed steganalysis techniques both for conventional LSB like embedding

used in the context of a passive warden model and for watermarking which can
be used to embed secret messages in the context of an active warden. The
techniques we present are novel and to the best of our knowledge, the first
attempt at designing general purpose tools for steganalysis.

We proposed a novel technique that unifies progressive transmission and nearlossless compression in one single bit stream. The proposed technique produces
a bitstream that results in progressive reconstruction of the image just like what
one can obtain with a reversible wavelet codec. In addition, the proposed scheme
provides near-lossless reconstruction with respect to a given bound after each
layer of the successively refinable bitstream is decoded.

Furthermore, the

compression performance provided by the proposed technique is superior or

comparable to the best-known lossless and near-lossless techniques proposed in
the literature. The originality of the method consists in looking at the image data
compression as one of asking the optimal questions to determine the interval in
which the current pixel lies.

1.4. Outline

We present a set of image quality measures and analyze them statistically with
respect to coding, blurring and noise distortions in the second section. The measures are
categorized into pixel difference-based, correlation-based, edge-based, spectral-based,
context-based and HVS-based (Human Visual System-based) measures. We conduct a
statistical analysis of the sensitivity and consistency behavior of objective image quality
measures. The mutual relationships between the measures are visualized by plotting their
Kohonen maps. Their consistency and sensitivity to coding as well as additive noise and
blur are investigated via analysis of variance of their scores.

Using the discrimination power concept of image quality measures, we develop

steganalysis techniques for watermarking and steganographic applications in section three.
We first describe the steganography problem in terms of prisoners problem. Next, we
show that the distance between an unmarked image and its filtered version is different from
the distance between a marked image and its filtered version. This is the critical finding on

which we build the steganalyzer by using a small subset of image quality measures and
multivariate regression analysis. Later, we give the design principles and extensively
describe the experiments to test the performance of the steganalyzer with a variety of best
known watermarking and mostly cited steganographic algorithms on a rich image set.

Key problem in lossless compression is accurate probability mass function

estimation. We describe probability mass function estimation methods, data models, and
our novel approach to the design of an embedded lossless and near-lossless image
compression scheme in section four. This section is compact and the knowledge it requires
from sections two and three is at minimal.

Section five concludes the thesis and explores directions for future work.

2. STATISTICAL EVALUATION OF IMAGE QUALITY MEASURES

2.1. Introduction

Image quality measures are figures of merit used for the evaluation of imaging
systems or of coding/processing techniques. We consider several image quality metrics and
study their statistical behavior when measuring various compression and/or sensor
artifacts.

A good objective quality measure should well reflect the distortion on the image due
to, for example, blurring, noise, compression, sensor inadequacy. One expects that such
measures could be instrumental in predicting the performance of vision-based algorithms
such as feature extraction, image-based measurements, detection, tracking, segmentation
etc. tasks. Our approach is different from companion studies in the literature focused on
subjective image quality criteria, such as in [1, 2, 3]. In the subjective assessment of
measures characteristics of the human perception becomes paramount, and image quality is
correlated with the preference of an observer or the performance of an operator on some
specific task.

In the image coding and computer vision literature, the raw error measures based on
deviations between the original and the coded images are overwhelmingly used [4, 5, 6],
Mean Square Error (MSE) or alternatively Signal to Noise Ratio (SNR) varieties being the
most common measures. The reason for their widespread choice is their mathematical
tractability and it is often straightforward to design systems that minimize the MSE. Raw
error measures such as MSE may quantify the error in mathematical terms, and they are at
their best with additive noise contamination, but they do not necessarily correspond to all
aspects of the observers visual perception of the errors [7, 8], nor do they correctly reflect
structural coding artifacts.

For multimedia applications and very low bit rate coding, quality measures based on
human perception are being more frequently used [9, 10, 11, 12, 13, 14]. Since a human
observer is the end user in multimedia applications, an image quality measure that is based

on a human vision model seems to be more appropriate for predicting user acceptance and
for system optimization. This class of distortion measures gives in general a numerical
value that will quantify the dissatisfaction of the viewer in observing the reproduced image
in place of the original (though Dalys VPD map [13] is a counter example to this). The
alternative is subjective tests where subjects view a series of reproduced images and rate
them based on the visibility of artifacts [15, 16]. Subjective tests are tedious and time
consuming and the results depend on various factors such as observers background,
motivation, etc., and furthermore actually only the display quality is being assessed.
Therefore an objective measure that accurately predicts the subjective rating would be a
useful guide when optimizing image compression algorithms.

Recently there have been ITU (International Telecommunications Union) efforts to

establish objective measurement of video quality. Thus in the context of distribution of
multimedia documents, video programming in particular, in-service continuous evaluation
of video quality is needed. This continuous video quality indicator would be an input to
the network management, which must guarantee a negotiated level of service quality.
Obviously such a quality monitoring can only be realized with objective methods [17, 18].
It must be pointed out, however, that subjective assessment, albeit costly and timeconsuming, if not impractical, is accurate. Objective methods, on the other hand, can at
best try to emulate the performance of subjective methods, utilizing the knowledge of the
human visual system.

Similarly for computer vision tasks, prediction of algorithmic performance in terms

of imaging distortions is of great significance [19, 20]. In the literature the performance of
feature extraction algorithms, like lines and corners [19], propagation of covariance
matrices [20], quantification of target detection performance and ideal observer
performance [21, 22, 23] have been studied under additive noise conditions. It is of great
interest to correlate coding and sensor artifacts with such algorithmic performance. More
specifically one would like to identify image quality metrics that can predict accurately and
consistently the performance of computer vision algorithms in distorted image records, the
distortions being due to compression, sensor inadequacy etc.. An alternative use of image
quality metrics is in the inverse mapping from metrics to the nature of distortions [24]. In
other words given the image quality metrics, one tries to reconstruct the distortions (e.g.,

10

blur, noise, etc. amount in a distortion space) that could have resulted in the measured
metric values.

Thus in this study we aim to study objective measures of image quality and to
investigate their statistical performance. Their statistical behavior is evaluated first, in
terms of how discriminating they are to distortion artifacts when tested on a variety of
images using Analysis of Variance (ANOVA) method.

Second, the measures are

investigated in terms of their mutual correlation or similarity, this being put into evidence
by means of Kohonen maps.

Some 26 image quality metrics are described and summarized in Table 2.1. These
quality metrics are categorized into six groups according to the type of information they
are using. The categories used are:

Pixel difference-based measures such as mean square distortion;

Correlation-based measures, that is, correlation of pixels, or of the vector angular

directions;

Edge-based measures, that is, displacement of edge positions or their consistency

across resolution levels;

Spectral distance-based measures, that is Fourier magnitude and/or phase spectral

discrepancy on a block basis;

Context-based measures, that is penalties based on various functionals of the

multidimensional context probability;

Human Visual System-based measures, measures either based on the HVSweighted spectral distortion measures or (dis)similarity criteria used in image
database browsing functions.

We define several distortion measures in each category. The specific measures are
named as D1, D2.. in the pixel difference category, as C1, C2 .. in the correlation category
etc. for ease of reference in the results and discussion sections.

11

Table 2.1. List of Symbols and Equation Numbers of the Quality Metrics
SYMBOL
D1
D2
D3
D4
D5
D6
C1
C2
C3
C4
C5
E1
E2
S1
S2
S3
S4
S5
Z1
Z2
Z3
Z4
H1
H2
H3
H4

DESCRIPTION
Mean Square Error
Mean Absolute Error
Modified Infinity Norm
L*a*b* Perceptual Error
Neighborhood Error
Multiresolution Error
Normalized Cross-Correlation
Image Fidelity
Czenakowski Correlation
Mean Angle Similarity
Mean Angle-Magnitude Similarity
Pratt Edge Measure
Edge Stability Measure
Spectral Phase Error
Spectral Phase-Magnitude Error
Block Spectral Magnitude Error
Block Spectral Phase Error
Block Spectral Phase-Magnitude Error
Rate Distortion Measure
Hellinger distance
Generalized Matusita distance
Spearman Rank Correlation
HVS Absolute Norm
HVS L2 Norm
Browsing Similarity
DCTune

EQUATION
2.3
2.1
2.3
2.4
2.5
2.6
2.7
2.8
2.9
2.10
2.11
2.12
2.13
2.14
2.15
2.16
2.17
2.18
2.19
2.20
2.21
2.22
2.23
2.24
2.26

2.2. Image Quality Measures

We define and describe the multitude of image quality measures considered. In these
definitions the pixel lattices of images A, B will be referred to as A(i, j ) and B(i, j ) , i, j =
1...N, as the lattices are assumed to have dimensions NxN. The pixels can take values
from the set {0,..., G} in any spectral band. The actual color images we considered had G =
255 in each band. Similarly we will denote the multispectral components of an image at the
pixel position i, j, and in band k as Ck (i, j ) , where k = 1,..., K . The boldface symbols
(i, j ) will indicate the multispectral pixel vectors at position (i,j). For
C(i, j ) and C
example for the color images in the RGB representation one has C(i,j) = [R(i,j) G(i,j)
B(i,j)]T. These definitions are summarized as follows:

12

C k (i, j )

(i, j)th pixel of the kth band of image C

C(i, j )

(i, j)th multispectral (with K bands) pixel vector

A multispectral image

Ck

The kth band of a multispectral image C

k = C k C k

Error over all the pixels in the kth band of a multispectral image C.

Thus for example the power in the k'th band can be calculated as k2 =

N 1

C (i, j )

i , j =0

. All

etc., will correspond to the

these quantities with an additional hat, i.e., C k (i, j ) , C
distorted versions of the same original image.
2

(i, j ) =
C(i, j ) C

[C (i, j ) C (i, j )]
K

k =1

As a case in point, the expression

will denote the sum of errors in the spectral

components at a given pixel position i, j. Similarly the error expression in the last row of
N

2
the above table expands as k2 = C k (i, j ) C k (i, j ) . In the specific case of RGB
i =1 j =1

color images we will occasionally revert to the notations {R, G , B} and R , G , B .

2.2.1. Measures Based on Pixel Differences

These measures calculate the distortion between two images on the basis of their
pixelwise differences or certain moments of the difference (error) image.

2.2.1.1. Minkowsky Metrics. The

norm of the dissimilarity of two images can be

calculated by taking the Minkowsky average of the pixel differences spatially and then
chromatically (that is over the bands):

1
=
K

1
2

k =1 N
K

N 1

C (i, j ) C (i, j )

i , j =0

1/

(2.1)

Alternately the Minkowsky average can be first carried over the bands and then
spatially, as in the following expression:

13

1
= 2
N

N 1 1

i , j =0 K

C k (i, j ) C k (i, j )

k =1

1/

(2.2)

In what follows we have used the pixel-wise difference in the Minkowsky sum as
given in Eq. (2.1). For = 2, one obtains the well-known Mean Square Error (MSE)
expression, denoted as D1:

D1 =

1 1
K N2

N 1

(i, j )
C(i, j ) C

i , j =0

1
K

k =1

2
k

(2.3)

An overwhelming number of quality results in the literature is in fact given in terms

of the Signal to Noise Ratio (SNR) or the Peak SNR (PSNR), which are obtained,
respectively, by dividing the image power by D1, and by dividing the peak power G2 by
D1. Though the SNR and the PSNR are very frequently used in quantifying coding
distortions, their shortcomings have been pointed out in various studies [13]. However,
despite these oft-cited criticisms of the MSE-based quality measures there has been a
recent resurgence of the SNR/PSNR metrics [17, 18]. For example studies of the Video
Quality Expert Group (VQEG) [17] have shown that the PSNR measure is a very good
indicator of subjective preference in video coding.
For = 1 one obtains the absolute difference denoted as D2. For = power in
the Minkowski average the maximum difference measure

1
(i, j) ||
Ck (i, j ) C k (i, j) = max|| C(i, j) C
,
i
j
k =1 K

= max
i, j

(2.4)

is obtained. Recall that in signal and image processing the maximum difference or the
infinity norm is very commonly used [6]. However given the noise-prone nature of the
maximum difference, this metric can be made more robust by considering the ranked list of

)
) denotes the l
Here (C C

, l = 1 ! N 2 , resulting in a modified Minkowsky infinity

pixel differences l C C
metric, called D3.

th

largest deviation among all pixels [25].

14

is simply the error expression above. Similarly correspond to the

Thus 1 C C
2
second largest term etc. Finally a modified maximum difference measure using the first r
of m terms can be constructed by computing the root mean square value of the ranked
largest differences, 1 ! r .

D3 =

1 r 2
m C C
r m =1

(2.5)

2.2.1.2. MSE in Lab Space. The choice of the color-space for an image similarity metric
is important, because the color-space must be uniform, so that the intensity difference
between two colors must be consistent with the color difference estimated by a human
observer. Since the RGB color-space is not well-suited for this task two color spaces are
defined: 1976 CIE L*u*v* and the 1976 CIE L*a*b* color spaces [26]. One recommended
color-difference equation for the Lab color-space is given by the Euclidean distance [27].
Let
L* (i, j ) = L* (i, j ) L* (i, j )

(2.6)

a* (i, j ) = a* (i, j ) a * (i, j )

(2.7)

b* (i, j ) = b* (i, j ) b* (i, j )

(2.8)

denote the color component differences in L*a*b* space. Then the Euclidean distance is:

D4 =

1
N2

[L (i, j )
N 1

i , j =0

+ a * (i, j ) + b * (i, j ) .
2

(2.9)

Note that (2.37) is intended to yield perceptually uniform spacing of colors that
exhibit color differences greater than JND threshold but smaller than those in Munsell
book of color [27]. This measure applies obviously to color images only and cannot be
generalized to arbitrary multispectral images. Therefore it has been used only for the face
images and texture images, and not in satellite images.

15

2.2.1.3. Difference over a Neighborhood. Image distortion on a pixel level can arise from
differences in the gray level of the pixels and/or from the displacements of the pixel. A
distortion measure that penalizes in a graduated way spatial displacements in addition to
gray level differences, and that allows therefore some tolerance for pixel shifts can be
defined as follows [28, 29]:

D5 =

1
2( N w) 2

N w / 2

i, j =w / 2

{(

)}

{(

)}

(l , m ) ] 2 + [ min d C
(i, j ), C(l , m ) ] 2
[ min d C(i, j ), C
l , mwi , j

l , mwi , j

(2.10)

where d (, ) is some appropriate distance metric. Notice that for w=1 this metric reduces to
the mean square error as in D1. Thus for any given pixel C(i, j ) , we search for a best
(i, j ) ,
matching pixel in the d-distance sense in the wxw neighborhood of the pixel C
(i, j ) . The size of the neighborhood is typically small e.g., 3x3 or 5x5, and
denoted as C
w
one can consider a square or a cross-shaped support. Similarly one calculates the distance
from C (i, j ) to C w (i, j ) where again C w (i, j ) denotes the pixels in the wxw neighborhood

(i, j ) is not equal to

of coordinates (i,j) of C (i, j ) . Note that in general d C(i, j ), C
w

(i, j ), C (i, j ) . As for the distance measure d ( , ) , the city metric or the chessboard
dC
w
metric can be used. For example city block metric becomes

city

(i l + j m ) C(i, j ) C (l , m )

C(i, j ), C(l , m ) =
+
N
G

(2.11)

(i, j) vectors. Thus both

where ||.|| denotes the norm of the difference between C(i, j) and C
the pixel color difference and search displacement are considered. In this expression N and
G are one possible set of normalization factors to tune the penalties due to pixel shifts and
pixel spectral differences, respectively. In our measurements we have used the city block
distance with 3x3 neighborhood size.

2.2.1.4 Multiresolution Distance Measure. One limitation of standard objective measures

of distance between images is that the comparison is conducted at the full image
resolution. Alternative measures can be defined that resemble image perception in the

16

human visual system more closely, by assigning larger weights to low resolutions and
smaller weights to the detail image [30]. Such measures are also more realistic in machine
vision tasks that often use local information only.
Consider the various levels of resolution denoted by r 1 . For each value of r the
image is split into blocks b1 to bn where n depends on scale r. For example for r = 1, at the
lowest resolution, only one block covers the whole image characterized by its average gray
level g. For r = 2 one has four blocks each of size (

N N
x ) with average gray levels g11,
2 2

g12, g21 and g22. For the rth resolution level one would have than 2 2 r 2 blocks of size
(

N
N
x r 1 ), characterized by the block average gray levels gij, i, j = 1,...,22 r 2 . Thus for
r 1
2
2

each block bij of the image C , take gij as its average gray level and g ij to corresponds to its
component in the image C (For simplicity a third index denoting the resolution level has
been omitted). The average difference in gray level at the resolution r has weight

1
.
2r

Therefore the distortion at this level is

1 1
d r = r 2r 2
2 2

2 r 1

i , j =1

ij

g ij

(2.12)

where 2 r 1 is the number of blocks along either the i and j indices. If one considers a total
of R resolution levels, then a distance function can simply be found by summing over all

( )

resolution levels, r = 1,.., R, that is D C , C = d r . The actual value of R (the number of

r =1

resolution levels) will be set by the initial resolution of the digital image. For example for
a 512x512 images one has R = 9. Finally for multispectral images one can extend this
definition in two ways. In the straightforward extension, one sums the multiresolution
distances d rk over the bands:

D6 =

1
K

d
k =1 r =1

k
r

(2.13)

17

where d rk is the multiresolution distance in the kth band. This is the multiresolution distance
definition that we used in the experiments. Alternatively the Burt pyramid was constructed
to obtain the multiresolution representation. However in the tests it did not perform as well
as the pyramid described in [30].

A different definition of the multiresolution distance would be to consider the vector

difference of pixels:

( )

R
= d with d = 1 1
D C, C

r
r
2 r 2 2r 2
r =1

[(g
2 r 1

i , j =1

R
ij

g ijR

) + (g
2

G
ij

g ijG

) + (g
2

B
ij

g ijB

)]
2

(2.14)

where, for example, g ijR is the average gray level of the ij'th block in the "red" component
of the image at the (implicit) resolution level r. Notice that in the latter equation the
Euclidean norm of the differences of the block average color components R, G, B have
been utilized.

Notice that the last two measures, that is, the neighborhood distance measure and the
multiresolution distance measure have not been previously used in evaluating compressed
images.

2.2.2. Correlation-Based Measures

2.2.2.1 Image Correlation Measures. The closeness between two digital images can also
be quantified in terms of correlation function [5]. These measures measure the similarity
between two images, hence in this sense they are complementary to the difference-based
measures: Some correlation based measures are as follows:

Structural content
N 1

C1 =

1
K

k =1

C (i, j )

i, j =0
N 1

2
C k (i, j )

i, j =0

(2.15)

18

normalized cross-correlation measure

N 1

1
C2 =
K

k =1

C (i, j )C (i, j )
k

i, j =0

N 1

C (i, j )

i, j =0

(2.16)

A metric useful to compare vectors with strictly non-negative components as in the case of
images is given by the Czekanowski distance [31]:

2
min C k (i, j ) , C k (i, j )

N 1
1
.
C 3 = 2 1 k =K1

N i , j =0
C k (i, j ) + C k (i, j )

k =1

(2.17)

The Czenakowski coefficient (also called the percentage similarity) measures the similarity
between different samples, communities, quadrates.
Obviously as the difference between two images tends to zero = C C 0 , all the
correlation-based measures tend to 1, while as 2 G 2 they tend to 0. Recall also that
distance measures and correlation measures are complementary, so that under certain
conditions, minimizing distance measures is tantamount to maximizing the correlation
measure [32].

2.2.2.2. Moments of the Angles. A variant of correlation-based measures can be obtained

by considering the statistics of the angles between the pixel vectors of the original and
coded images. Similar "colors" will result in vectors pointing in the same direction, while
significantly different colors will point in different directions in the C space. Since we
, we are constrained to one quadrant of the Cartesian space.
deal with positive vectors C, C
Thus the normalization factor of 2 / is related to the fact that the maximum difference
attained will be / 2 .

The combined angular correlation and magnitude difference

between two vectors can be defined as follows [33, 31]:

19

(i, j )

C(i, j ), C
2
ij = 1 1 cos 1
(i, j )

C(i, j ) C

(i, j )

C(i, j ) C
1

3 255 2

(2.18)

We can use the moments of the spectral (chromatic) vector differences as distortion
measures. To this effect we have used the mean of the angle difference (C4) and the mean
of the combined angle-magnitude difference (C5) as in the following two measures:

1
C4 = = 1 2
N

(i, j )
C(i, j ), C
2
1
),
( cos
(i, j )
i , j =1
C(i, j ) C
N

C5 =

1
N2

(2.19)

ij

(2.20)

i, j=1

where is the mean of the angular differences. These moments have been previously
used for the assessment of directional correlation between color vectors.

2.2.3. Edge Quality Measures

According to the contour-texture paradigm of images, the edges form the most
informative part in images. For example, in the perception of scene content by human
visual system, edges play the major role. Similarly machine vision algorithms often rely
on feature maps obtained from the edges. Thus, task performance in vision, whether by
humans or machines, is highly dependent on the quality of the edges and other twodimensional features such as corners [9, 34, 35]. Some examples of edge degradations are:
Discontinuities in the edge, decrease of edge sharpness by smoothing effects, offset of
edge position, missing edge points, falsely detected edge points etc [32]. Notice however
that all the above degradations are not necessarily observed as edge and corner information
in images is rather well preserved by most compression algorithms.

Since we do not possess the ground-truthed edge map, we have used the edge map
obtained from the original uncompressed images as the reference. Thus to obtain edge-

20

based quality measures we have generated edge fields from both the original and
compressed images using the Canny detector [36]. We have not used any multiband edge
detector; instead a separate edge map from each band has been obtained. The outputs of the
derivative of gaussians of each band are averaged, and the resulting average output is
interpolated, thresholded and thinned in a manner similar to that in [12]. The parameters
are set1 as in [36].

In summary for each band k=1...K, horizontal and vertical gradients and their norms,
2

G xk , G ky and N k = G xk + G yk are found. Their average over bands is calculated and

thresholded
Tmin =

1
K

with

min(N ), = 0.1 .
k

T = Tmax Tmin + Tmin ,

Tmax =

where

1
K

max(N )
k

and

Finally they are thinned by using interpolation to find the

pixels where the norms of gradient constitute the local maxima.

2.2.3.1. Pratt Measure. A measure introduced by Pratt [32] considers both edge location
accuracy and missing / false alarm edge elements.

This measure is based on the

knowledge of an ideal reference edge map, where the reference edges should have
preferably a width of one pixel. The figure of merit is defined as:

E1 =

nd
1
1

max{nd , nt } i =1 1 + adi2

(2.21)

where nd and nt are the number of detected and ground-truth edge points, respectively, and
di is the distance to the closest edge candidate for the ith detected edge pixel. In our study
the binary edge field obtained from the uncompressed image is considered as the ground
truth, or the reference edge field. The factor max{nd , nt } penalizes the number of false
alarm edges or conversely missing edges.

This scaling factor provides a relative weighting between smeared edges and thin but
offset edges. The sum terms penalize possible shifts from the correct edge positions. In
1

At http://robotics.eecs.berkeley.edu/~sastry/ee20/cacode.html

21

summary the smearing and offset effects are all included in the Pratt measure, which
provides an impression of overall quality.

2.2.3.2. Edge Stability Measure. Edge stability is defined as the consistency of edge
evidences across different scales in both the original and coded images [37]. Edge maps at
different scales have been obtained from the images by using the Canny [36] operator at
different scale parameters (with the standard deviation of the Gaussian filter assuming
values m = 1.19, 1.44, 1.68, 2.0, 2.38

(The output of this operator at scale m is

thresholded with Tm, where T m = 0.1(C max C min ) + C min . In this expression Cmax and Cmin
denote the maximum and minimum values of the norm of the gradient output in that band.
Thus the edge map at scale m of the image C is obtained as
1 Cm (i, j ) > T m at
E(i, j, m ) =
otherwise
0

(i, j)

(2.22)

where Cm (i, j ) is the output of the Derivative of Gaussian operator at the mth scale. In other
words using a continuous function notation one has Cm (x, y) = C(x, y) Gm (x, y) where

x2 + y2
1
G m ( x, y ) =
xy exp

2
2 m4
2 m

(2.23)

and ** denotes two dimensional convolution. An edge stability map Q(i, j ) is obtained by
considering the longest subsequence E (i, j , m ),..., E (i, j , m +l 1 ) of edge images such that
Q(i, j ) = l

where l = arg max

l

" E (i, j, ) = 1 .
k

(2.24)

m k m + l 1

The edge stability index calculated from distorted image at pixel position i,j will be
denoted by Q (i, j ) . We have used five scales to obtain the edge maps of five band-pass
filtered images.

Finally a fidelity measure called Edge Stability Mean Square Error

(ESMSE) can be calculated by summing the differences in edge stability indexes over all

22

edge pixel positions, nd, that is the edge pixels of the ground-truth (undistorted) image at
full resolution.

E2 =

1
nd

(Q(i, j ) Q (i, j ))
nd

(2.25)

i, j =0

For multispectral images the index in (2.51) can be simply averaged over the bands.
Alternatively a single edge field from multiband images [38, 39] can be obtained and the
resulting edge discrepancies measured as Eq. (2.51).

A property complementary to edge information could be the surface curvature [40],

which is a useful feature for scene analysis, feature extraction and object recognition.
Estimates of local surface types [41], based on the signs of the mean and Gaussian
curvatures, have been widely used for image segmentation and classification algorithms. If
one models a gray level image as a 3-D topological surface, then one can analyze this
surface locally using differential geometry. A measure based on the discrepancy of mean
and Gaussian curvatures between an image and its distorted version has been used in [42].
However this measure was not pursued further due to the subjective assignment of weights
to the surface types and the fact that this measure didn't perform particularly well in
preliminary tests.

2.2.4. Spectral Distance Measures

In this category we consider the distortion penalty functions obtained from the
complex Fourier spectrum of images [10].
Let the Discrete Fourier Transforms (DFT) of the kth band of the original and coded
image be denoted by k (u, v) and k (u, v) , respectively. The spectra are defined as:

k (u , v ) =

N 1

u
v

C (m, n )exp 2im N exp 2in N , k = 1...K

m,n =0

(2.26)

23

Spectral distortion measures, using difference metrics as for example given in (2.12.3) can be extended to multispectral images. To this effect considering the phase and
magnitude spectra, that is

(u , v) = arctan((u, v) ) ,

(2.27)

M (u , v) = (u , v) ,

(2.28)

the distortion occurring in the phase and magnitude spectra can be separately calculated
and weighted. Thus one can define the spectral magnitude distortion

1
S= 2
N

N 1

M (u, v ) M (u, v )

(2.29)

u ,v = 0

the spectral phase distortion

S1 =

1
N2

N 1

(u, v ) (u, v )

(2.30)

u ,v = 0

and the weighted spectral distortion

1
S2 = 2
N

2
2
N 1
N 1

(u, v ) (u , v ) + (1 ) M (u , v ) M (u , v )
u ,v = 0

u ,v = 0

(2.31)

where is to be judiciously chosen e.g., to reflect quality judgment. These ideas can be
extended in a straightforward manner to multiple band images, by summing over all band
distortions. In the following computations, is chosen so as to render the contributions of
the magnitude and phase terms commensurate, that is = 2.5x10 5 .
Due to the localized nature of distortion and/or the non-stationary image field,
Minkowsky averaging of block spectral distortions may be more advantageous. Thus an
image can be divided into non-overlapping or overlapping L blocks of size b x b, say

24

16x16, and blockwise spectral distortions as in (2.14-2.15) can be computed. Let the DFT
of the lth block of the kth band image Ckl (m, n ) be kl (u , v ):

kl (u , v ) =

b 1

C (m, n)exp 2im b exp 2in b

l
k

m, n = 0

(2.32)

b b
where u , v = ... and l = 1,..., L ,or in the magnitude-phase form
2 2
kl (u , v ) = kl (u , v ) e j k (u ,v ) = M kl (u , v )e k (u , v ) .
l

(2.33)

Then the following measures can be defined in the transform domain over the lth block.

J Ml

1
=
K

1
J =
K
l

b 1 l
k (u , v ) kl (u , v )

k =1 u , v = 0
K

b 1 l
k (u, v ) kl (u , v )

k =1 u , v = 0
K

J l = J Ml + (1 )Jl

(2.34)

(2.35)

(2.36)

with the relative weighting factor of the magnitude and phase spectra. Obviously
measures (2.36)-(2.38) are special cases of the above definitions for block size b covering
the whole image. Various rank order operations on the block spectral differences J M and /
or J can prove useful. Thus let J (1),..., J (L ) be the rank ordered block distortions, such

{ }

that for example J (L ) = max J l .

l

Then one can consider the following rank order

1
averages: Median block distortion J 2 + J
2
L

and Average block distortion:

L +1

, Maximum block distortion, J (L ) ;

1 L (i )
J . We have found that median of the block
L i =1

25

distortions is the most effective averaging of rank ordered block spectral distortions and we
have thus used:
S 3 = median J Ml

(2.37)

S 4 = median J l

(2.38)

S 5 = median J l

(2.39)

In this study we have averaged the block spectra with =2 and as for the choice of
the block size we have found that block sizes of 32 and 64 yield better results than sizes in
the lower or higher ranges.

2.2.5. Context Measures

Most of the compression algorithms and computer vision tasks are based on the
neighborhood information of the pixels. In this sense any loss of information in the pixel
neighborhoods, that is, damage to pixel contexts could be a good measure of overall image
distortion. Since such statistical information lies in the context probabilities, that is the
joint probability mass function (p.m.f.) of pixel neighborhoods, changes in the context
probabilities should be indicative of image distortions.

A major hurdle in the computation of context distortions is the requirement to

calculate the high dimensional joint probability mass function. Typical p.m.f. dimensions
would be of the order of s = 10 at least. Consequently one incurs curse of dimensionality
problems. However as detailed in [43, 44], this problem can be solved by judicious usage
of kernel estimation and cluster analysis. A modification of the kernel method is to
identify the important regions in a s-dimensional space X s by cluster analysis and to fit
region-specific kernels to these locations. The result is a model that well represents both
mode and tail regions of p.m.f.s, while combining the summarizing strength of histograms
with generalizing strength of kernel estimates.

26

In what follows we have used the causal neighborhood of pixels i.e., Ck (i, j ) ,
Ck (i 1, j ) , Ck (i, j 1) , Ck (i 1, j 1) , k = 1, 2, 3.

Hence we have derived s = 12

dimensional p.m.f.s obtained from 4-pixel neighborhoods in the 3-bands.

2.2.5.1. Rate-Distortion Based Distortion Measure. A method to quantify the changes in

context probabilities is the relative entropy, defined as
D( p p ) =

p(x )

p( x ) log p (x )

(2.40)

x X s

where

X s denotes a s-pixel neighborhood and x = [x 1,...,x s ] a random vector.

Furthermore p and p are the p.m.f.s of the original image contexts and that of the
distorted (e.g., blurred, noisy, compressed etc.) image. The relative entropy is directly
related to efficiency in compression and error rate in classification. Recall also that the
optimal average bit rate is the entropy of x
H (X ) =

p(X) log p(X) = R( p ).

(2.41)

X X s

If instead of the true probability, a perturbed version p , that is the p.m.f. of the
distorted image, is used, then the average bit rate R( p ) becomes
R( p ) =

p( X) log

p ( X) = H (X ) + D( p p ) .

(2.42)

X X s

The increase in the entropy rate is also indicative of how much the context
probability differs from the original due to coding artifacts. However we do not know the
true p.m.f. p, hence its rate. We can bypass this problem by comparing two competing
compression algorithms, in terms of the resulting context probabilities p 1 and p 2 . If p 1
and p 2 are the p.m.f.s resulting from the two compressed images, then their difference in
relative entropy

27

Z1 = D( p p 1 ) D( p p 2 ) = R( p 1 ) R( p 2 )

(2.43)

is easily and reliably estimated from a moderate-size sample by subtracting the sample
average of log p 2 from that of log p1 [44].

The comparison can be carried out for

more than two images compressed to different bit rates in a similar way, that is comparing
them two by two since the unknown entropy term is common to all of them.

As a quality measure for images we have calculated Z1 for each image when they
were compressed at two consecutive bit rates, for example, R( p1 ) at the bit rate of quality
factor 90 and R( p 2 ) at the bit rate of quality factor 70, etc. Alternatively the distortion was
calculated for an original image and its blurred or noise contaminated version.

2.2.5.2.

f-divergences.

Once the joint p.m.f. of a pixel context is obtained, several

information theoretic distortion measures [45] can be used. Most of these measures can be
expressed in the following general form
p
d ( p, p ) = g E p f
p

where

(2.44)

p
is the likelihood ratio between, p , the context p.m.f. of the distorted image, p the
p

p.m.f. of the original image, and Ep is the expectation with respect to p. Some examples
are as follows:

Hellinger distance, f (x ) =

x 1 , g (x ) =
2

Z2 =

1
2

generalized Matusita distance, f (x ) = 1 x

1
x
2

(
1

r
r

p d ,

, g (x ) = x

(2.45)

28

Z3 =

1/ r

p 1 / r d

r 1.

(2.46)

Notice that integration in (2.44)-(2.45) are carried out in s-dimensional space. Also we
have found according to ANOVA analysis that the choice of r = 5 in the Matusita distance,
yields good results. Despite the fact that the p.m.f.'s do not reflect directly the structural
content or the geometrical features in an image, they perform sufficiently well to
differentiate artifacts between the original and test images.

2.2.5.3. Local Histogram Distances. In order to reflect the differences between two
images at a local level, we calculated the histograms of the original and distorted images
on the basis of 16x16 blocks. To this effect we considered both the Kolmogorov-Smirnov
(KS) distance and the Spearman Rank Correlation (SRC).

For the KS distance we calculated the maximum deviation between the respective
cumulatives. For each of the 16x16 blocks of the image, the maximum of the KS distances
over the K spectral components was found and these local figures were summed over all
b

the blocks to yield

u =1

{ }

max KS uk where KS kb denotes the Kolmogorov-Smirnov distance

k =1.. K

of the block number u and of the kth spectral component. However the KS distance did not
turn out to be effective in the ANOVA tests. Instead the SRC measure had a better
performance. We again considered the SRC on a 16x16 block basis and we took the
maximum over the three spectral bands. The block SRC measure was computed by
computing the rank scores of the gray levels in the bands and their largest for each pixel
neighborhood. Finally the correlation of the block ranks of the original and distorted
images is calculated.

Z 4 = max SRC uk
u =1

k =1.. K

(2.47)

where SRCuk denotes the Spearman Rank Correlation for the uth block number and kth
spectral band.

29

2.2.6. Human Visual System Based Measures

Despite the quest for objective image distortion measure it is intriguing to find out
the role of HVS based measures. The HVS is too complex to be fully understood with
present psychophysical means, but the incorporation of even a simplified HVS model into
objective measures reportedly [7, 46, 10, 14] leads to a better correlation with the
subjective ratings. It is conjectured therefore that in machine vision tasks they may have as
well some relevance.

2.2.6.1. HVS Modified Spectral Distortion. In order to obtain a closer relation with the
assessment by the human visual system, both the original and coded images can be
preprocessed via filters that simulate the HVS. One of the models for the human visual
system is given as a band-pass filter with a transfer function in polar coordinates [46]:

0.05e
H ( ) = 9[ log log 9 ]2.3
10
10
e

<7

0.554

where = u 2 + v 2

1/ 2

(2.48)

. Image processed through such a spectral mask and then inverse

DCT transformed can be expressed via the U {}

operator, i.e.,
U {C (i, j )} = DCT

{H ( u

+ v 2 (u , v )

(2.49)

where (u, v ) denotes the 2-D Discrete Cosine Transform (DCT) of the image and
DCT

is the 2-D inverse DCT.

Some possible measures [5, 47, 48, 49] for the K component multispectral image are

normalized absolute error

H1 =

1
K

N 1
N 1

(i, j ) /

{
(
)
}
,
U
C
i
j

U
C
U {C k (i, j )} ,

k
k

k =1 i , j = 0
i , j =0

(2.50)

30

L2 norm

1
H2 =
K

1
2

k =1 N
K

1/ 2

2
(i, j )
{
(
)
}

,
U
C
i
j
U
C
k
k
i , j =0

N 1

(2.51)

normalized mean square HVS error

H=

1
K

}] / [U {C (i, j )}] .

N 1
U {C k (i, j )} U C k (i, j )

k =1 i , j = 0
K

N 1

i, j =0

(2.52)

2.2.6.2. A Distance Metric for Database Browsing. The metric proposed in [14, 50] based
on a multiscale model of the human visual system, has actually the function of bringing
forth the similarities between image objects for database search and browsing purposes.
This multiscale model includes channels, which account for perceptual phenomena such as
color, contrast, color-contrast and orientation selectivity. From these channels, features are
extracted and then an aggregate measure of similarity using a weighted linear combination
of the feature differences is formed. The choice of features and weights is made to
maximize the consistency with similarity.

We have adopted this database search algorithm to measure discrepancies between

an original image and its distorted version. In other words an image similarity metric that
was conceived for browsing and searching in image databases was adapted to measure the
similarity (or the difference) between an image and its distorted version.

More specifically, we exploit a vision system designed for image database browsing
and object identification to measure image distortion. The image similarity metric in [14]
uses 102 feature vectors extracted at different scales and orientations both in luminance
and color channels. The final (dis)similarity metric is

102

H 3 = i d i
i =1

(2.53)

31

where i are their weights as attributed in [50] and d i are the individual feature
discrepancies. We call this metric browsing metric for the lack of a better name. For
example the color contrast distortion at scale l is given by

1
d =
Nl N l

(
Nl

i , j =0

K (i, j ) K (i, j )

(2.54)

where N l xN l is the size of the image at scale l. K (i, j ) and K (i, j ) denote any color or
contrast channel of the original image and of the coded image at a certain level l. The
lengthy details of the algorithm and its adaptation to our problem are summarized in [14,
50]. Finally note that this measure was used only for color images, and not in the case of
satellite three-band images.

The last quality measure we used that reflects the properties of the human visual system
was the DCTune algorithm [56]. DCTune is in fact a technique for optimizing JPEG still
image compression. DCTune calculates the best JPEG quantization matrices to achieve
the maximum possible compression for a specified perceptual error, given a particular
image and a particular set of viewing conditions. DCTune also allows the user to compute
the perceptual error between two images in units of JND (just-noticeable differences)
between a reference image and a test image. This figure was used as the last metric (H4)
in Table 2.1.
2.3. Goals and Methods

2.3.1 Quality Attributes

Objective video quality model attributes have been studied in [17, 18]. These
attributes can be directly translated to the still image quality measures in the multimedia
and computer vision applications.

Prediction Accuracy: The accurate prediction of distortion, whether for algorithmic

performance or subjective assessment. For example when quality metrics are shown in
box plots as in Fig. 2.1., an accurate metric will possess a small scatter plot.

32

Prediction Monotonicity: The objective image quality measures scores should be

completely monotonic in their relationship to the performance scores.

Prediction Consistency: This attribute relates to the objective quality measures

ability to provide consistently accurate predictions for all types of images and not to fail
excessively for a subset of images.

These desired characteristics reflect on the box plots and the F scores of the quality
metrics, as detailed in the sequel.

2.3.2. Test Image Sets and Rates

All the image quality measures are calculated in their multiband version. In the
study of the quality measures in image compression, we used the two well-known
compression algorithms:

The popular DCT based JPEG [51] and wavelet zero-tree

method, Set Partitioning in Hierarchical Trees (SPIHT), due to Said and Pearlman [52].
The other types of image distortions are generated by the use of blurring filters with
various support sizes and by the addition of white Gaussian noise at various levels.

The rate selection scheme was based on the accepted rate ranges of JPEG. It is wellknown that the JPEG quality factor Q between 80-100 corresponds to visually
imperceptible impairment, Q between 60-80 is the perceptible but not annoying distortion,
Q between 40-60 becomes slightly annoying, Q between 20-40 is annoying, and finally 0120 is the Q range where the quality is very annoying.

Thus each image class was

compressed with 5 JPEG Q factors of 90, 70, 50, 30 and 10. For each class the average
length of compressed files was calculated and the corresponding bit rate (bit/pixel) was
accepted as the class rate. The same rate as obtained from the JPEG experiment was also
used in the SPIHT algorithm.

The test material consisted of the following image sets: 1) Ten three-band remote
sensing images, which contained a fair amount of variety, i.e., edges, textures, plateaus and

33

contrast range, 2) Ten color face images from Purdue University Face Images database
[53], 3) Ten texture images from MIT Texture Database (VISTEX)2.

2.3.3. Analysis of Variance

The Analysis of Variance (ANOVA) [54] was used as a statistical tool to put into the
evidence merits of quality measures. In other words ANOVA was used to show whether
the variation in the data could be accounted for by the hypothesized factor, for example the
factor of image compression type, the factor of image class etc.

The purpose of a one-way ANOVA is to find out whether data from several groups
have a common mean. That is, to determine whether the groups are actually different in the
measured characteristic. In our case each "compression group" consists of quality scores
from various images at a certain bit rate, and there are k = 5 groups corresponding to the 5
bit rates tested. Each group had 30 sample vectors since there were 30 multispectral test
images (10 remote sensing, 10 faces, 10 textures). Similarly three "blur groups" were
created by low-pass filtering the images with 2-D Gaussian-shaped filters with increasing
support. Finally three "noise groups" were created by contaminating the images with
Gaussian noise with increasing variance, that is 2 = 200, 600, 1700. This range of noise
values spans the noisy image quality from the just noticeable distortion to annoying
degradation.

The purpose of two-way ANOVA is to find out whether data from several groups
have a common mean. One-way ANOVA and two-way ANOVA differ in that the groups
in two-way ANOVA have two categories of defining characteristics instead of one. Since
we have two coders (i.e., JPEG and SPIHT algorithms) two-way ANOVA is appropriate.
The hypotheses for the comparison of independent groups are:

H01:

11 = 12 = ... = 1k

means of the groups w.r.t. first factor are equal,

HA1:

1i 1 j

means of the groups w.r.t. first factor are not equal.

H02:

21 = 22 = ... = 2 k

means of the groups w.r.t. second factor are equal,

http://www-white.media.edu/vismod/imagery/VisionTexture/vistex.html

34

HA2:

2i 2 j

means of the groups w.r.t. second factor are not equal.

It should be noted that the test statistic is an F test with k-1 and N-k degrees of
freedom, where N is the total number of compressed images. ANOVA returns the p-value
for the null hypothesis that the means of the groups are equal [54]. A low p-value (high F
value) for this test indicates evidence to reject the null hypothesis in favor of the
alternative. In other words, there is evidence that at least one pair of means are not equal.
We have opted to carry out the multiple comparison tests at a significance level of 0.05.
Thus any test resulting in a p-value under 0.05 would be significant, and therefore, one
would reject the null hypothesis in favor of the alternative hypothesis. This is to assert that
the difference in the quality metric arises from the image coding artifacts and not from
random fluctuations in the image content.

To find out whether the variability of the metric scores arises predominantly from the
image quality, and not from the image set, we considered the interaction between image set
and the distortion artifacts (i.e., compression bit rate, blur etc.).

To this effect we

considered the F-scores with respect to the image set as well. As discussed in sub-section
2.3.1. and shown in Tables 2.2.-2.3, metrics that were sensitive to distortion artifacts were
naturally sensitive to image set variation as well. However for the good measures that
identified the sensitivity to image set variation was always inferior to the distortion
sensitivity.

Boxplot is a garphical way of looking at the distribution of the data in different

groups. Boxplot produces a box and whisker plot for each group. The box has lines at the
lower quartile, median, and upper quartile values. The whiskers are lines extending from
each end of the box to show the extent of the rest of the data. Outliers are data with values
beyond the ends of the whiskers. If the F-value is high, there will be little overlap between
the two or more groups. If the F-value is not high, there will be a fair amount of overlap
between all of the groups. In the box plots, a steep slope and little overlap between boxes,
as illustrated in Figure 2.1, are both indicators of a good quality measure. In order to
quantify the discriminative power of a quality measure, we have normalized the difference
of two successive group means normalized by respective variances, i.e.,

35

a) H2 JPEG F=2291, p=0

b) D1 JPEG F=104.6, p=0

c) C4 SPIHT F=7.91, p=0

Figure 2.1. Box plots of quality measure scores. a) good measure, b) moderate measure,
c) poor measure. The F-scores as well as the significance level p are given.

Q r , r +1 =

r r +1

Q = Ave{Q r , r +1 }

r r +1
r = 1,..., k 1

(2.55)

(2.56)

36

where r denotes the mean value of the image quality measure for the images compressed
at rate r and r is the standard deviation, k is the number of different bit rates at which
quality measures are calculated. A good image quality measure should have high Q value,
which implies little overlap between groups and/or large jumps between them hence high
discriminative power of the quality measure. It should be noted that the Q values and the
F-scores yielded totally parallel results in our experiments. In Figure 2.1. we give box plot
examples of a good, a moderate and a poor measure. For the box plot visualization the
data has been appropriately scaled without any loss of information. The horizontal axis
corresponds to bitrate variation and the vertical axis is the normalized IQM scores.

2.3.4. Visualization of Quality Metrics

Since we would like to visualize the quality metrics data, we organize them as
vectors and feed them to a SOM (Self-Organizing Map) algorithm. The elements of the
vectors are the corresponding quality scores. For example, consider the MSE error (D1)
for a specific compression algorithm (e.g., JPEG) at a specific rate. The corresponding
vector D1 is M dimensional, where M is the number of images, and it reads as:
D1(JPEG, bitrate) = [D1(1| JPEG, bitrate) .... D1(M| JPEG, bitrate]T

(2.57)

There will be 5 such vectors, one for each bit rate considered. Overall for training of SOM
we utilize 30 images x 5 bit rates x 2 compressors x 26 metrics = 7800 vectors.

Recall that the self-organizing map (SOM) is a tool for visualization of high
dimensional data.

It maps complex, non-linear high dimensional data into simple

geometric relationships on a low dimensional array and thus serves to produce

abstractions. Among the important applications of the SOM one can cite the visualization
of high dimensional data, as the case in point, and discovery of categories and abstractions
from raw data.
Let the data vectors be denoted as X = [x 1 ,..., x M ] R M , where M is the number of
T

images considered (M = 30 in our case).

With each element in the SOM array, a

37

parametric real vector m i = [ i1 ,..., iM ] R M is associated. The image of an input

T

vector X on the SOM array is defined by the decoder function d(X, m i ) , where d (.,.) is a
general distance measure. The image of the input vector will have the array index c
definedas c = arg min d(X, m i ) . A critical part of the algorithm is to define the m i in such
i

a way that the mapping is ordered and descriptive of distribution of X . Finding such a set
of values that minimize the distance measure resembles in fact the standard VQ problem.
In contrast, the indexing of these values is arbitrary, whereby the mapping is unordered.
However if the minimization of the objective functional based on the distance function is
implemented under the conditions described in [55], then one can obtain ordered values of
m i , almost as if the m i were lying at the nodes of an elastic net. With the elastic net
analogy in mind, SOM algorithm can be constructed as
m i (t + 1) = m i (t ) + ( t )[ X( t ) m i (t )]

(2.58)

where ( t ) is a small scalar, if the distance between units c and i in the array is smaller
than or equal to a specified limit (radius), and ( t ) = 0 otherwise. During the course of
ordering process, (t ) is decreased from 0.05 to 0.02, while radius of neighborhood is
decreased from 10 to 3. Furthermore scores are normalized with respect to the range.
The component planes j of the SOM, i.e., the array of scalar values ij representing
the j'th components of the weight vectors m i and having the same format as the SOM
array is displayed as shades of gray.

2.4. Statistical Analysis of Image Quality Measures

Our first goal is to investigate the sensitivity of quality measures to distortions

arising from image compression schemes. In other words to find out the degree to which a
quality measure can discriminate the coding artifacts and translate it into a meaningful
score.

We establish similarly the response sensitivity of the measures to such other

distortion effects as blur and noise. Our second goal is to establish how various quality
measures are related to each other and to show the degree to which measures respond to

38

(dis)similarly to coding and sensor artifacts. As the outcome of these investigations we

hope to extract a small subset of measures that hopefully satisfies the above desiderata.

2.4.1. ANOVA Results

The two-way ANOVA results of the image quality measures for the data obtained
from all image classes (Fabrics, Faces, Remotes) are listed in Table 2.2. In these tables
the symbols of quality measures D1, D2...H3, H4 are listed in the first column while the Fscores of JPEG compression, of SPIHT compression, of blur and of noise distortions are
given, respectively, in the succeeding four columns. The first factor tested is the bitrate
variation and the second factor is the image set variation.
The metric that responds most strongly to one distortion type is called the
fundamental metric of that distortion type [24]. Similarly the metric that responds to all
sorts of distortion effects is denoted as the global metric. One can notice that:

The fundamental metrics for JPEG compression are H2, H1, S2, E2, that is, HVS
L2 norm, HVS absolute norm, spectral phase-magnitude, and edge stability
measures. These measures are listed in decreasing order of F-score.

The fundamental metrics for SPIHT compression are E2, S2, S5, H2, that is, edge
stability, spectral phase-magnitude, block spectral phase-magnitude, and HVS L2
norm.

The fundamental metrics for the BLUR effect are S1, E2, S2, H1, that is, spectral
phase, edge stability, spectral phase-magnitude, HVS absolute norm. Notice the
similarity of metrics between SPIHT and blur due, in fact, to the blurring artifact
encountered in wavelet-based compression.

The fundamental metric for the NOISE effect is, as expected, D1, the mean
square error.

39

Finally the image quality metric that is sensitive to all distortion artifacts are, in
order, E2, H1, S2, H2, S5, that is, edge stability, HVS absolute norm, spectral
phase-magnitude, HVS L2 norm, block spectral phase-magnitude.

Table 2.2. ANOVA results (F-scores) for the JPEG and SPIHT compression distortions as
well as additive noise and blur artifacts. For each distortion type the variation due to
image set is also established

ANOVA2
D1
D2
D3
D4
D5
D6
C1
C2
C3
C4
C5
E1
E2
S1
S2
S3
S4
S5
Z1
Z2
Z3
Z4
H1
H2
H3
H4

JPEG
Bitrate F
Imageset
Score
F Score
104.6
42.59
108.5
67.45
63.35
29.37
89.93
1.99
20.26
80.71
76.73
5.94
1.35
124.6
12.26
93.83
82.87
83.06
45.65
47.36
91.42
38.17
26.24
3.64
176.3
92.75
150.5
102.2
191.3
98.42
145.6
56.39
129.1
63.26
146.1
71.03
1.69
141.8
7.73
114.7
17.63
223
9.4
23.58
371.9
0.09
2291
5.46
123
1.2
78.83
7.14

SPIHT
Bitrate F
Imageset F
Score
Score
39.23
13.28
29.56
15.93
53.31
48.53
13.75
3.71
14.09
68.22
37.52
11.22
12.05
325.5
15.18
82.87
24.96
22.42
7.91
5.94
27.51
5.28
77.86
137
212.5
200.4
104
68.17
161
101.8
38.58
26.97
128
46.85
144.1
61.65
21.36
14
11.41
77.68
23.22
181.4
9.84
32.41
107.2
40.05
132.9
22.82
27.45
7.6
25.2
95.72

Blur F
Score
43.69
33.94
38.55
27.87
6.32
412.9
5.61
11.19
30.92
16.48
52.57
125.8
768.7
1128
572.2
24.28
215
333.6
35.9
10.17
17.26
8.45
525.6
47.28
67.31
12.55

BLUR
Imageset
F Score
2.06
17.76
24.13
0.96
55.11
45.53
107.2
39.77
1.71
0.77
2.44
21.09
23.41
60.04
17.95
6.39
11.17
27.84
62.5
1.80
8.31
14.74
69.98
101.7
6.77
2.11

NOISE
Noise F
Imageset
Score
F Score
9880
17.32
6239
20.4
1625
11.15
166.4
9.88
1981
43.51
44.61
4.38
3.82
6.17
58.04
45.63
567.5
52.01
198.8
19.03
704
10.8
87.76
27.87
158.5
24.84
47.29
38.42
107.1
4.83
2803
8.59
56.04
55.1
78.04
26.53
44.89
110.9
3.03
11.36
14.71
21.12
24.99
3.31
230.7
19.57
624.3
21.32
117.3
0.50
29.06
6.69

To establish the global metrics, we gave rank numbers from 1 to 26 to each one
metric under the four types of distortion as in Table 2.2. For example for JPEG the metrics
are ordered as H2, H1, S2, E2, etc. if we take into consideration their F-scores. Then we
summed the rank numbers and the metrics for which the sum of the scores were the
smallest were declared as the global metric, that is the one that qualifies well in all
discrimination tests.

40

The metrics that were the least sensitive to image set variation are D4, H3, C4, C5,
D6 etc.. However it can be observed that these metrics show in general poor performance
in discriminating distortion effects. On the other hand for the distortion sensitive metrics,
even though their image set dependence is higher than the so-called image independent
metrics, more of the score variability is due to distortion than to image set change. This can
be observed based on the higher F-scores for distortion effects as compared to image set
related F-scores.

These observations are summarized in Table 2.3. where one-way

ANOVA results are given for each image class (Fabrics, Faces, Remote Sensing)
separately, and two-way ANOVA results are presented for the combined set. In the two
bottom rows of Table 2.3. the metrics that are least sensitive to the coder type and to the
image set are given.

Table 2.3. One-way ANOVA results for each image class and two-way ANOVA results
for the distortions on the combined and image set independence
1-way
IMAGE SET
ANOVA Fabrics
Faces
Remote Sensing
2-way
Combined Set
ANOVA Image Set
Independence
Coder Type
Independence

JPEG
H4,H2,E2,S4
H2, D1,S3,H1
H2,H4,S4,S5
H2,H1,S2,E2
H1,H3

SPIHT
E1,S1,E2,S2
H4,D3,H2,C1
S2,S5,S4,S1
E2,S2,S5,H2
D4,C5

BLUR
S1,S5,E2,S4
S2,H1,S1,E2
D6,S5,S4,S1
S1,E2,S2,H1
C4,D4

NOISE
D1,D2,D5,D3
D1,S3,D2,D3
D1,D2,C3,C5
D1,D2,S3,D5
H3,Z4

D2,D1,Z4,D3

We also investigated the metrics with respect to their ability to respond to bit rate and
coder type. More specifically the first factor tested was bitrate variation and the second
was coder type variation. For this analysis the scores of the JPEG and SPIHT compressors
were combined. It was observed in Table 2.4. that:

The metrics that were best in discriminating compression distortion as

parameterized by the bit rate, whatever the coder type, that is JPEG or SPIHT,
were H2, H1, S2, S5 (HVS L2 norm, HVS absolute norm, spectral phasemagnitude, block spectral phase-magnitude etc.

41

The metrics that were capable of discriminating the coder type (JPEG versus
SPIHT) were quite similar, that is, D6, H2, H4 and H1 (Multiresolution error,
HVS L2 norm, DCTune, HVS L1 norm).

Finally the metrics that were most sensitive to distortion artifacts, but at the same
time, least sensitive to image set variation were C5, D1, D3, S3, D2, C4..., (Mean
angle-magnitude similarity, Mean square error, Modified infinity norm, Block
spectral magnitude error, Mean absolute error, Mean angle similarity...).

These metrics were identified by summing the two rank scores of metrics, the first
being the ranks in ascending order of distortion sensitivity, the second being in descending
order the image set sensitivity. Interestingly enough almost all of them are related to the
mean square error varieties. Despite its many criticisms, this may explain why mean square
error or signal-to-noise ratio measures have proven so resilient over time.

Table 2.4. ANOVA scores for the bit rate variability (combined JPEG and SPIHT scores)
and coder variation

ANOVA2
D1
D2
D3
D4
D5
D6
C1
C2
C3
C4
C5
E1
E2
S1
S2
S3
S4
S5
Z1
Z2
Z3
Z4
H1
H2
H3
H4

JPEG+SPIHT
Bitrate
Coder
89.79
0.75
74.98
2.72
71.55
1.21
70.52
43.85
17.07
0.0005
85.22
118.8
2.66
45.47
12.28
18.27
56.48
1.56
31.3
2.43
78.98
2.23
42.69
11.61
122.4
26.28
99.12
5.29
140.1
12.37
92.99
9.27
115.5
39.1
124.8
43.09
4.28
41.6
9.54
0.83
12.87
0.56
9.39
6.64
278.6
52.87
493
87.21
97.94
16.19
21.13
57.72

42

As expected the metrics that are responsive to distortions are also almost always
responsive to the image set. Conversely the metrics that do not respond to the image set
variation are also not very discriminating with respect to the distortion types. The fact that
the metrics are sensitive, as should be expected, to both the image content and distortion
artifacts does not eclipse their capability to potential as quality metrics. Indeed when the
metrics were tested within more homogeneous image sets (that is only within Face images
or Remote Sensing images etc.) the same high-performance metrics scored consistently
higher. Furthermore when one compares the F-scores of the metrics with respect to bit rate
variation and image set variation, even though there is a non-negligible interaction factor,
one can notice that the F-score due to bit rate is always larger than the F-score due to
Image sets.

2.4.2. Self Organizing Map of Quality Measures

Our second investigation was on the mutual relationship between measures. It is

obvious that the quality measures must be correlated with each other as most of them must
respond to compression artifacts in similar ways. On the other hand one can conjecture
that some measures must be more sensitive to blurring effects, while others respond to
blocking effects, while still some others reflect additive noise.

Self Organizing Map (SOM) [55] is a pictorial method to display similarities and
differences between statistical variables, such as quality measures. We have therefore
obtained spatial organization of these measures via Kohonens self-organizing map
algorithm. The input to the SOM algorithm was vectors whose elements are the scores of
the measure resulting from different images.

More explicitly, consider one of the

measures, D1, and a certain compression algorithm, e.g., JPEG. The instances of this
vector will be 60-dimensional, one for each of the images in the set.

The first 30

components consist of 30 images compressed with JPEG, the next 30 juxtaposed

components of the same images compressed with SPIHT. Furthermore there will be five
such vectors, one for each one of the bit rates.

43

Figure 2.2. SOM of distortion measures for JPEG and SPIHT

The SOM organization of the measures in the 2-D space for pooled data from JPEG
and SPIHT coders is shown in Figure 2.2. The map consists of 70 X 70 cells. These maps
are useful for the visual assessment of possible correlation present in the measures. One
would expect that measures with similar trends and that respond in similar ways to artifacts
would cluster together spatially. The main conclusions from the observation of the SOM
and the correlation matrix are as follows:

Clustering tendency of pixel difference based measures (D1, D2, D4, D5) and
spectral magnitude based method (S3) is obvious in the center portion of the map,
a reflection of the Parseval relationship, that is, distortion in image energy in
spatial domain matches in the same way in the frequency domain. However
notice that spectral phase measures (S2, S5) stay distinctly apart from these
measures.

44

The human visual system based measures (H2, H3, H4), multiresolution pixeldifference measure (D6), E2 (edge stability measure) and C5 (mean anglemagnitude measure) are clustered in the right side of the map. The correlation of
the multiresolution distance measure, D6 with HVS based measures (H2, H3, H4)
is not surprising since the idea behind this measure is to mimic image comparison
by eye more closely, by assigning larger weight to low resolution components
and less to the detailed high frequency components.

The three correlation based measures (C1, C2, C3) are together in the lower part
of the map while the two spectral phase error measures (S2, S5) are concentrated
separately in the upper part of the map.

It is interesting to note that all the context-based measures (Z1, Z2, Z3, Z4) are
grouped in the upper left region of the map together with H1 (HVS filtered
absolute error).

The proximity between the Pratt measure (E1) and the maximum difference
measures (D3) is meaningful, since the maximum distortions in reconstructed
images are near the edges.

The constrained maximum distance or sorted

maximum distance measures can be used in codec designs to preserve the two
dimensional features, such as edges, in reconstructed images.

2.5. Conclusions

We have presented collectively a set of image quality measures in their multispectral

version and categorized them. Statistical investigation of 26 different measures using a
ANOVA analyses has revealed that local phase-magnitude measures (S2 or S5), HVSfiltered L1, L2 norms, edge stability measure are most sensitive to coding, blur and
artifacts, while the mean square error (D1) remains as the measure for additive noise. One
can state that combined spectral phase-magnitude measures and HVS filtered error norms
should be paid more attention in the design of coding algorithms and sensor evaluation.
On the other hand the pixel-difference based measures remain still to be the measures
responsive to distortions and least affected by image variety.

45

The Kohonen map of the measures has been useful in depicting similar ones, and
identifying the ones that are sensitive possibly to different distortion artifacts in
compressed images. The correlation between various measures has been depicted via
Kohonens Self-Organizing Map. The placement of measures in the two-dimensional map
has been in agreement with ones intuitive grouping.

Future work will address subjective experiments and prediction of subjective image
quality using the above salient measures identified. Another possible avenue is to combine
various fundamental metrics for better performance prediction.

46

3. STEGANALYSIS USING IMAGE QUALITY METRICS

3.1. Introduction

Steganography refers to the science of invisible communication.

Unlike

cryptography, where the goal is to secure communications from an eavesdropper,

steganographic techniques strive to hide the very presence of the message itself from an
observer. Although steganography is an ancient subject, the modern formulation of it is
often given in terms of the prisoners problem [57, 58, 59] where Alice and Bob are two
inmates who wish to communicate in order to hatch an escape plan.

However, all

communication between them is examined by the warden, Wendy, who will put them in
solitary confinement at the slightest suspicion of trouble. Specifically, in the general
model for steganography, we have Alice wishing to send a secret message m to Bob. In
order to do so, she embeds m into a cover-object c, to obtain the stego-object s. The
stego-object s is then sent through the public channel.

The warden Wendy who is free to examine all messages exchanged between Alice
and Bob can be passive or active. A passive warden simply examines the message and
tries to determine if it potentially contains a hidden message. If it appears that it does, then
she takes appropriate action, else she lets the message through without any action. An
active warden, on the other hand, can alter messages deliberately, even though she does not
see any trace of a hidden message, in order to foil any secret communication that can
nevertheless be occurring between Alice and Bob. The amount of change the warden is
allowed to make depends on the model being used and the cover-objects being employed.
For example, with images, it would make sense that the warden is allowed to make
changes as long as she does not alter significantly the subjective visual quality of a
suspected stego-image.

It should be noted that the main goal of steganography is to communicate securely in

a completely undetectable manner. That is, Wendy should not be able to distinguish in any
sense between cover-objects (objects not containing any secret message) and stego-objects
(objects containing a secret message). In this context, steganalysis refers to the body of

47

techniques that are designed to distinguish between cover-objects and stego-objects. It

should be noted that nothing might be gleaned about the contents of the secret message m.
When the existence of hidden message is known, revealing its content is not always
necessary.

Just disabling and rendering it useless will defeat the very purpose of

steganography. In this paper, we present a steganalysis technique for detecting stegoimages, i.e. still images containing hidden messages, using image quality metrics.
Although we focus on images, the general techniques we discuss would also be applicable
to audio and video data.

Given the proliferation of digital images, and given the high degree of redundancy
present in a digital representation of an image (despite compression), there has been an
increased interest in using digital images as cover-objects for the purpose of
steganography. The simplest of such techniques essentially embed the message in a subset
of the LSB (least significant bit) plane of the image, possibly after encryption [60]. It is
well known that an image is generally not visually affected when its least significant bit
plane is changed. Popular steganographic tools based on LSB like embedding vary in their
approach for hiding information. Methods like Steganos and Stools use LSB embedding in
the spatial domain, while others like Jsteg embed in the frequency domain.
techniques include the use of quantization and dithering.
steganography techniques, the reader is referred to [60].

Other

For a good survey of

What is common to these

techniques is that they assume a passive warden framework. That is they assume the
warden Wendy will not alter the image. We collectively refer to these techniques as
conventional steganography techniques or for brevity, more simply as steganography
techniques.

Conventional steganography techniques like LSB embedding techniques are not

useful in the presence of an active warden as the warden can simply randomize the LSB
plane to thwart communication. In order to deal with an active warden Alice must embed
her message in a robust manner. That is, Bob should be able to accurately recover the
secret message m despite operations like LSB randomizing, compression, filtering, rotation
by small degrees etc. performed by the active warden Wendy. Indeed, the problem of
embedding messages in a robust manner has been the subject of active research in the
image processing community under the name of digital watermarking [61, 62, 63].

48

A digital watermark is an imperceptible signal added to digital content that can be

later detected or extracted in order to make some assertion about the content. For example,
the presence of her watermark can be used by Alice to assert ownership of the content.
Given the proliferation of content in digital form, recent years have seen an increasing
interest in digital watermarking and in the past few years, many different watermarking
algorithms have been proposed for different applications. Although the main applications
for digital watermarking appear to be copyright protection and digital rights management,
watermarks have also been proposed for secret communication, that is, steganography.
Essentially digital watermarks provide a means of image-based steganography in the
presence of an active warden since modifications made by the warden will not affect the
embedded watermark as long as the visual appearance of the image is not significantly
degraded.

However, despite this obvious and commonly observed connection to steganography,

there has been very little effort aimed at analyzing or evaluating the effectiveness of
watermarking techniques for steganographic applications. Instead, most work has focused
on analyzing or evaluating the watermarking algorithms for their robustness against
various kinds of attacks that try to remove or destroy them.

However, if digital

watermarks are to be used in steganography applications, detection of their presence by an

unauthorized agent defeats their very purpose. Even in applications that do not require
hidden communication, but only robustness, we note that it would be desirable to first
detect the possible presence of a watermark before trying to remove or manipulate it. This
means that a given signal would have to be first analyzed for the presence of a watermark.
Based on this analysis there could then be attempts made to remove the watermark.

In this thesis, we develop steganalysis techniques both for conventional LSB like
embedding used in the context of a passive warden model and for watermarking which can
be used to embed secret messages in the context of an active warden.

In order to

distinguish between these two models, we will be using the terms watermark and message
when the embedded signal is in the context of watermarking and conventional
steganography, respectively. Furthermore, we simply use the terms marking or embedding
when the context of discussion is general to include both watermarking and steganography.

49

The techniques we present are novel and to the best of our knowledge, the first
attempt at designing general purpose tools for steganalysis. General detection techniques
as applied to steganography have not been devised and methods beyond visual inspection
and specific statistical tests for individual techniques [64, 65, 66, 67] are not present in the
literature. Since too many images have to be inspected visually to sense hidden messages,
the development of a technique to automate the detection process will be very valuable to
the steganalyst.

Our approach is based on the fact that hiding information in digital media requires
alterations of the signal properties that introduce some form of degradation, no matter how
small. These degradations can act as signatures that could be used to reveal the existence
of a hidden message. For example, in the context of digital watermarking, the general
underlying idea is to create a watermarked signal that is perceptually identical but
statistically different from the host signal. A decoder uses this statistical difference in
order to detect the watermark. However, the very same statistical difference that is created
could potentially be exploited to determine if a given image is watermarked or not. In this
thesis, we show that addition of a watermark or message leaves unique artifacts, which can
be detected using Image Quality Measures [68, 69, 70, 71, 72].

The rest of this section is organized as follows. In Section 3.2., we discuss the
selection of the image quality measures to be used in the steganalysis and the rationale of
utilizing concurrently more than one quality measure. We then show that the image
quality metric based distance between an unmarked image and its filtered version is
different as compared to the distance between a marked image and its filtered version.
Section 3.3. describes the regression analysis that we use to build a composite measure of
quality to indicate the presence or absence of a mark. Statistical tests and experiments are
given in Section 3.4. and, finally, conclusions are drawn in Section 3.5.
3.2. Choice of Image Quality Measures

As stated in the introduction, the main goal of this is to develop a discriminator for
message or watermark presence in still images, using an appropriate set of IQMs. Image
quality measurement continues to be the subject of intensive research and experimentation

50

[4, 5, 2, 73, 47].

Objective image quality measures are based on image features, a

functional of which, correlates well with subjective judgment, that is, the degree of
(dis)satisfaction of an observer [13]. The interest in developing objective measures for
assessing multimedia data lies in the fact that subjective measurements are costly, timeconsuming and not easily reproducible.

Objective measures are also utilized in

performance prediction of vision algorithms against quality loss due to sensor inadequacy
or compression artifacts [24]. In this work, however, we want to exploit image quality
measures, not as predictors of subjective image quality or algorithmic performance, but as
steganalysis tools, that is, as detection features of watermarks or hidden messages.

A good IQM should be accurate, consistent and monotonic in predicting quality as

already mentioned in second section. In the context of steganalysis, however, prediction
accuracy can be interpreted as the ability of the measure to detect the presence of hidden
message with minimum error on average. Similarly, prediction monotonicity signifies that
IQM scores should ideally be monotonic in their relationship to the embedded message
size or watermark strength. Finally, prediction consistency relates to the quality measures
ability to provide consistently accurate predictions for a large set of watermarking or
steganography techniques and image types. This implies that the spread of quality scores
due to image variety should not eclipse the score differences arising from message
embedding artifacts.

The steganalysis technique we develop is based on regression analysis of a number

of relevant IQMs. Hence, we seek IQMs that are sensitive specifically to watermarking
and steganography effects. In other words, those measures for which the variability in
score data can be explained better because of treatment rather then as random variations
due to the image set. The idea behind detection of watermark or hidden message presence
is to obtain a consistent distance metric for images containing a watermark or hidden
message and those without, with respect to a common reference processing. The reference
processing we have used was low-pass filtering based on a Gaussian kernel. The filter was
chosen as a Gaussian smoothing filter

H (m, n ) = Kg (m, n )

(3.1)

51

where

g (m, n ) = 2 2

{(

exp m 2 + n 2 / 2 2

(3.2)

is the 2-D Gaussian kernel and

K = ( g (m, n ) ) 1 / 2
2

(3.3)

is the normalizing constant. We experimentally choose = 0.5 , implemented via a 3x3

filter. The reason why Gaussian blurring works fine as a common reference is that it gives
us the local mean which is also the maximum likelihood (ML) estimate of the image under
Gaussian assumption [74]. Under Laplacian distribution assumption the median would
have been the ML estimate. Therefore the blurred image minus the original yields, in fact,
the maximum likelihood estimate of the additive watermark. In fact we have tested both
the mean and median filters as the ML estimates of the image and we have found out that
the former performs slightly better in the detection tests.

Most watermarking techniques or steganographic message embedding techniques,

whether by spread-spectrum or quantization modulation or LSB insertion, can be
represented as a signal addition to the cover image, as shown in Figure 3.1. Let f be the
cover image, g = f + w be the stego-image, and w the inserted watermark. Let H be the
ML operator for the estimate of the watermark sequence. For the two ML estimators that
we have tested, H obviously corresponds to the subtraction from the received stego-image
of its local mean or median. In the absence of any watermark or stego-signal Hg = f
corresponds to the high-frequency content f of the image, while for a marked signal it
yields Hg = f + w where w denotes the ML estimate of the mark. The image quality
metrics, in fact, are trained to differentiate between these two signals f and f + w .
Figure 3.2. gives an instance of the watermarked-nonwatermarked class separability based
on a scatter diagram of the three image quality metrics used.

52

Watermark Signal w
Cover Signal f

Stego Signal g = f + w

+
(a)

f + w

(b)

(c)

Figure 3.1. Schematic descriptions of (a) watermarking or stegoing, (b) filtering an unmarked image, (c) filtering a marked image.

As for the selection of quality measures we have gleaned out the ones that serves
well the purpose of our steganalysis. The rationale of using several quality measures is
that different measures respond with differing sensitivities to artifacts and distortions. For
example, some measures like mean square error respond more to additive noise, others
such as spectral phase or mean square HVS-weighted (Human Visual System) error are
more sensitive to pure blur, while the gradient measure reacts to distortions concentrated
around edges and textures. Recall that some watermarking algorithms inject noise in block
DCT coefficients, others in a narrow-band of global DCT or Fourier coefficients, still
others operate in selected localities in the spatial domain. Since we want our steganalyzer
to be able to work with a variety of watermarking and steganography algorithms, a
multitude of quality features are needed so that the steganalyzer has the chance to probe all
features in an image that are significantly impacted by the embedding watermarking or
steganographic process.

In order to identify specific quality measures that are useful in steganalysis, we use
ANOVA [54] test. ANOVA helped us to distinguish measures that are most consistent and
accurate vis--vis the effects of watermarking and of steganography. Various quality
measures are subjected to a statistical test to determine if the fluctuations of the measures

53

unmarked
marked

1.002

1.001

M6

0.999

0.998
0.01
0.99

0.005
0.995
0

M3
M5

Figure 3.2. Scatter plots of the three Image Quality Measures (M3: Czenakowski measure,
M5: Image Fidelity, M6: Normalized Cross-correlation).

result from image variety or whether they arise due to treatment effects, that is,
watermarking or stegoing.

We performed three different ANOVA tests: The first was for watermarking, the
second for steganography, and the last one for both watermarking and steganography.

For watermarking, the first group consisted of the IQM scores computed from plain
images and their filtered versions. The remaining three groups consisted of the IQM
scores computed from watermarked images by Digimarc [75], PGS [76] and COX [77]
techniques, respectively, and their filtered versions. The data given to the ANOVA
algorithm consisted of four vectors, each of dimensions N , where N = 12 is the number
of images used in the test from the training set. More specifically, consider a typical
quality measure, say M ( i ) , where the parametric dependence upon the watermarking
algorithm is shown with i , i = 0! 3 , for plain images, Digimarc, PGS and COX
techniques, respectively. The N-dimensional vector M reads as:

54

M ( i ) = [M (1 i )! M (N i )] .
T

(3.4)

For steganography, the first group consisted of the IQM scores computed from plain
(non-marked) images and their filtered versions. The remaining three groups consisted of
the IQM scores computed from stegoed images by Steganos [78], Stools [79] and Jsteg
[80], respectively, and their filtered versions.

For the joint watermarking and steganography analysis, the first group consisted of
the IQM scores computed from plain images and their filtered versions. The remaining six
groups consisted of the IQM scores computed from watermarked images by Digimarc,
PGS and COX technique, stegoed images by Steganos, Stools and Jsteg, respectively, and
their filtered versions.

In Table 3.1. we give ANOVA results with respect to watermarking, steganography

and combined techniques. The measures that have relatively higher discriminative power,
measures that catch the statistical evidence of watermarking or steganography, are shown
in bold. These measures, in fact, sense the statistical difference between the populations of
marked and non-marked images and hence they can be used to separate the two classes.

The implications of the result are two fold. One is that, using these features a
steganalysis tool can be designed to detect the watermarked or stegonagraphically marked
images, as we show in Section 3.3., using multivariate regression analysis. The other is
that, current watermarking or steganographic algorithms should exercise more care on
those statistically significant image features to eschew detection. For instance, the relative
ordering of the statistically significant IQMs for watermarking and steganographic
algorithms are different. While the Minkowsky measures were not statistically significant
for steganographic algorithms, they were for the watermarking algorithms. Minimizing the
Mean Square Error (MSE) or the Kullbeck-Leibler distance between the original (cover)
image and the stego image is not necessarily enough to achieve covert communication as
the evidence can be caught by another measure such as spectral measures. The selected
subset of image quality measures in the design of steganalyzer with respect to their
statistical significance were as follows:

55

Table 3.1. One-way ANOVA tests for watermarking, steganography and pooled
waterwarking and steganography

Image Quality Measures

Minkowsky Metric = 2
Minkowsky Metric = 1
Maximum Difference
Sorted Maximum Difference
Czenakowski
Structural Content
Cross Correlation
Image Fidelity
Angle Mean
Angle Standard Deviation
Spectral Magnitude
Spectral Phase
Weighted Spectral Distance
Median Block Spectral Magnitude
Median Block Spectral Phase
Median Block Weighted Spectral Dist.
Normalized Absolute Error (HVS)
Normalized MS ERROR (HVS)
HVS Based L2

Stego

Watermark&Stego

Watermark
F
p
0.01
6.06

F
0.56

p
0.58

F
5.28

p
0.00

3.28

0.05

0.57

0.58

3.07

0.03

0.13
0.14
4.63
0.62
2.08
2.67
1.95
0.45
5.50
5.49
1.12
0.79
0.47
0.45
0.16
3.30
0.19

0.93
0.93
0.02
0.61
0.14
0.08
0.17
0.72
0.03
0.03
0.37
0.51
0.72
0.72
0.92
0.05
0.90

0.31
0.07
1.08
0.15
0.21
0.40
4.20
3.27
0.02
0.02
0.06
0.001
3.95
3.96
1.16
4.93
0.46

0.74
0.92
0.37
0.86
0.81
0.68
0.04
0.08
0.98
0.98
0.94
0.99
0.05
0.05
0.35
0.02
0.64

0.25
0.13
4.66
0.58
0.74
1.14
3.40
2.36
4.35
4.34
0.66
0.44
4.24
4.22
0.74
2.69
0.47

0.93
0.98
0.01
0.71
0.60
0.37
0.02
0.08
0.01
0.01
0.65
0.81
0.02
0.02
0.61
0.05
0.79

Watermarking: Mean Absolute Error D2, Mean Square Error D1, Czekanowski
Correlation Measure C3, Image Fidelity C2, Cross Correlation C1, Spectral Magnitude
Distance S, Normalized Mean Square HVS Error H.

We denote this feature set as

= {D1 , D2 , C1 , C2 , C3 , S , H } for future reference in the experiments in Section 3.4.

Steganography: Angle Mean C4, Median Block Spectral Phase Distance S4, Median
Block Weighted Spectral Distance S5, Normalized Mean Square HVS Error H. We denote
this feature set as = {C4 , S4 , S5 , H }.

Pooled Watermarking And Steganography: Mean Absolute Error D2, Mean Square
Error D1, Czekanowski Correlation Measure C3, Angle Mean C4, Spectral Magnitude
Distance S, Median Block Spectral Phase Distance S4, Median Block Weighted Spectral
Distance S5, Normalized Mean Square HVS Error H. We denote this feature set as
= {D1 , D2 , C 3 , C 4 , S 4 , S 5 , H }.

56

3.3. Regression Analysis of the Quality Measures

The steganalysis we propose is based on the observation in Section 3.2. that an
embedded and filtered image would differ significantly from a non-embedded but filtered
image. In other words, both the embedded and non-embedded images are compared
against the common reference of their filtered images. It has been observed that filtering
an image with no message causes changes in the IQMs differently than the changes
brought about on embedded images. This differential behavior is in part because
watermarking or steganographic embedding is not in general a global operation, but is
local in nature. The watermark or message signal is either injected locally, e.g., on a block
basis, or the signal is subjected to a perceptual mask. In any case, we consistently obtained
statistically different quality scores from filtered-and-embedded images and from filteredbut-not-embedded sources. For the hypothesis test, we used various measured quality
scores, which are either due to the difference between an originally non-embedded image
and its filtered version, or due to the difference between embedded image and its filtered
version. In conclusion the selected statistically significant IQMs form a multidimensional
feature space whose points cluster well enough to do a classification of embedded and nonembedded images.

In the design phase of the steganalyzer, we regressed the normalized IQM scores to,
respectively, -1 and 1, depending upon whether an image did not or did contain a message.
Similarly, IQM scores were calculated between the original images and their filtered
versions. In the regression model [54], we expressed each decision label y in a sample of
n observations as a linear function of the IQM scores x s plus a random error, :
y1 = 1 x11 + 2 x12 + ... + q x1q + 1
y2 = 1 x21 + 2 x22 + ... + q x2 q + 2
"
yn = 1 xn1 + 2 xn 2 + ... + q xnq + n

(3.5)

57

In this expression, x ij denotes the IQM score, where the first index indicates the i'th image
and the second one the quality measure. The total number of quality measures considered
is denoted by q . The s denote the regression coefficients. The complete statement of
the standard linear model is

y = X nxq +

rank (X ) = q

E [ ] = 0 .
Cov[ ] = 2 I

such that

(3.6)

The corresponding optimal MMSE linear predictor can be obtained by

= X T X

) (X y ).
1

(3.7)

Once the prediction coefficients are obtained in the training phase, these coefficients can
be used in the testing phase. Given an image in the test phase, first it is filtered and the q
IQM scores are obtained using the image and its filtered version.

Then using the

prediction coefficients, these scores are regressed to the output value. If the output exceeds
the threshold 0 then the decision is that the image is embedded, otherwise the decision is
that the image is not embedded. That is
y = 1 x1 + 2 x 2 + ... + q x q
for y 0

the image contains watermark, and for y < 0

(3.8)

it does not. The schematic

diagram of the steganalyzer is given in Figure 3.3.

3.4. Simulation Results

The watermarking techniques we used were the following: 1) Photoshop plug-in

Digimarc [75], Cox et. al.s technique [77], and the technique from Swiss Federal Institute
of Technology, PGS [76]. One reason for the selection of these techniques was their free
availability on the Internet and the fact that they were all popularly known algorithms. A
more important reason was that with these techniques it was possible to embed watermarks

58

at different strengths, which was instrumental to extract the sensitive IQMs.

The

steganographic tools were Steganos [78], S-Tools [79] and Jsteg [80]. These tools were
among the most cited ones for their pleasing results with respect to steganographic
applications. We used the image database for the simulations. The database contained a
variety of images including computer generated images, images with bright colors, images
with reduced and dark colors, images with textures and fine details like lines and edges,
and well-known images like Lena, peppers etc. We performed eight experiments.

The first three experiments involved watermarking only, namely: 1) The steganalysis
of individual watermarking algorithms, Digimarc, PGS and Cox et. al. for admissible
watermark strengths; 2) The steganalysis of pooled watermarking algorithms at admissible
watermark strengths; 3) In the third experiment the steganalyzer was trained on images
watermarked by Digimarc, and tested on images watermarked by PGS and Cox et. al.

The next three experiments involved steganography only: 4) The steganalysis of

individual steganography algorithms, Steganos, Stools and Jsteg for different embedded
I i
H

1 ()
"
q ()

X i = xi1 # xiq

= X T X

y i

) (X y )
1

(a)

1 ()
"
q ()

x = x1 # x q

y = 1 x1 + ... + q xq

(b)

Figure 3.3. Schematic description of (a) training, and (b) testing.

59

message sizes; 5) The steganalysis of pooled steganography algorithms for different

message sizes; 6) In the sixth experiment the steganalyzer was trained on images
embedded with Steganos and Stools, and tested on images embedded with Jsteg.

The final two experiments involved both watermarking tools and steganography
algorithms. 7) The seventh experiment was the steganalysis of pooled three steganographic
and three watermarking algorithms for admissible levels of watermark strength and for
different message lengths. 8) In the last eighth experiment steganalyzer was trained on
images embedded with Steganos, Stools, watermarked by Digimarc and tested on images
embedded with Jsteg and watermarked by Cox et. al. The aim of the last three experiments
was to see the generalizing ability of the steganalyzer in case an image was to be analyzed
unknown to it in the learning phase. In experiments 1, 2 and 3 the feature set was
which was defined in Section II, for the experiments 4, 5 and 6 the feature set was ,
while the feature set was for the remaining experiments 7 and 8. In the training phase
of every experiment, the feature sets were regressed to 1 and +1.

The organizations of the training and testing samples for the experiments are given in
Tables 3.2.-3.12. The images in the training and test sets are denoted by numbers. More
specifically the training set is = {1, ! ,12} and the test set is = {13, ! ,22}. There were
four levels of watermark strength for Digimarc and PGS. We used the original settings of
Coxs technique; modified the 1000 most significant coefficients in spectral domain. The
embedded message sizes were 1/10 and 1/40 of the cover image size for Steganos and
Stools, while the message sizes were 1/100 of the cover image size for Jsteg.

Table 3.2. Training and test samples for Digimarc and PGS for experiment 1

Training samples
Test samples

Level 1
1,2,3
13,14,15

Level 2
4,5,6
16,17

Level 3
7,8,9
18,19,20

Table 3.3. Training and test samples for Cox for experiment 1

Training samples
Test samples

1000 coefficients
1...12
1322

Level 4
10,11,12
21,22

60

Table 3.4. Training and test samples for pooled watermarking algorithms for experiment 2
(L1: Level 1 etc.)

Levels
Train
Test

L1
1
13

Digimarc
L2
L3
2
3
14
15

L4
4
16

L1
5
17

L2
6
18

PGS
L3
7
19

COX
L4
8
20

9,10,11,12
21,22

Table 3.5. Training and test samples for experiment 3: Train on Digimarc, test on PGS and
COX

Training
WM Levels
Training samples
Testing
WM Levels
Test samples

Digimarc
L1
13
PGS
L1
1315

L2
46
L2
1618

L3
79
COX
L3
19,20

L4
1012

21,22

Table 3.6. Training and test samples for Stools for experiment 4

Message size
Training samples
Test samples

1/40 of image size

16
1317

1/10 of image size

712
1822

Table 3.7. Training and test samples for Jsteg for experiment 4

Message size
1/100 of image size
Training samples
112
Test samples
1322
Table 3.8. Training and test samples for Steganos for experiment 4. (Note: In certain
images the Steganos did not let the messages to be embedded no matter what their size)

Message size
1/40 of image size
Training samples
2,4,8
Test samples
15,17

1/10 of image size

10,11,13
19,20,21

61

Table 3.9. Training and test samples for pooled steganography algorithms for experiment 5

Message size
Training samples
Test samples

Steganos
1/40
2,4
13,15

Stools
1/40
1,3
14,16

1/10
8,10
17,19

1/10
5,6
18,20

Jsteg
1/100
7,9,11,12
21,22

Table 3.10. Training and test samples for experiment 6: train on Steganos and Stools, test
on Jsteg

Training
Msg. Size
Training samples
Testing
Msg. Size
Test samples

Steganos
1/40
2,4,8
Jsteg
1/100
1322

1/10
10,11

Stools
1/40
1,3,5,6

1/10
7,9,12

Table 3.11. Training and test samples for pooled watermarking and steganography
algorithms for experiment 7

Digimarc
PGS
COX
Steganos
Stools
Jsteg
Level or msg size L2
L3 L2 L3 1000 cof 1/40 1/10 1/40 1/10 1/100
Training samples 7
8
9 10 11,12
2
4
1
3
5,6
Test samples
18
19 20 21
22
13
15 14 16
17
Table 3.12. Training and test samples for experiment 8: train on Steganos, Stools and
Digimarc, test on Jsteg and Cox

Training
Level, msg. size
Train samples
Testing
Level, msg. size
Test samples

Digimarc
PGS
COX
Steganos
Stools
L2
L3 L2 L3 1000 cof 1/40 1/10 1/40 1/10
7
9
11 12
2,4 8,10 1,3 5,6
Digimarc
PGS
COX
Steganos
Stools
L2
L3 L2 L3 1000 cof 1/40 1/10 1/40 1/10
1317

Jsteg
1/100
Jsteg
1/100
18..22

62

Table 3.13. Performance of the Steganalyzer for All the Experiments

Experiment

False
Alarm
Rate
1. a. Digimarc
2/10
1. b. PGS
2/10
1. c. Cox
4/10
2. Pooled Watermarking
3/10
3 Train on Digimarc, Test on PGS and Cox
5/10
4. a. Steganos
2/5
4. b. Stools
4/10
4. c. Jsteg
3/10
5 Pooled Steganography
5/10
6 Train on Steganos and Stools, Test on Jsteg
3/10
7 Pooled Watermarking and Steganography
5/10
8 Train on Digimarc, PGS, Steganos, Stools
4/10
Test on Cox and Jsteg

Miss
Rate

Correct
Detection

2/10
1/10
2/10
3/10
2/10
1/5
1/10
3/10
0/10
3/10
1/10
3/10

16/20
17/20
14/20
14/20
13/20
7/10
15/20
14/20
15/20
14/20
14/20
13/20

The performance of the steganalyzer is given in Table 3.13. Simulation results

indicate that the steganalyzer is well enough to do a classification of marked and nonmarked images. The classifier is still able to do a classification when the tested images
come from an embedding technique unknown to it, indicating that it has a generalizing
capability of capturing the intrinsic characteristics of watermarking and steganographic
techniques. As we have noted, the features were regressed to output labels 1 and +1 in
every experiment. The false alarms, especially in experiments 3, 5 and 7, can be fixed to
desired rate by choosing the output regression labels asymmetrically.
3.5. Conclusions

In this section, we have addressed the problem of steganalysis of watermarked and

stegonagraphically marked images. That is, we develop techniques for discriminating
between cover-images and stego-images. Our approach is based on the hypothesis that a
particular message embedding scheme leaves statistical evidence or structure that can be
exploited for detection with the aid of proper selection of image features and multivariate
regression analysis. We showed that the distance between an unmarked image and its

63

filtered version is different than the distance between a marked image and its filtered
version. We used image quality metrics as the feature set to distinguish between marked
and non-marked images. To identify specific quality measures, which provide the best
discriminative power, we used ANOVA technique. We have also pointed out the image
features that should be taken more seriously into account in the design of more successful
watermarking or steganographic techniques to eschew detection.

After selecting an

appropriate feature set, we used multivariate regression techniques to get an optimal

classifier using an image and its filtered version. Simulation results with well known and
commercially available watermarking and steganographic techniques indicate that the
proposed technique is successful in classification of marked and non-marked images.

64

4. LOSSLESS AND NEAR-LOSSLESS IMAGE COMPRESSION

WITH SUCCESSIVE REFINEMENT

4.1. Introduction

Lossless or reversible compression refers to compression approaches in which the

reconstructed data exactly matches the original.

Near-lossless compression denotes

compression methods, which give quantitative guarantees on the nature of the loss that is
introduced. Typically, most of the near-lossless compression techniques proposed in the
literature provide a guarantee that no pixel difference between the original and the
compressed image is above a given value [81]. Near-lossless compression is potentially
useful in remote sensing, medical imaging, space imaging and image archiving
applications, where the huge data size could require lossy compression for efficient storage
or transmission. However, the need to preserve the validity of subsequent image analysis
performed on the data set to derive information of scientific or clinical value puts strict
constraints on the error between compressed image pixel values and their originals. In
such cases, near-lossless compression can be used as it yields significantly higher
compression ratios compared to lossless compression and at the same time, the quantitative
guarantees it provides on the nature of loss introduced by the compression process are
more desirable compared to the uncertainties that are faced when using lossy compression.

Another way to deal with the lossy-lossless dilemma faced in applications such as
medical imaging and remote sensing is to use a successively refinable compression
technique that provides a bitstream that leads to a progressive reconstruction of the image.
The increasingly popular wavelet based image compression techniques, for example,
provide an embedded bit stream from which various levels of rate and distortion can be
obtained. With reversible integer wavelets, one gets a progressive transmission capability
all the way to lossless reconstruction. Hence such techniques have been widely cited for
potential use in applications like tele-radiology where a physician can request portions of
an image at increased quality (including lossless reconstruction) while accepting
unimportant portions at much lower quality, thereby reducing the overall bandwidth

65

required for transmitting an image [82, 83]. Indeed, the new still image compression
standard, JPEG 2000, provides such features in its extended forms [84].

Although reversible integer wavelet based image compression techniques provide

integration of lossless and lossy compression in one single framework, the compression
performance they provide is typically inferior to state-of-the-art non-embedded and DPCM
based techniques like CALIC [85]. In addition, although lossless compression is possible
by receiving the entire bit stream (corresponding to a block or the entire image), the lossy
reconstruction at intermediate stages provides no guarantees on the nature of the distortion
that may be present. Near-lossless compression in such a framework is only possible
either by an appropriate pre-quantization of the wavelet coefficients and lossless
transmission of the resulting bit stream, or by truncation of the bit stream at an appropriate
point followed by transmission of a residual layer to provide the near-lossless bound. Both
these approaches have been shown to provide inferior compression as compared to nearlossless compression in conjunction with DPCM coding [81].

We propose a technique that unifies the above two approaches. The proposed
technique produces a bitstream that results in progressive reconstruction of the image just
like what one can obtain with a reversible wavelet codec. In addition, the proposed
scheme provides near-lossless reconstruction with respect to a given bound after each layer
of the successively refinable bitstream is decoded (note, however that these bounds need to
be pre-decided at compression time and cannot be changed during decompression).
Furthermore, the compression performance provided by the proposed technique is superior
or comparable to the best-known lossless and near-lossless techniques proposed in the
literature [86, 87, 88].

This section is organized as follows:

We review the concepts of successive

refinement, density estimation and the data model in Section 4.2. The compression method
is described in Section 4.3. In Section 4.4. we give experimental results and conlusions are
drawn in Section 4.5.

66

4.2. Problem Formulation

The key problem in lossless compression involves estimating the p.m.f. (probability
mass function) of the current pixel based on previously known pixels (or previously
received information). With this in mind, the problem of successive refinement can then
be viewed as the process of obtaining improved estimates of the p.m.f.s with each pass of
the image. If we also restrict the "support" of the p.m.f. to a given length we then integrate
near-lossless compression and successive refinement with lossless compression in one
single framework.

That is we obtain a bitstream, which gives us near-lossless

reconstruction after each pass in the sense that each pixel is within counts of its original
value.

The length of the interval, = 2 + 1 , in which the pixel is, decreases with

successive passes and in the final pass we have lossless reconstruction, = 1 and = 0 .
In order to design a compression technique with these properties we consider image data
compression as asking the optimal question to determine the exact value or the interval of
the pixel depending on whether we are interested in lossless or near-lossless compression,
respectively. Our aim is to find the minimum description length of every pixel based on
the knowledge we have about its neighbors. We know from the Kraft Inequality that a
code length is just another way to express a probability distribution. Massey [89] observed
that the average number of guesses to determine the value of a random variable is
minimized by a strategy that guesses the possible values of the random variable in
decreasing order of probability. Our strategy is to estimate the probability density of the
current pixel using previous information, and based on this density to determine the
interval of the pixel by questioning the most probable interval where the pixel lies.

In the first pass, we assume that the data in use at a coarse level is stationary and
Gaussian in a small neighborhood and we hence use linear prediction. We fit a Gaussian
density for the current pixel, with the linear prediction value taken as the optimal estimate
of its mean, and linear prediction error as its variance. We divide the support of the current
pixels p.m.f., [0,255], into equal 2 + 1 length intervals, being integer. The intervals
are sorted with respect to their probability mass. If the pixel is found to lie in the interval
with highest probability mass the probability mass outside the interval is zeroed out and the
event 1 is fed to the entropy coder; otherwise the next question is asked whether it lies in
the next highest probability interval. Every time one receives a negative answer, the

67

probability mass within the given interval is zeroed out and the event 0 is fed to the
entropy coder till the right interval is found. At the end of the first pass, we have a "crude"
approximation of the image but the maximum error in reconstructed image e

is when

the midvalue of the interval is selected as the reconstructed pixel value.

In the remaining passes we then refine the p.m.f. for each pixel by narrowing the size
of the interval in which it is now known to lie. The key problem here is how to refine the
p.m.f. of each pixel based on p.m.f.'s of its neighboring pixels. Note that the causal pixels
already have a refined p.m.f. but the non-causal pixels do not. The non-causal pixels now
give us vital information (like the presence of edges and texture patterns), which can be
used to get better estimates of the refined p.m.f.. However care must be taken, as the
available information is less redundant than in the first pass with respect to the current
p.m.f. estimation. That is we know in which interval the current pixel is, we have more
precise information of causal pixels and less precise information of non-causal pixels. We
need to estimate/refine the current p.m.f. within the constraint of its support.

The

refinement of the current p.m.f. should take all these into account. The p.m.f. estimation
method for second and remaining passes, outlined in the next section, which is simply a
causal and non-causal p.m.f. summation over the current p.m.f.s support takes
successfully all the information into account. Once the p.m.f. is estimated/refined for the
current pixel the same strategy, guessing the correct interval or the value depending on
their probability, is applied to constrain the pixel to narrower intervals or to their exact
values.

In the following sub-sections we review some key concepts and results from known
literature and show how we propose to use these in order to develop the proposed
technique for successively refinable lossless and near-lossless image compression.
4.2.1. Successive Refinement

Successive refinement of information consists of first approximating data using a few

bits of information, and then iteratively improving the approximation as more and more
information is supplied. In their paper on successive refinement of information [90] Equitz
and Cover state that rate distortion problem is successively refinable if and only if the

68

individual solutions of the rate distortion problems can be written as Markov chain. Then
they give examples of signals along with distortion measures for which successive
refinement is possible, i.e. if the source is Gaussian and MSE (Mean Square Error) is used
as the distortion measure the source is successively refinable. Massey [89] considered the
problem of guessing the value of a realization of a random variable X by asking the
questions of the form Is X equal to x until the answer is Yes. It is observed that the
average number of guesses is minimized by a guessing strategy that guesses the possible
values of X in decreasing order of probability. In near-lossless compression we are
interested in intervals where the pixel lies rather than in their exact values, so the optimal
strategy for minimizing the average number of guesses is to guess the interval in
decreasing order of probability masses contained in the intervals. In either case, we first
need to construct probability mass estimates in order to use this strategy. In what follows,
we describe probability mass estimation for different passes.
4.2.2. P.m.f. Estimation in the First Pass

The Gaussian Model: Natural images in general do not satisfy Gaussianity or stationarity
assumptions. But at a coarse level, in a reasonable size neighborhood, the statistics can be
assumed not to differ from the above assumptions and the results of Gauss-Markov
property can be used. We use linear prediction in the first pass assuming the data in a
small neighborhood as stationary and Gaussian. We fit a Gaussian density for the current
pixel, with the linear prediction value taken as the optimal estimate of its mean.

We use causal pixel to predict the current pixel via normal linear regression model.
Suppose X i 1 , X i 2 , !, X i N are random variables representing the causal neighbors of the
current pixel X i , shown in Figure 4.1. Let x ( i 1) k , x ( i 2 ) k , ! , x ( i N ) k , k = 1,,K denote
their realizations. We assume a discrete-time scalar-valued random process {X i } that
satisfies the Nth-order linear prediction equation

X i = j (xi j ) + i
N

j =1

(4.1)

69

where { j }j =1 are real-valued linear prediction coefficients of the process, and { i } is a

N

sequence that consists of i.i.d. random variables having a Gaussian density with zero mean
and variance 2 . Optimal MMSE (Minimum MSE) linear prediction for an Nth order
stationary Gauss-Markov process {X i } can be formulated as:
E [X i X i 1 , X i 2 , ! , X i N ] = j (xi j ).
N

(4.2)

j =1

For this standard linear model, according to Gauss-Markov theorem, the minimum
variance linear unbiased estimator = [ 1 ... N ] is the least square solution of (4.2) and is
given by [91, 54]

= XT X

) (X y )
1

(4.3)

where y = [X i 1 , X i 2 ,! , X i K ] denote the K context pixels given in Figure 4.2., while the
data matrix X
X i 11 ! X i 1 N
X= "
#
"

X i K 1 ! X i K N
consists of the prediction neighbors of y . The expected value of X i is given by (4.2) and
an unbiased estimator of prediction error variance 2 can be obtained [54] as

2 =

1
y T y T XT y .
K N 1

(4.4)

Based on the principle that the mean-square prediction for a normal random variable
is its mean value, then the density of X i conditioned on causal neighbors is given by

70

1
exp 2
2
2

f (x i xi 1 , xi 2 ,! , xi N ) =

$ i6
$ i 2

$ i 3
$ i 1

$ i 5

xi j (x i j )

j =1

(4.5)

$ i 4

Figure 4.1. Ordering of the causal prediction neighbors of the current pixel i , N=6.

%
%
%
%
%

"

"

"

"

"

"

"

"

"

$
$

$
$

%
%
%
%

Figure 4.2. The context pixels, denoted by and $ , used in the covariance estimation of
the current pixel . The number of context pixels is K=40.
4.2.3. P.m.f. Estimations in the Second Pass

L2 Norm Minimizing Probability Estimation: In finer quantal resolutions after the first pass
we have to leave aside the gaussian assumption since the image data at finer resolutions
behaves more randomly lacking correlation. We thus assume data is independent and
update the estimate of the current pixel density by using neighboring densities, that is by
minimizing the L2 norm of causal and non-causal densities. This stems from the fact that
in the first pass, most of the time the interval is guessed correctly in one question, leading
to Gaussian distributions which fit well to pixels at low resolution (large ). In later
passes the data becomes more independent of each other as more of the redundancy is
removed after each pass, resulting in decreased correlation. At this stage we can use the
non-causal densities as well, which are densities from the non-causal neighborhood of the
pixel from the previous pass. Several probability mass update methods are presented for
the second and higher passes.

The prediction neighbors used in probability mass

71

estimation in the second and higher passes are given in Figure 4.3. Note that we have the
chance to use the non-causal pixels for prediction.
$ i 3
$ i 1

$ i2

i
i 6

i 7

$ i 4

i 5
i 8

Figure 4.3. Causal, $ , and non-causal, , neighbors of the current pixel , , used for
probability mass estimation in the second and higher passes.
Let pi (x ) denote the probability mass function to be estimated given the causal and
non-causal distributions
constraint

{p }

i j

j =1

(p (x ) p (x))

Minimizing

i j

subject to the

+ ( pi (x )) .

(4.6)

p (x ) = 1 and using Lagrange multipliers we have

i

J ( pi ) =

(p (x ) p (x ))

1 j N

i j

Using the variational derivative with respect to pi (x ) one finds the distribution to be of the
form

pi (x ) =
*

1 N
pi j (x ) .

N j =1

(4.7)

The method has some desirable properties. If the neighboring interval censored
p.m.f.s do not overlap with the current one then they have no negative effect on the
estimate. If there exist some overlapping, then an evidence gathering from the causal and
non-causal neighbors for the indices of the current interval occurs as they give rise to
higher accumulated probabilities for some indices in the interval. Notice this method of
summing neighboring densities gives automatically more importance to more precise
information residing in the causal neighbor p.m.f.'s concentrated in narrower intervals than
to the less precise information in the non-causal ones.

72

Turnbull Probability Estimator: An interval-censored observation of X is of the form

( X L , X R ] with X L < X R , where the actual value of X lies in ( X L < X X R ] . Suppose a
sample of N i.i.d observations ( X Li , X Ri ] , for i = 1,..., N , is given. Define the indicator

variables si = I x s ( X Li , X Ri ] , s = 1,...,255 . With this notation, a self-consistent density

estimator p , adopted from Turnbull [92], is given by

p[ s ] =

si p[ s ]
1

N 1i N li p[l ]

(4.8)

1l 255

Turnbull also gives an iterative algorithm to compute (4.8). A heuristic justification

of this method can be made by multiplying both sides by N. Then the left hand side of
(4.6) is the expected number of individuals in the neighborhood having X = s , while the
right-hand side is the conditional expected number of individuals in the neighborhood
having X = s given the observed interval data.
Note that Turnbulls estimator accumulates the likelihood for the current index s the
same way L2 minimizing estimator does, that is the probability of the indices s for which
i.i.d. intervals overlap is higher than the other indices for which the intervals do not
overlap. This seems intuitively reasonable thing to do when the samples are i.i.d.
Hellinger Norm Minimizing Probability Estimation: The relative entropy D( p q ) is a
measure of distance between two distributions. It is a measure of the inefficiency of
assuming that the distribution is q when the true distribution is p . For example if we
knew true distribution of the random variable, then we could construct a code with average
length H ( p ) .

If instead, we used the code for distribution q , we would need

H ( p ) + D( p q ) bits on the average to describe the random variable. The squared Hellinger
norm between distributions with densities p and q is defined as

H 2 ( p, q ) =

p(x ) q(x ) dx
2

(4.9)

73

Many,

if

not

all,

smooth

function

classes

satisfy

the

equivalence

D( p q ) H 2 ( p, q ). The advantage of H 2 is that it satisfies triangle inequality while D

does not.

However D brings in clean information theoretic identities [93] such as

minimum description length principle, stochastic complexity, etc. Taking advantage of the
equivalence between D and H 2 we can use one for the other in the derivation of the
optimal pi (x ) .
*

When we have a class of candidate densities { p i j : j = 1, ! , N } and want to find the

pi (x ) , which minimizes the inefficiency of assuming the distribution was pi j , we can

*

minimize the total extra bits to obtain the shortest description length on the average:

J ( pi ) =

pi (x ) pi j (x ) dx + pi (x )dx
2

1 j N

(4.10)

where is the Lagrange multiplier. Again finding the variational derivative with respect
to p i (x ) and setting it equal to zero, we get

pi (x ) = T pi j (x )
1 j N

(4.11)

where T is the normalizing constant. In general the relative entropy or Kullback-Leibler

distance has a close connection with more traditional statistical estimation measures such
as L2 norm (MSE) and Hellinger norm when the distributions are bounded away from zero,
and is equivalent to MSE when both p and q are Gaussian distributions with the same
covariance structure [93]. Like Turnbulls method this method is not used in p.m.f.
estimation because it is similar in performance to (4.5) but computationally more
expensive.

74

4.2.4. Multi-hypothesis Testing

We can treat the problem of estimating the interval where the current pixel value is
in, within the framework of multi-hypotheses testing [94]. Let the H 1 ,..., H M denote the M
hypotheses where every hypothesis m is associated with an interval (Lm , Rm ] that has a
length of 2 + 1 . The random variable X has a probability mass under each hypothesis
H = m and denote this probability mass by

p X H (x m ) =

i(Lm , Rm ]

(i)

(4.12)

When each hypothesis has an a priori probability, p m = Pr{H = m}, the cumulative
probability mass of H = m and X = x is then p X H (x m )p m . The a posteriori probability
that H = m conditional on X = x is

pH

(m x ) =

p X H (x m )p m

p (x m)p
M

m =1

X H

(4.13)

The rule for maximizing the probability of being correct, so as to minimize the
number of guesses in finding the correct interval, is to choose that m for which p H

(m x )

is maximized. This is denoted by

H = arg max p m p H X (m x )
m

(4.14)

and known as maximum a posteriori (MAP) rule. For equal probabilities, this becomes
the maximum likelihood (ML) rule where we simply choose the hypothesis with the largest
likelihood

H = arg max p X H (x m ) .
m

(4.15)

75

We use ML rule in the first pass, while MAP rule is used in the following passes
since we have the a priori probability mass from the previous passes.

Defining the

indicator function as

xm = I {x (Lm , Rm ]}

(4.16)

where after hypothesis test the (Lm , Rm ] is the correct interval for the current pixel with
highest probability mass in it, the entropy coder is fed with one or zero, respectively, if the
indicator function is one or zero.

yes

yes

Succesive
Mode ?

no

Lossless
Mode?

Gaussian pmf estimate,

ML test, 1

L2 norm pmf estimate,

MAP test, 2

Gaussian pmf estimate,

ML test, = 1
L2 norm pmf estimate,
MAP test, 3 = 1
Succesive
Mode ?

yes
no
Gaussian pmf estimate,
ML test,

embedded
bitstream

Gaussian pmf estimate,

ML test, 1

L2 norm pmf estimate,

MAP test, 2
embedded
bitstream

Figure 4.4. Schematic description of the overall compression scheme.

4.3. The Compression Method

The schematic description of our algorithm is given in Figure 4.4.

We have

integrated near-lossless compression and successive refinement with lossless compression

76

in one single framework. Lossless or near-lossless compression can be achieved either

with successive refinement or in one step, called the direct method.

In successive mode lossless compression, the support of the p.m.f. is successively

refined and narrowed down to one. In the first pass, for every pixel the p.m.f. is estimated
with (4.5) using linear prediction and multi-hypothesis testing (4.15) to constrain the
support (or length) of the current p.m.f. to 1 .

The details of the p.m.f. support

constraining are given in Figure 4.5.

pmf estimate,
hypothesis test,

bitstream

pmf estimate

hypothesis test

xm 1

(Lm , Rm ]c

yes
hypothesis test

p =0

(Lm , Rm ]

=0

Entropy
Coder

0
bitstream

Figure 4.5 Details of the encoder block used in Figure 4.4. Here is the length of the
interval (Lm , Rm ] .

The probability mass within an interval is zeroed out every time we fail in making
the correct guess and feed an arithmetic encoder with zero until we find the correct
interval. The whole probability mass is within a fixed length interval when we proceed to
the next pixel.

77

Successive near-lossless compression is similarly performed in several passes. The

first pass is the same as the direct near-lossless case, but in the following passes, for every
pixel the p.m.f.s are successively refined and the lengths of the supports for the pixels are
narrowed down to the desired precision. P.m.f.s are estimated with method (4.7) and the
intervals where the pixel lies are determined by MAP test (4.14) as we have the a priori
probability mass for the current pixel from the previous pass.

In successive mode, for both lossless and near-lossless compression, the interval of
the current pixel at second or the following passes can be narrowed in two ways: One way
is to split it into two intervals and to perform binary hypothesis test. The other way is to
split the current interval into more than two equal length intervals and to perform multiple
hypothesis test.

bitstream Entropy
Decoder
z = {0,1}
pmf estimate

hypothesis test

z 1
hypothesis test

no

mid arg p
0

yes
p =0

(Lm , Rm ]

Figure 4. 6. The decoder is a replica of the encoder.

The decoder given in Figure 4.6. is similar to the encoder. Based on the estimated
p.m.f. and decoded sequences of successes and failures, the p.m.f. support for the current
pixel is found. Within the support of the estimated p.m.f. the intervals are sorted with
respect to their probability mass and the correct interval is found using decoded failure and
success events. Recall that the decoded events are the outcomes of the tests whether the

78

current pixel lies in that interval or not. Whenever the decoded event z is not success for
the given interval, this interval is discarded, until the correct interval is tested with success.
The reconstructed value of the current pixel is taken as the midvalue of the successfully
tested interval, which guarantees that the error in reconstructed value is not more than half
of the interval length minus one, that is e

= ( 1) / 2 , where is the length of the

interval.
4.4. Experimental Results

Successive and direct mode lossless compression results, which are obtained with
three passes are given in Table 4.1. The interval lengths of the pixel values are taken 8 in
the first pass, 4 in the second pass and 1 in the final third pass in the successive mode.
One-pass results are also given in the same table. Near-lossless results obtained with one
pass for = 1 , = 3 and = 7 , are given in Tables 4.2., 4.3., and 4.4. respectively.
Table 4.1. Comparison of lossless compression results: proposed method versus CALIC.

Proposed
CALIC [85]
3-pass 1-pass
BARB
4.18
4.21
4.45
LENA
4.07
4.07
4.11
ZELDA
3.81
3.79
3.86
GOLDHILL
4.65
4.69
4.63
BOAT
4.06
4.08
4.15
MANDRILL
5.89
5.96
5.88
PEPPERS
4.41
4.37
4.38
MEAN
4.44
4.45
4.49
=0

Table 4.2. Comparison of 4 different methods of near-lossless compression ( = 1 )

Image

SPIHT [52]

Bpp PSNR

CALIC [85]
e

bpp PSNR

CALIC [95]
e

bpp PSNR

Proposed

Bpp PSNR

Hotel

2.70 76.93 6

2.76 49.95 1

2.70 50.07 1

2.84 49.92 1

Ultrasound

1.60 46.33 8

1.99 52.36 1

1.60 51.76 1

1.70 49.84 1

Caf

3.22 47.09 7

3.30 49.97 1

3.22 50.09 1

3.44 49.85 1

Barbara
Finger

2.94 49.91 1
3.82 47.27 5

3.90 49.89 1

2.65 49.88 1
3.82 49.98 1

3.80 49.89 1

79

Table 4.3. Comparison of 4 different methods of near-lossless compression ( = 3 )

Image

SPIHT [52]

Bpp PSNR

CALIC [85]
e

Bpp PSNR

CALIC [95]
e

bpp PSNR

Proposed

Bpp PSNR

Hotel

1.69 41.77 11

1.74 42.31 3

1.67 42.37 3

1.75 42.30 3

Ultrasound

1.09 41.42 16

1.51 45.04 3

1.09 44.86 3

1.30 42.22 3

Caf

2.19 41.42 16

2.27 42.41 3

2.19 42.49 3

2.31 42.21 3

Barbara

1.92 42.23 3

Finger

2.67 40.71 15

1.61 42.27 3

2.73 42.11 3

2.67 42.18 3

2.63 42.10 3

Table 4.4. Comparison of 4 different methods of near-lossless compression ( = 7 )

Image

SPIHT [52]

Bpp PSNR

CALIC [85]
e

Bpp PSNR

CALIC [95]
e

bpp PSNR

Proposed

Bpp PSNR

Hotel

0.95 37.76 20

0.97 36.50 7

0.95 36.75 7

0.96 36.42 7

Ultrasound

0.72 37.17 25

0.99 37.19 7

0.72 38.55 7

0.78 36.48 7

Caf

1.43 36.49 30

1.50 36.31 7

1.43 36.45 7

1.48 35.80 7

Barbara

1.19 36.21 7

Finger

1.77 35.90 21

1.80 35.43 7

0.91 36.27 7
1.77 35.59 7

1.73 35.43 7

Table 4.5. Comparison of bit/pixel efficiency and peak signal to noise ratio in dB of the
proposed algorithm versus the CALIC [85] algorithm.

=1
=2
=3
=4
=5
=6
=7

bit/pixel gain
0.05
0.07
0.08
0.09
0.10
0.11
0.10

PSNR gain
-0.01
0.01
0.03
0.05
0.07
0.07
0.08

Notice that the lossless compression results in bits per pixel obtained with one pass
and three passes are almost the same, contrary to ones expectation that more than one pass
would yield better performance as the non-causal information is available in the following
passes.

This is because the non-causal information is less precise than the causal

information.

80

4.5. Conclusion

In this section, we have presented a technique that unifies progressive transmission

and near-lossless compression in one single bit stream. The proposed technique produces a
bitstream that results in progressive reconstruction of the image just like what one can
obtain with a reversible wavelet codec. In addition, the proposed scheme provides nearlossless reconstruction with respect to a given bound after each layer of the successively
refinable bitstream is decoded. Furthermore, the compression performance provided by
the proposed technique is superior or comparable to the best-known lossless and nearlossless techniques proposed in the literature [81, 85, 95, 52].

The originality of the method consists in looking at the image data compression as
one of asking the optimal questions to determine the interval in which the current pixel
lies. With successive passes of the image, the length of this interval is gradually narrowed
until it becomes of length one, in which case we have lossless reconstruction. Stopping the
process in any intermediate stage gives near-lossless compression. Although our
experimental results show that the proposed method brings in only modest gains in dB
measure or bit per pixel efficiency, we believe that there is room for improvement. Our
future work will explore different avenues for improving upon the results given in this
paper. For example, we have no mechanism for learning global or repeated patterns in the
image. Context based techniques like CALIC, keep a history of past events in a suitably
quantized form and use these to better model subsequent events.

We believe such

mechanisms when incorporated within our framework will give additional improvements.

The proposed techniques provides a flexible framework and many variations of the
basic method are possible. For example, the quality of reconstruction as defined by the
near-lossless parameter k can be made to vary from region to region or even from pixel to
pixel based on image content or other requirements. Given this fact, different regions in
the image can be refined to different desired precision using HVS properties. To this
effect, flat regions where compression artifact visibility is higher can be refined more
accurately, thus achieving perceptually lossless compression. In addition, it would be
interesting to extend the technique to multispectral images.

81

5. CONCLUSIONS AND FUTURE PERSPECTIVES

Firstly, we have presented collectively a set of image quality measures in their

multispectral version and categorized them.

Statistical investigation of 26 different

measures using a ANOVA analyses has revealed that local phase-magnitude measures (S2
or S5), HVS-filtered L1, L2 norms, edge stability measure are most sensitive to coding,
blur and artifacts, while the mean square error (D1) remains as the measure for additive
noise. One can state that combined spectral phase-magnitude measures and HVS filtered
error norms should be paid more attention in the design of coding algorithms and sensor
evaluation. On the other hand the pixel-difference based measures remain still to be the
measures responsive to distortions and least affected by image variety.

The Kohonen map of the measures has been useful in depicting similar ones, and
identifying the ones that are sensitive possibly to different distortion artifacts in
compressed images. The correlation between various measures has been depicted via
Kohonens Self-Organizing Map. The placement of measures in the two-dimensional map
has been in agreement with ones intuitive grouping.

Future work will address subjective experiments and prediction of subjective image
quality using the above salient measures identified. Another possible avenue is to combine
various fundamental metrics for better performance prediction.

Secondly, we have addressed the problem of steganalysis of watermarked and

steganographically marked images. That is, we developed techniques for discriminating
between cover-images and stego-images. Our approach is based on the hypothesis that a
particular message embedding scheme leaves statistical evidence or structure that can be
exploited for detection with the aid of proper selection of image features and multivariate
regression analysis. We showed that the distance between an unmarked image and its
filtered version is different than the distance between a marked image and its filtered
version. We used image quality metrics as the feature set to distinguish between marked
and non-marked images. To identify specific quality measures, which provide the best
discriminative power, we used ANOVA technique. We have also pointed out the image

82

features that should be taken more seriously into account in the design of more successful
watermarking or steganographic techniques to eschew detection.

After selecting an

appropriate feature set, we used multivariate regression techniques to get an optimal

classifier using an image and its filtered version. Simulation results with well known and
commercially available watermarking and steganographic techniques indicate that the
selected IQMs form a multidimensional feature space whose points cluster well enough to
do a classification of marked and non-marked images. The classifier is still able to do a
classification when the tested images come from an embedding technique unknown to it,
indicating that it has a generalizing capability of capturing the general intrinsic
characteristics of watermarking and steganographic techniques.
Proposed steganalysis technique is its infancy. The measures that are incorporated in
the watermarking and steganographic system designs should be investigated for
steganalysis. We have reduced the problem to classification, so the wealth of tools in
pattern recognition i.e., feature selection methods, nonlinear or neural classifiers can be
used for a better classification performance, since the real world problems are non-linear.
In fact, one of the consequences of this work is that current watermarking and
steganographic techniques are so weak for steganographic applications that they lend
themselves easily to linear solutions.
Thirdly, we have presented a technique that unifies progressive transmission and
near-lossless compression in one single bit stream. The proposed technique produces a
bitstream that results in progressive reconstruction of the image just like what one can
obtain with a reversible wavelet codec. In addition, the proposed scheme provides nearlossless reconstruction with respect to a given bound after each layer of the successively
refinable bitstream is decoded. Furthermore, the compression performance provided by
the proposed technique is superior or comparable to the best-known lossless and nearlossless techniques proposed in the literature.
The originality of the method consists in looking at the image data compression as
one of asking the optimal questions to determine the interval in which the current pixel
lies. With successive passes of the image, the length of this interval is gradually narrowed
until it becomes of length one, in which case we have lossless reconstruction. Stopping the
process in any intermediate stage gives near-lossless compression.

Although our

83

experimental results show that the proposed method brings in only modest gains in dB
measure or bit per pixel efficiency, we believe that there is room for improvement. Our
future work will explore different avenues for improving upon the results given in this
thesis. For example, we have no mechanism for learning global or repeated patterns in the
image. Context based techniques like CALIC, keep a history of past events in a suitably
quantized form and use these to better model subsequent events.

We believe such

mechanisms when incorporated within our framework will give additional improvements.
The proposed techniques provides a flexible framework and many variations of the
basic method are possible. For example, the quality of reconstruction as defined by the
near-lossless parameter k can be made to vary from region to region or even from pixel to
pixel based on image content or other requirements. Given this fact, different regions in
the image can be refined to different desired precision using HVS properties. To this
effect, flat regions where compression artifact visibility is higher can be refined more
accurately, thus achieving perceptually lossless compression. It would be interesting to
extend the technique to multispectral images.
Finally, multidimensional pmf estimation and optimal question approach may result
in several birds with one shot.

84

REFERENCES

1.

Perlmutter, S.M., P.C. Cosman, R.M. Gray, R.A. Olshen, D. Ikeda, C.N. Adams, B.J.
Betts, M.B. Williams, K.O. Perlmutter, J. Li, A. Aiyer, L. Fajardo, R. Birdwell and
B.L. Daniel, Image Quality in Lossy Compressed Digital Mammograms, Signal
Processing, Vol. 59, pp. 189-210, 1997.

2.

Lambrecht, C. B., Ed., Special Issue on Image and Video Quality Metrics, Signal
Processing, Vol. 70, 1998.

3.

Lehmann, T., A. Sovakar, W. Schmitt and R. Repges, A comparison of Similarity

Measures for Digital Subtraction Radiography, Computational Biological Medicine,
Vol. 27, No. 2, pp. 151-167, 1997.

4.

Eskiciolu, A. M., Application of Multidimensional Quality Measures to

Reconstructed Medical Images, Optical Engineering Vol. 35, No. 3, pp. 778-785,
1996.

5.

Eskiciolu, A. M. and P. S. Fisher, Image Quality Measures and Their Performance,

IEEE Transactions on Communications, 43(12), 2959-2965 (1995).

6.

Ridder, H., Minkowsky Metrics as a Combination Rule for Digital Image Coding
Impairments, in Proceedings SPIE 1666: Human Vision, Visual Processing, and
Digital Display III, pp. 17-27 1992.

7.

Watson, A. B. (Ed.), Digital Images and Human Vision, Cambridge, MA, MIT Press,
1993.

8.

Girod, B., Whats Wrong with Mean-squared Error, in A. B. Watson (Ed.), Digital
Images and Human Vision, Chapter 15, Cambridge, MA, MIT Press 1993.

85

9.

Miyahara, M., K. Kotani and V. R. Algazi, Objective Picture Quality Scale (PQS) for
Image Coding, IEEE Transactions On Communications, Vol. 46, No. 9, pp. 12131226, 1998.

10. Nill, N. B. and B. H. Bouzas, Objective Image Quality Measure Derived From
Digital Image Power Spectra, Optical Engineering, Vol. 31, No. 4, pp. 813-825,
1992.
11. Franti, P., Blockwise Distortion Measure for Statistical and Structural Errors in
Digital Images Signal Processing: Image Communication, 13, 89-98 (1998).
12. Winkler, S., A perceptual distortion metric for digital color images. in Proceedings
of the Fifth International Conference on Image Processing, vol. 3, pp. 399-403,
Chicago, Illinois, October 4-7, 1998.
13. Daly, S., The visible differences predictor: An algorithm for the assessment of image
fidelity, in A. B.Watson (Ed.), Digital Images and Human Vision, Cambridge, MA,
MIT Press, pp. 179-205, 1993.
14. Frese, T., C. A. Bouman and J. P. Allebach, Methodology for Designing Image
Similarity Metrics Based on Human Visual System Models, Proceedings of
SPIE/IS&T Conference on Human Vision and Electronic Imaging II, Vol. 3016, pp.
472-483, San Jose, CA, 1997.
15. CCIR, Rec. 500-2, Method for the Subjective Assessment of the Quality of Television
Pictures, 1986.
16. Van Dijk, M. and J. B. Martens, Subjective Quality Assessment of Compressed
Images, Signal Processing, Vol. 58, pp. 235-252, 1997.
17. Rohaly, A.M., P. Corriveau, J. Libert, A. Webster, V. Baroncini, J. Beerends, J.L Blin,
L. Contin, T. Hamada, D. Harrison, A. Hekstra, J. Lubin, Y. Nishida, R. Nishihara, J.
Pearson, A. F. Pessoa, N. Pickford, A. Schertz, M. Visca, A. B. Watson and S.

86

Winkler: "Video Quality Experts Group: Current results and future directions."
Proceedings SPIE Visual Communications and Image Processing, Vol. 4067, Perth,
Australia, June 21-23, 2000.
18. Corriveau, P. and A. Webster, "VQEG Evaluation of Objective Methods of Video
Quality Assessment", SMPTE Journal, Vol. 108, pp. 645-648, 1999.
19. Kanugo, T. and R. M. Haralick, A Methodology for Quantitative Performance
Evaluation of Detection Algorithms, IEEE Transactions on Image Processing, Vol.
4, No. 12, pp. 1667-1673, 1995.
20. Matrik, R., M. Petrou and J. Kittler, Error-Sensitivity Assessment of Vision
Algorithms, IEE Proceedings-Vis. Image Signal Processing, Vol. 145, No. 2, pp.
124-130, 1998.
21. Grim, M. and H. Szu, Video Compression Quality Metrics Correlation with Aided
Target Recognition (ATR) Applications, Journal of Electronic Imaging, Vol. 7, No.
4, pp. 740-745, 1998.
22. Barrett, H. H., Objective Assessment of Image Quality: Effects of Quantum Noise
and Object Variability, Journal of Optical Society of America, Vol. A, No. 7, pp.
1261-1278, 1990.
23. Barrett, H. H., J. L. Denny, R. F. Wagner and K. J. Myers, Objective Assessment of
Image Quality II: Fisher Information, Fourier-Crosstalk, and Figures of Merit for Task
Performance, Journal of Optical Society of America, Vol. A, No. 12, pp. 834-852,
1995.
24. Halford, C.E., K.A. Krapels, R.G. Driggers and E.E. Burroughs, Developing
Operational Performance Metrics Using Image Comparison Metrics and the Concept
of Degradation Space, Optical Engineering, Vol. 38, No. 5, pp. 836-844, 1999.

87

25. Dubuisson, M. P. and A. K. Jain, A Modified Hausdorff Distance for Object

Matching, International Conference on Pattern Recognition, Vol. A, pp. 566-569,
Jerusalem, 1994.
26. International Commission of Illumination (CIE), Recommendations on Uniform Color
Spaces, Color Difference Equations, Psychometric Color Terms, Publication CIE 15
(E.-1.3.1), Supplement No. 2, Bureau Central de la CIE, Vienna, 1971.
27. Jain, A. K., Fundamentals of Digital Image Processing, Prentice Hall, New Jersey,
1989.
28. DiGesu, V., and V. V. Staravoitov, Distance-based Functions for Image
Comparison, Pattern Recognition Letters, Vol. 20, No. 2, pp. 207-213, 1999.
29. Starovoitov, V. V., C. Kse and B. Sankur, Generalized Distance Based Matching of
Nonbinary Images, IEEE International Conference on Image Processing, Chicago,
1998.
30. Juffs, P., E. Beggs and F. Deravi, A Multiresolution Distance Measure for Images,
IEEE Signal Processing Letters, Vol. 5, No. 6, pp.138-140, 1998.
31. Andreutos, D., K. N. Plataniotis and A. N. Venetsanopoulos, Distance Measures for
Color Image Retrieval, IEEE International Conference On Image Processing,
Chicago, 1998.
32. Pratt, W. K., Digital Image Processing, New York, Wiley, 1978.
33. Trahanias, P. E., D. Karakos, A. N. Venetsanopoulos, Directional Processing of
Color Images: Theory and Experimental Results, IEEE Transactions Image
Processing, Vol. 5, No. 6, pp. 868-880, 1996.
34. Zetsche, C., E. Barth and B. Wegmann, The Importance of Intrinsically TwoDimensional Image Features in Biological Vision and Picture Coding, in A. B.

88

Watson (Ed.), Digital Images and Human Vision, Cambridge, MA, MIT Press, pp.
109-138, 1993.
35. Rajan, P. K. and J. M. Davidson, Evaluation of Corner Detection Algorithms,
Proceedings of Twenty-First Southeastern Symposium on System Theory, pp. 29-33,
1989.
36. Canny, J., A Computational Approach to Edge Detection, IEEE Transactions on
Pattern Analysis and Machine Intelligence, Vol. 8, No. 6, pp. 679-698, 1986 .
37. Carevic, D. and T. Caelli, Region Based Coding of Color Images Using KLT,
Graphical Models and Image Processing, Vol. 59, No. 1, pp. 27-38, 1997.
38. Tao, H. and T. Huang, Color Image Edge Detection using Cluster Analysis, IEEE
International Conference On Image Processing, pp. 834-836, California, 1997.
39. Trahanias, P. E. and A. N. Venetsanopoulos, Vector Order Statistics Operators as
Color Edge Detectors, IEEE Transactions on System Man and Cybernetics, Vol. 26,
No. 1, pp. 135-143, 1996.
40. Lipschutz, M. M., Theory and Problems of Differential Geometry, McGraw-Hill Inc.,
1969.
41. McIvor, M. and R. J. Valkenburg, A Comparison of Local Surface Geometry
Estimation Methods, Machine Vision and Applications, Vol. 10, pp.17-26, 1997.
42. Avcba, ., B. Sankur and K. Sayood, New Distortion Measures for Color Image
Vector Quantization, ISAS'98: World Multiconference on Systemics, Cybernetics and
Informatics II, pp. 496-503, July 12-16, Orlando, USA, 1998.
43. Duda, R. O. and P. E. Hart, Pattern Recognition and Scene Analysis, New-York,
Wiley, 1973.

89

44. Popat, K. and R. Picard, Cluster Based Probability Model and Its Application to
Image and Texture Processing, IEEE Transactions on Image Processing, Vol. 6, No.
2, pp.268-284, 1997.
45. Basseville, M., Distance Measures for Signal Processing and Pattern Recognition,
Signal Processing Vol. 18, pp. 349-369, 1989.
46. Nill, N. B., A Visual Model Weighted Cosine Transform for Image Compression and
Quality Assessment, IEEE Transactions on Communications, Vol. 33, No. 6, pp.
551-557, 1985.
47. Avcba, . and B. Sankur, Statistical Analysis of Image Quality Measures, in
Proceedings EUSIPCO'2000: 10th European Signal Processing Conference, pp.
2181-2184, 5-8 September, Tampere, Finland, 2000.
48. Avcba, . and B. Sankur, okbantli Imgelerde Nitelik llerinin Istatistiksel
Davranisi, SIU'99: IEEE 1999 Sinyal leme ve Uygulamalar Kurultay, Bilkent,
Ankara, 1999.
49. Avcba, . and B. Sankur, Kodlama Algoritmalarinin Bilgi Ierigini Koruma
Basarimlari, SIU98: 6. Sinyal Isleme ve Uygulamalari Kurultayi, pp. 248-253,
Kizilcahamam, Ankara, 1998.
50. Frese, T., C. A. Bouman and J. P. Allebach, A Methodology for Designing Image
Similarity Metrics Based on Human Visual System Models, Tech. Rep. TR-ECE 97-2,
Purdue University, West Lafayette, IN, 1997.
51. Wallace, G. K., The JPEG Still Picture Compression Standard, IEEE Transactions
on Consumer Electronics, Vol. 38, No. 1, pp. 18-34, 1992.
52. Said, A. and W. A. Pearlman, A New Fast and Efficient Image Codec Based on Set
Partitioning in Hierarchical Trees, IEEE Transactions Circuits and Systems: Video
Technology, Vol. 6, No. 3, pp. 243-250, 1996.

90

53. Martinez, A.M. and R. Benavente, The AR Face Database, CVC Technical Report No.
24, June 1998.
54. Rencher, A. C., Methods of Multivariate Analysis, John Wiley, New York, 1995.
55. Kohonen, T., Self-Organizing Maps, Springer-Verlag, Heidelberg, 1995.
56. Watson, A. B., DCTune: A Technique for Visual Optimization of DCT Quantization
Matrices for Individual Images, Society for Information Display Digest of Technical
Papers, Vol. 24, pp. 946-949, 1993.
57. Simmons, G. J., The prisoners problem and the subliminal channel, in Proceedings
IEEE Workshop Communications Security CRYPTO83, Santa Barbara, CA, 1983, pp.
5167.
58. Simmons, G. J., The history of subliminal channels, IEEE Journal of Selected Areas
on Communications, vol. 16, pp. 452462, May 1998.
59. Petitcolas, F. A. P., R. J. Anderson and M. G. Kuhn, Information HidingA
Survey, Proceedings of the IEEE, Vol. 87, pp.1062-1078, July 1999.
60. Johnson, N. F. and S. Katzenbeisser, A Survey of steganographic techniques, in S.
Katzenbeisser and F. Petitcolas (Eds.): Information Hiding, pp. 43-78. Artech House,
Norwood, MA, 2000.
61. Dugelay, J-L. and S. Roche, A Survey of current watermarking techniques, in S.
Katzenbeisser and F. Petitcolas (Eds.): Information Hiding, pp. 43-78. Artech House,
Norwood, MA, 2000.
62. Langelaar, G. C., I. Setyawan and R. L. Lagendijk, Watermarking Digital Image and
Video Data, A State of-the-Art Overview, IEEE Signal Processing Magazine, pp. 2046, September 2000.

91

63. Hartung, F. and M. Kutter, Multimedia Watermarking Techniques, Proceedings of

the IEEE, Vol. 87, pp.1079-1107, July 1999.
64. Westfield, A. and A. Pfitzmann, Attacks on Steganographic Systems, in A.
Pfitzmann (Ed.): Information Hiding, LNCS 1768, pp. 61-76, Springer-Verlag Berlin
Heidelberg, 1999.
65. Johnson, N. F. and S. Jajodia, Steganalysis: The investigation of Hidden
Information, IEEE Information Technology Conference, Syracuse, NY, USA, 1-3
September 1998.
66. Johnson, N. F. and S. Jajodia, Steganalysis of Images created using current
steganography software, in D. Aucsmith (Ed.): Information Hiding, LNCS 1525, pp.
32-47. Springer-Verlag Berlin Heidelberg 1998.
67. Fridrich, J., R. Du and M. Long, Steganalysis of LSB Encoding in Color Images,
ICME 2000, New York City, July 31-August 2, New York, USA, 2000.
68. Avcba, ., N. Memon and B. Sankur, Steganalysis Using Image Quality Metrics,
(under review), IEEE Transactions on Image Processing, 2001.
69. Avcba, ., N. Memon and B. Sankur, Automatic Detection of the Presence of Stegosignals and Watermarks in Images, to be presented, ECVP2001: European
Conference on Visual Perception, Antalya, Turkey, 2001.
70. Avcba, ., N. Memon and B. Sankur,

Steganalysis Based On Image Quality

Metrics, to be presented, MMSP2001: IEEE Workshop on Multimedia Signal

Processing, Cannes, France, 2001.
71. Avcba, ., N. Memon and B. Sankur, Staganalysis of Watermarking Techniques
Using Image Quality Metrics, SPIE Conference 4314: Security and Watermarking of
Multimedia Contents III, San Jose, USA, 2001.

92

72. Avcba, ., B. Sankur and N. Memon, mge Nitelik ltlerine Dayanarak

Damgalama Sezimi, SIU2001: IEEE 2001 Sinyal leme ve Uygulamalar Kurultay,
Lefkoa, KKTC, 2001.
73. Avcba, ., B. Sankur and K. Sayood, Statistical Evaluation of Image Quality
Measures, (under review), Journal of Electronic Imaging, 2001.
74. Kutter, M., S. Voloshynovskiy and A. Herrigel, The Watermark Copy Attack, SPIE
Conf. on Security and Watermarking of Multimedia Contents II, pp. 371-380, San
Jose, USA, 2000.
75. PictureMarc, Embed Watermark, v.1.00.45, Copyright 1996, Digimarc Corporation.
76. Kutter, M. and F. Jordan, JK-PGS (Pretty Good Signature). Signal Processing
Laboratory at Swiss Federal Institute of Technology (EPFL), Lausanne, Switzerland,
1998, http://ltswww.epfl.ch/~kutter/watermarking/JK_PGS.html.
77. Cox, J., J. Kilian, T. Leighton and T. Shamoon, Secure Spread Spectrum
Watermarking For Multimedia, IEEE Transactions Image Processing, Vol. 6, pp.
1673-1686, 1997.
78. Steganos II, 2001, http://www.steganos.com/english/steganos/download.htm.
79. Brown, A., S-Tools 4.0, 1996, http://members.tripod.com/steganography/stego/stools4.html.
80. Korejwa, J., Jsteg Shell 2.0, 2001, http://www.tiac.net/users/korejwa/steg.htm.
81. Ansari, R., N. Memon and E. Ceran, Near-lossless Image Compression Techniques,
Journal of Electronic Imaging, Vol. 7, No. 3, pp. 486-494, July 1998.
82. Ebrahimi, T. and M. Kunt, Visual data compression for multimedia applications,
Proceedings of the IEEE, vol. 86, pp.1109-1125, May 1999.

93

83. Egger, O., P. Fleury, T. Ebrahimi, and M. Kunt, High-Performance Compression of

Visual InformationA Tutorial ReviewPart I: Still Pictures, Proceedings of the
IEEE, Vol. 87, No. 6, June 1999.
84. Marcellin, M. W., M. J. Gormish, A. Bilgin and M. P. Boliek, An overview of JPEG2000, Proceedings 2000 Data Compression Conference, Snowbird, Utah, March
2000.
85. Wu, X. and N. Memon, Context-Based, Adaptive, Lossless Image Coding, IEEE
Transactions On Communications, pp. 437-444, April 1997.
86. Avcba, ., N. Memon and B. Sankur, Lossless and Near-Lossless Image
Compression with Successive Refinement, (under review), IEEE Signal Processing
Letters, 2001.
87. Avcba, ., N. Memon, B. Sankur and K. Sayood, Lossless and Near-Lossless
Compression with Successive Refinement, SPIE Conference 4310: Visual
Communications and Image Processing, San Jose, USA, 2001.
88. Avcba, ., B. Sankur and N. Memon, Yitimsiz ve Snrl -Yitimli Kodlamada
Balam Olaslklarnn Kullanm, SIU2001: IEEE 2001 Sinyal leme ve
Uygulamalar Kurultay, Lefkoa, KKTC, 2001.
89. Massey, J. L., Guessing and Entropy, in Proceedings 1994 IEEE International
Symposium on Information Theory, p. 204, Trondheim, Norway, 1994.
90. Equitz, H. R. and T. Cover, Successive Refinement of Information, IEEE
Transactions on Information Theory, Vol. 37, No. 2, pp. 269-275, March 1991.
91. Jayant, N. S. and P. Noll, Digital Coding of Waveforms, Prentice-Hall, 1984.

94

92. Turnbull, B. W., The Empirical Distribution Function with Arbitrary Grouped,
Censored and Truncated Data, Journal Royal Statist. Soc. Ser. B, Vol. 38, pp. 290295, 1976.
93. Barron, A., J. Rissanen and B. Yu, The Minimum Description Length Principle in
Coding and Modeling, IEEE Transactions on Information Theory, Vol. 44, No. 6, pp.
2743-2760, October 1998.
94. Van Trees, H. L., Detection Estimation and Modulation Theory Part I, Wiley, 1968.
95. Wu, X. and P. Bao, L Constrained High-Fidelity Image Compression via Adaptive
Context Modeling, IEEE Transactions on Image Processing, pp.536-542, Apr. 2000.

95

REFERENCES NOT CITED

Arikan, E., An Inequality on Guessing and its Application to Sequential Decoding, IEEE
Transactions On Information Theory, Vol. 42, No. 1, pp. 99-105, January 1991.
Avcba, ., B. Sankur, K. Sayood and N. Memon, Component Ratio Preserving
Compression for Remote Sensing Applications, Proceedings SPIE Conf. 3974:
Image and Video Communications and Processing, pp. 700-708, San Jose, USA,
2000.
Barni, M., F. Bartolini, and A. Piva, Digital Waermarking Of Visual Data: State of the Art
and New Trends, EUSIPCO'2000: 10th European Signal Processing Conference,
5-8 September, Tampere, Finland, 2000.
Barron, A. R. and T. Cover, Minimum Complexity Density estimation, IEEE
Transactions on Information Theory, Vol. 37, No. 4, pp. 1034-1054, July 1991.
Berger, T. and J. D. Gibson, Lossy Source Coding, IEEE Transactions On Information
Theory, Vol. 44, No. 6, pp. 2693-2723, October 1998.
Brocket, P. L., A. Charnes and K. H. Paick, Information Theoretic Approach to Unimodal
Density Estimation, IEEE Transactions on Information Theory, Vol. 41, No. 3, pp.
824-829, January 1991.
Cover, T. M. and J. A. Thomas, Elements of Information Theory, New York, Wiley, 1991.
Derin, H., and P. A. Kely, Discrete-Index Markov Type Random Processes, Proceedings
of the IEEE, Vol. 77. No. 10, pp. 1485-1510, October 1989.
Gubner, J. A., Distributed Estimation and Quantization, IEEE Transactions on
Information Theory, Vol. 39, No. 4, pp. 1456-1459, July 1993.

96

Hoffman, R. L. and A. K. Jain, Segmentation and Classification of Range Images, IEEE

Transactions on Pattern Analysis and Machine Intelligence, Vol. 9, No. 5, pp. 608620, September 1987.
Hoover, An Experimental Comparison of Range Image Segmentation Algorithms,
IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 18, No. 7,
July 1996.
Howard, P. G. and J. S. Vitter, Arithmetic Coding for Data Compression, Proceedings of
the IEEE, Vol. 82, No. 6, pp. 857-865, June 1994.
Rissanen, J. J., A universal data compression system, IEEE Transactions on Information
Theory, Vol. 29, No. 5, pp. 656-664, 1983.
Jayant, N., J. Johnston and R. Safranek, Signal Compression based on models of human
perception, Proceedings of the IEEE, Vol. 81. No. 10., pp. 1385-1422, 1993.
Johnson, N. and S. Jajodia, Exploring Steganography: Seeing the Unseen, IEEE
Computer, pp. 26-34, February 1998.
Karray, L., P. Duhamel and O. Rioul, Image Coding with an L-Infinity Norm and
Confidence Interval Criteria, IEEE Transactions on Image Processing Vol.7, pp.
621-631, 1998.
Kohonen, T., E. Oja, O. Simula, A. Sisa and J. Kangas, Engineering Applications of the
Self-Organizing Map, Proceedings of the IEEE, Vol. 84, No. 10, pp. 1358-1378,
October 1996.
Linde, Y., A. Buzo and R. M. Gray, An Algorithm for Vector Quantizer Design, IEEE
Transactions On Communications, Vol. COM-28, pp. 84-95, 1980.
Lohmann, A. W., D. Mendelovic and G. Shabtay, Significance of Phase and Amplitude in
the Fourier Domain, JOSA, Vol. 14, pp. 29012904, 1997.

97

Memon, N. and X. Wu, Recent Developments in Context-Based Predictive Techniques

for Lossless Image Compression, The Computer Journal, Vol. 40, No. 2/3, pp.
127-136, 1997.
Olives, J. L., B. Lamiscarre, M. Gazalet, Optimization of Electro-Optical System with an
Image Quality Measure, SPIE Vol. 3025, 1997.
Ortega, A. and K. Ramchandran,Rate Distortion Methods for Image and Video
Compression, IEEE Signal Processing Magazine, pp. 23-50, November 1998.
Popat, K. and R. Pickard, Exaggerated Consensus in Lossless Image Compression, in
Proceedings IEEE ICIP-1994, Austin, Texas, November 1994.
Rissanen, J., T. P. Speed and B. Yu, Density Estimation by Stochastic Copmplexity,
IEEE Transactions on Information Theory, Vol. 38, No. 2, pp. 315-323, March 1992.
Sayood, K., Introduction to Data Compression, Morgan Kaufman, Publishers, Inc., 1996.
Schafer, R., G. Heising, A. Smolic, Improving Image Compression- Is It Worth the
Effort?, EUSIPCO'2000: 10th European Signal Processing Conference, 5-8
September, Tampere, Finland, 2000.
Shapiro, J. M., Embedded Image Coding Using Zerotrees of Wavelet Coefficients, IEEE
Transactions on Signal Processing, Vol. 41, No. 12, pp. 3445-3462, December 1993.
Fischer, T. R., J. D. Gibson, and B. Koo, Estimation and noisy source coding, IEEE
Transactions on Acoustics, Speech, and Signal Processing, vol. ASSP-38, pp. 23-34,
Jan. 1990.
Torres, L. and E. Delp, New Trends in Image and Video Compression, EUSIPCO'2000:
10th European Signal Processing Conference, 5-8 September, Tampere, Finland,
2000.

98

Weinberger, M. J., J. J. Rissanen, and R. B. Arps, Applications of Universal Context

Modeling to Lossless Compression of Gray-Scale Images, IEEE Transactions on
Image Processing, Vol. 5, No. 4, pp. 575-586, April 1996.
Weinberger, M. J., G. Seroussi and G. Sapiro, The LOCO-I Lossless Image Compression
Algorithm: Principles and Standardization into JPEG-LS, IEEE Transactions on
Image Processing, Vol. 9, No. 8, pp. 1309-1324, August 2000.