Accepted Manuscript

Title: An efficient image compression technique using

Tchebichef bit allocation

Author: Ferda Ernawan Nomani Kabir Kamal Zuhairi Zamli

PII: S0030-4026(17)30899-9
DOI: http://dx.doi.org/doi:10.1016/j.ijleo.2017.08.007
Reference: IJLEO 59470

To appear in:

Received date: 23-9-2016

Revised date: 20-7-2017
Accepted date: 1-8-2017

Please cite this article as: Ferda Ernawan, Nomani Kabir, Kamal Zuhairi
Zamli, An efficient image compression technique using Tchebichef bit allocation,
<![CDATA[Optik - International Journal for Light and Electron Optics]]> (2017),
http://dx.doi.org/10.1016/j.ijleo.2017.08.007

This is a PDF file of an unedited manuscript that has been accepted for publication.
As a service to our customers we are providing this early version of the manuscript.
The manuscript will undergo copyediting, typesetting, and review of the resulting proof
before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.
 An Efficient Image Compression Technique Using
Tchebichef Bit Allocation

t
Ferda Ernawan, Nomani Kabir and Kamal Zuhairi Zamli

ip
IBM Centre of Excellence, Faculty of Computer Systems & Software Engineering,
Universiti Malaysia Pahang, Lebuhraya Tun Razak, 26300 Gambang, Kuantan, Malaysia

cr
us
Abstract

The psychovisual technique has brought about significant improvement in pur-

an
suing image analysis and image reconstruction. The psychovisual threshold can
be utilized to find the optimal bits-budget for image signals. The psychovisual
system is developed based on noticeable distortion of the compressed image
M
from an original image in frequency order. This paper proposes an image com-
pression technique using Tchebichef psychovisual threshold for generating an
optimal bits-budget of image signals. The bits-budget is designed to replace
d

the main role of quantization tables in image compression. The experimental

te

results show that the proposed bits-budget technique can improve the visual
quality of image output by 42% of JPEG compression. The visual image qual-
p

ity of Thcebichef bits allocation produces less artifact effect and distortion of
the image pixels. A set of bits-budgets gives excellent improvement in the image
ce

quality at a low average bit length of Huffman code than the image coding using
quantization tables.
Ac

Keywords: Image Compression, Bit Allocation, Psychovisual

Threshold, Tchebichef moments

∗ FerdaErnawan
Email address: [email protected] or [email protected] (Ferda Ernawan, Nomani
Kabir and Kamal Zuhairi Zamli)

Preprint submitted to International Journal for Light and Electron Optics August 2, 2017

Page 1 of 32
 1. Introduction

Nowdays, bilions of new images are uploaded to the internet everyday [1]

t
and they require large storage to store the image data. Image compression

ip
provides a compact storage and a fast transmission for bandwidth-limited ap-
5 plications of the image data [2]. Over the last two decades, researchers have

cr
developed efficient image compression techniques to improve the compression
rates and the quality of images. JPEG [3] and JPEG2000 [4] are the most

us
popular standard image compression techniques along with many other non-
standard compression algorithms. JPEG compression technique implements
the sequential block coding which is easy to implement on most digital cam-

an
10

eras, providing a compact image size. However, JPEG compression introduces

discontinuity of intensities among adjacent image blocks causing a visual ar-
M
tifact due to inter-block correlations [5]. JPEG2000 is the most recent image
compression standard. JPEG2000 is able to produce higher image quality than
15 the standard JPEG compression. However, this significant contribution does
d

not come without price [6]. JPEG2000 has high computational complexity than
the standard JPEG compression [7]. Moreover, some parts of JPEG2000 are
te

patented and thus they are not found as open source, hindering straightforward
experimentation for performance comparison.
p

20 In this paper, we have chosen JPEG compression for the comparison to

ce

the our proposed Tchebichef bit allocation technique rather than JPEG 2000.
JPEG 2000 is not widely supported in web browsers, and it is not generally
used on the World Wide Web, while the standard JPEG compression has been
Ac

widely adopted as an efficient electronic way of storing and transmitting images

25 [8]. The most important part of the image coding is the amount of frequency
image signals to represent the detail information of the image data. The redun-
dancy of image data usually can be reduced by the quantization tables [9]. The
quantization tables can play a major role of image compression performance.
This paper pays an attention to the block based coding and the role of quan-
30 tization tables in image compression. In particular, in this paper we propose

Page 2 of 32
 the optimum bits budget based on psychovisual thresholds to replace the main
role of quantization tables in image compression. In the past, psychovisual ex-
periments have been conducted for subjective measurement test [10] [11]. The

t
ip
subjective evaluation is laborious, since the results need to be obtained by many
35 observers through repeated viewing sessions [12]. The results from the subjec-

cr
tive quality assessment depend on the test condition of human visual systems.
The image is judged for suitability by visual inspection, however, the differ-

us
ence in observer conditions may give significant influences to the testing results
[13]. Some researchers worked on the psychovisual threshold based on quanti-
40 tative experiment to improve the image compression performance [14],[15],[16].

an
Recently, the psychovisual experiment has become an important issue and a
big challenge in image compression. Crandall et. al.[17] proposed a psycho-
visual image compression technique by discarding psychovisually unnecessary
M
bits from the image pixels. The psychovisual image compression techniques
45 compress pixel data of an image by a fixed compression ratio such that small or
no perceptual loss of the image information occurs [17].
d

Image signals require the certain amount of bits at a time in order to fully
te

store the image data. Redundant frequency image signals not only require
amount of bits, but also make no impact on the quality of image reconstruc-
p

50 tion. A psychovisual threshold can be used to determine the optimal frequency

image signals in image compression. Psychovisual threshold can produce the
ce

optimal quantization tables in image compression [18][19]. However a bit al-

location technique can be used to change the role of quantization tables used
for image compression [20]. The goal of this work is to generate an optimal
Ac

55 bit allocation for frequency image signals based on psychovisual thresholds in

image compression in order to reduce redundancy. The assigning bit allocations
are of a great relevance to frequency image signals in image compression. Since
the psychovisual studies have not yet been fully explored in the bit allocation
of image signals.
60 This work investigates an ideal bits-budget for Tchebichef Moments Trans-
form (TMT) image compression. A bit allocation technique is formulated to

Page 3 of 32
 change the role of the quantization process in TMT image compression. An op-
timal bit allocation is generated based on the principle of psychoacoustic model.
The psychoacoustic model has been widely implemented in audio coding such as

t
ip
65 MPEG-Audio Layer-3 (mp3), Dolby Digital and Advanced Audio Coding [21].
In psychoacoustics, the sound can only be heard at certain or higher sound

cr
pressure levels (SPL) across frequency order [22]. A strong signal makes a weak
signal inaudible and the weak signal is imperceptible when occurs under ideal

us
listening conditions. The slope of masking threshold is steeper towards lower
70 frequencies. In audio coding [23], a strong signal at the local frequency image
signals is identified as the local spectral peaks and the surrounding of the peak

an
frequency (weak signals) as the noise components. Psychoacoustics indicates
that human hearing sensation has a selective sensitivity to different frequencies
[24]. Similar to audio coding, in our technique, the frequency image signals are
M
75 categorized as strong signals (peak) and weak signals. The bits of weak image
signals are reduced by the presence of peak signals without introducing image
distortion on the image quality. Visual masking occurs when visibility of target
d

stimulus is reduced by the presence of mask stimulus [25]. A set of bits-budget

te

is used to reduce bits of the weak image signals without distorting the image
80 quality. A set of bits-budget with the optimized frequency image signal is finally
p

obtained to achieve the compact image representation.

ce

2. Psychovisual threshold on large tchebichef moment transform

Discrete orthonormal Tchebichef Moment is a compact support in the large

Ac

sample image block [26]. The orthonormal Tchebichef moment has its own
85 advantage when used for the large sample data. The orthonormal Tchebichef
moment is applied when the moment order becomes large and it can reduce
the accumulation of numerical errors [26]. The discrete orthonormal Tchebichef
moment has energy compactness properties of a large image block. Tchebichef
Moments for a set {tn (x)} of input values (image intensity values) of size M ×N

Page 4 of 32
90 with an order m + n are given as follows:
M −1 N −1
X X tm (x) tn (y)
Tmn = f (x, y) (1)
ρ(m, M ) ρ(n, N )

t
x=0 y=0

ip
f (x, y) denotes the intensity value of an image at the pixel position x and y.
tn (x) is defined using the following recursive relation:

cr
tn (x) = α1 xtn−1 (x) + α2 tn−1 (x) + α3 tn−2 (x) (2)

us
for n = 2, 3, ..., N − 1 and x = 0, 1, ..., N − 1. Where,
r
2 4n2 − 1
α1 = (3)
n N 2 − n2

an
r
1−N 4n2 − 1
α2 = (4)
n N 2 − n2
M
95 r
r
n−1 2n + 1 N 2 − (n − 1)2
α3 = (5)
n 2n − 3 N 2 − n2
The start values for the above recursion are obtained from the equations as
d

follows:
1
te

t0 (x) = √ (6)
N
s
p

3
t1 (x) = (2x + 1 − N ) (7)
N (N 2 − 1)
ce

The recurrence relation to compute the polynomial value for tn (x) is given by:
r r
N − n 2n + 1
tn (0) = tn−1 (0) (8)
N + n 2n − 1
Ac

100

n(1 + n)
tn (1) = {1 + }tn (0) (9)
1−N

tn (x) = γ 1 tn (x − 1) + γ 2 tn (x − 2) (10)

For n = 1, 2, ..., N − 1and x = 2, 3, ..., N2 − 1 where

−n(n + 1) − (2x − 1)(x − N − 1) − x

γ1 = (11)
x(N − x)

Page 5 of 32
 (x + 1)(x − N − 1)
γ2 = (12)
x(N − x)

t
The above definition uses the following scale factor for the polynomial of degree

ip
105 n: r
N (N 2 − 1)(N 2 − 22 ) . . . (N 2 − n2 )
β(n, N ) = (13)
2n + 1

cr
The set tn (x) has a squared-norm given by
N
X −1
{tn (i)}2 = 1.0

us
ρ(n, N ) = (14)
i=0

The description of squared-norm ρ (·) and the properties of orthonormal Tchebichef

polynomials are explained by Mukundan [26]. The first four M × N TMT with

an
M = N = 256 is shown in Figure 1 for visual purposes. Image reconstruction
M
d
p te
ce
Ac

Figure 1: The first four discrete Tchebichef polynomials tn (x) for x = 0, 1, 2 and 3 .

110 from its moments is made by inverse TMT as follows:

M
X −1 N
X −1
f (x, y) = tmn tm (x)tn (y) (15)
m=0 n=0

Page 6 of 32
 For m, n = 0, 1, 2, ..., N − 1. Where f (x, y) denotes the reconstructed image
intensity and M, N denote the maximum orders of moments. Based on the
discrete orthogonal moments above, a compact representation of the moment

t
ip
ti (x)
coefficient K ∈ R(S×S) considering ki (x) = ρ(i,M ) is given by:
 
k0 (0) k1 (0) ... kS−1 (0)

cr
 
 k0 (1) k1 (1) ... kS−1 (1) 
 
 
K =  k0 (2) k1 (2) ... kS−1 (2) 
 

us
.. .. ..
 
 .. 
 . . . . 
 
k0 (S − 1) k1 (S − 1) . . . kS−1 (S − 1)

an
115 Image block matrix F ∈ R(S×S) with f (x, y) denotes the intensity values of the
image pixels for each color component:
 
M
f (0, 0) f (0, 1) ... f (0, S − 1)
 
 f (1, 0) f (1, 1) ... f (1, S − 1) 
 
 
F =  f (2, 0) f (2, 1) ... f (2, S − 1) 
 
.. .. ..
 
..
d
 
 . . . . 
 
f (S − 1, 0) f (S − 1, 1) ... f (S − 1, S − 1)
te

The matrix T ∈ R(S×S) of moments is defined using K and F as:

p

T = KT F K (16)
ce

This process is repeated for every block in the original image to generate coef-
ficient of discrete orthogonal TMT. The inverse moment is used to reconstruct
Ac

120 the image block from the above moments and is presented as follows:

G = KT K T (17)

where G ∈ R(S×S) denotes the matrix image of the reconstructed intensity

value. The frequency order is obtained by the counts of cross-diagonal elements
in matrix D. Each (S × S) block image is arranged in a linear order of the
moment coefficients. Implementation of moment order D ∈ R(S×S) is provided

Page 7 of 32
125 as:  
d(0,0) d(0,1) ... d(0,S−1)
 
 d(1,0) d(1,1) ... d(1,S−1) 
 

t
 
D =  d(2,0) d(2,1) ... d(2,S−1) 

ip
 
.. .. ..
 
 .. 
 . . . . 
 

cr
d(S−1,0) d(S−1,1) ... d(S−1,S−1)
The order zero d(0,0) on moment coefficient represents the average intensity

us
of an image. The first order on moment coefficients are represented by d(1,0)
and d(0,1) . The second order moment coefficients are denoted by d(2,0) , d(1,1)
and d(0,2) . Psychovisual threshold on M × N TMT is to support practically

an
130 local bit allocation in TMT image compression. The psychovisual threshold
on large TMT is developed based on the contribution of TMT coefficients to
the reconstruction error for each frequency order. The psycho visual thresholds
M
on 256 × 256 TMT coefficients for luminance and chrominance are shown in
Figure 2. The ideal curve of average reconstruction error for a given frequency
d
p te
ce
Ac

Figure 2: Psycho visual threshold on M × N Tchebichef Moments with M = N = 256.

135 order 0 to frequency order 510 for luminance channel is visualized by a solid
curve. For chrominance channel, an ideal curve of average reconstruction error
is depicted by a blue curve. The finer curves above of average reconstruction

Page 8 of 32
 Table 1: polynomial coefficients
i p q

t
0 0.0303 0.0303

ip
1 −0.0000506 −0.0000576
2 0.00001672 0.0000189

cr
3 0.000000014256 0.0000000162
4 −0.00000000024827 −0.0000000002815

us
5 0.000000000000305848 0.00000000000034721

error are the primitive of psychovisual threshold on 256 × 256 TMT. A just

an
noticeable difference of the compressed image from original image known as
140 psychovisual threshold on TMT luminance and chrominance. These curves are
then interpolated by polynomial functions for luminance fM L and chrominance
M
fM R as follows:

fM L (x) = p5 x5 + p4 x4 + p3 x3 + p2 x2 + p1 x + p0 (18)
d
te

fM R (x) = q5 x5 + q4 x4 + q3 x3 + q2 x2 + q1 x + q0 (19)

for x = 0, 1, 2, ..., 510 which represents frequency order under regular 256 × 256
p

145 discrete transform. Coefficient values p and q are given in Table 1. Furthermore,
ce

these thresholds are used as references for the maximum impact of assigning bit
allocation for local frequency blocks on each frequency order. The impact of bit
allocation for each local frequency block is measured by the reconstruction error
Ac

per pixel. The reconstruction error of assigning bit allocation should be less than
150 psycho visual threshold. Next section explains the generation of bit allocation
based on psycho visual threshold for various local frequency block-sizes in image
compression.

Page 9 of 32
 3. Experimentation with Tchebichef Bit Allocation

A color image has a high correlation between the different colors [27]. There-

t
155 fore, an RGB image is converted to a YUV image to reduce the correlation

ip
between the color components. The YUV image can be computed directly from
8 bit RGB as follows:

cr
       
Y 0.299 0.587 0.114 R 0
       
U  = 0.299 0.587 0.114 · G + 128
       

us
       
V 0.299 0.587 0.114 B 128

Inverse transformation from YUV to RGB is carried out by:

an
       
R 1 0 1.4021 Y 0
       
G = 1 −0.34414 −0.71414 · U  − 128
       
       
M
B 1 1.7718 0 V 128

The luminance (Y) and chrominances (U and V) components are very much
160 decor-related between each other. The most energy of the transformed YUV
d

image is contained in the luminance component [28] [29]. Consequently, an

te

efficient compression is achieved by reaching high compression ratios in the

chrominance components without losing the image quality when converted to the
original RGB color image. The chrominance channels contain much redundant
p

165 color information. The bits can be reduced to eliminate the redundancy. Now we
ce

investigate the optimum bits budget for each channel using TMT psychovisual
threshold.
First, an image is subdivided by 256 × 256 image block and it is transformed
Ac

into image signals using Tchebichef transform. The frequency image signals
170 under regular 256 × 256 Tchebichef Moment contain 65535 AC coefficients and
there is only one DC coefficient on each image block. The DC coefficient is
encoded by a 14-bit number since the average minimum value is about 3565
and the maximum average value is about 16384. Nevertheless, AC coefficients
are listed a traversing array with a zigzag pattern. A linear array of frequency
175 of AC coefficients is divided into various local frequency block-sizes. Then we

10

Page 10 of 32
 investigate an optimal amount of bits assigned to each local frequency blocks
with 8, 16 and 32 coefficients. Frequencies of AC coefficients on each local
frequency block are classified as peak frequencies and non-peak frequencies of

t
ip
image signals.
180 Referring to the audio coding, the tonal signal on audio coding produces the

cr
most energy of an audio signal which contains the important information to hu-
man auditory [23]. Thus, we assume that the peak frequency of AC coefficients

us
on each local frequency block provides a significant contribution to the image
quality. The local peak frequency is determined based on the maximum peak
185 of each local frequency block. The amount of peak frequency signals is main-

an
tained as the original peak frequency signals by rounding into integer values.
AC coefficients on the surrounding of peak signal on each local frequency block
are called the non-peak frequency signals. The non-peak frequency signals of
M
AC coefficients are assigned to get an ideal amount of bit allocation as follows:
 
C(x)
B = round 2n (20)
P
d

190 where B is the frequency masking, P is the local peak frequency signal, x is
te

the index of local frequency block, C represents non-peak frequency signals

and n represents the number of bits on local frequency blocks. An optimal bit
p

allocation can be done by increasing bit allocation by one at a time. The effect
of incremental bit allocation is measured by the Average Reconstruction Error
ce

195 (ARE) which can be defined as follows:

M −1 N −1 R−1
1 X X X
ARE(s) = |g (i, j, k) − f (i, j, k)| (21)
Ac

M N R i=0 j=0
k=0

where, the original image of size M × N × R contains the three RGB colors.
The average reconstruction error from incremental bit one at a time is always
set less than psychovisual error threshold for its frequency order. The inverse
of non-peak frequency coding is computed by:
 
B
C(x) = P (22)
2n

11

Page 11 of 32
200 Several experiments are conducted to evaluate the performance of the pro-
posed bits budget in image compression. 80 images including 40 color images
and 40 grayscale images are used to evaluate the proposed TMT bit allocation in

t
ip
image compression. This proposed bits-budget is tested on 40 color images from
a database [? ] and 40 grayscale images from another database [30] to prove

cr
205 the effectiveness of Tchebichef bits-budget in image compression. The sample
Lena and Baboon images selected from 40 color images in order to analyze the

us
quality output of the compressed image are shown Figure 3. For grayscale im-
ages, zelda and women images chosen as samples from 40 grayscale images are
depicted in Figure 4.

an
M
d

Figure 3: Original Lena and Baboon color image, 512 × 512, 24 bits/pixel.
p te
ce
Ac

Figure 4: Original Zelda (left) and Women (right) grayscale image, 512 × 512, 8 bits/pixel.

210 As a part of the evaluation process, the average bit length of Huffman code,
compression ratio and quality on image reconstruction have been chosen to mea-
sure the effectiveness of the TMT bit allocation in image compression. Average
bit length of Huffman code is used to predict that the number of bits on the
compressed image.

12

Page 12 of 32
215 4. Experimental Results and Analysis

A bit-allocation technique is used in image compression to replace the key

t
role of quantization process. The proposed TMT bits-budget in luminance chan-

ip
nel is designed to have greater bit allocation than chrominance channel. The
average reconstruction errors of the proposed TMT bit allocation on each fre-

cr
220 quency order for local blocks of 8, 16 and 32 coefficients are shown in Figure 5,
Figure 6 and Figure 7, respectively.

us
an
M
d

Figure 5: Average reconstruction error of TMT bits-budget on each frequency order for local
te

block of 8 coefficients
p
ce
Ac

Figure 6: Average reconstruction error of TMT bits-budget on each frequency order for local
block of 16 coefficients

It is observed that ARE continues to increase with higher frequency order

to a peak value and then falls with the rising frequency order. Chrominance

13

Page 13 of 32
 t
ip
cr
us
Figure 7: Average reconstruction error of TMT bits-budget on each frequency order for local
block of 32 coefficients

an
has slightly higher error than luminance. The error with low frequency order
225 is about 0.03 which increases rapidly to about 0.5 to 0.6 with higher frequency
M
order. The error then declines with further frequency order. Notice that the
error reduction is higher with increasing Tchebichef coefficients.
Experimental results are provided by the histogram of frequency distribu-
d

tion, probability distribution coefficients, average bit length of Huffman code,

quality of image reconstruction, compression ratio for our test images. In order
te

230

to measure the effectiveness of our TMT bit allocation in image compression,

JPEG compression is used to compare the proposed TMT compression method.
p

The frequency distributions from JPEG compression and TMT compression are
ce

summarized by histograms as depicted in Figure 8. The frequency distribution

235 shows how much frequency image signals are encoded by Huffman coding [31].
The frequency distribution of TMT compression provides the most significant
Ac

zeros for the AC coefficients than that of JPEG compression. In the proposed
TMT bit allocation, the frequency distribution of image signals is divided into
two parts, peak signals and non-peak signals as shown in Figure 9. Since peak
240 frequency of AC coefficients provides the significant contribution to the image
quality, the peak of AC coefficients is considered as the original frequency signals
by rounding into integer values. A set of bits-budgets is addressed to the non-
peak frequency of AC coefficients. From the histograms on the left of Figure 9,

14

Page 14 of 32
 the extreme distributions of the peak AC coefficients are shown in Figure 10.

t
ip
cr
us
Figure 8: Histogram of AC coefficients from JPEG compression (left) and TMT compression
(right) for 40 images.

an
M
Figure 9: Histogram of TMT bit allocation on AC coefficients for 40 images.
d
p te
ce
Ac

Figure 10: Extreme distribution on peak of AC coefficients.

245 Referring to Figure 10, a red curve represents generalized extreme distribu-
tion at each frequency. Thus, these results of generalized extreme distribution
are used to calculate the probability density function as shown in Figure 11.
For non-peak AC coefficients, the normal distribution is applied to estimate the
probability of AC coefficients from assigning TMT bits-budget which simplifies

15

Page 15 of 32
 t
ip
cr
us
Figure 11: Probability density function on peak of AC coefficients.

an
250 the computation of the encoding process. Huffman tree is not built repeatedly
to compute the probability of the frequency distribution when the new input
images come to be processed. The probability density function will cover the
M
zeros on the frequency distribution. The total area under the red graph of the
probability distribution in Figure 11 is equal to one. The probability density
function (pdf) estimates the probability distribution of the frequency image sig-
d
255

nals. Using probability density functions, we do not require to run the encoding
te

process repetitively for new images. The probability density function is utilized
to estimate the probability distribution in Huffman coding rapidly.
p

Huffman coding makes shorter code for frequently occurring symbols in the
AC coefficients. Each symbol is encoded with a variable length code from the
ce

260

Huffman table [32]. The Huffman codes are written in the Huffman tables.
Huffman tables used during the compression process are stored as header infor-
Ac

mation in the compressed image file in order to decode during the decompression
process.
265 The image quality is measured by Absolute Reconstruction Error (ARE),
Means Square Error (MSE), Peak Signal to Noise Ratio (PSNR) and Structural
Similarity Index Metric (SSIM) to validate the effectiveness of TMT bit alloca-
tion technique in image compression. MSE calculates the average of the squares

16

Page 16 of 32
 of the errors defined by:
M −1 N −1 R−1
1 X X X 2
M SE = [(i, j, k) − f (i, j, k)] (23)
M N R i=0 j=0

t
k=0

ip
270 The standard PSNR is calculated to obtain the quality of image reconstruction.
A higher PSNR means that the image reconstruction has more similarity with

cr
the original image [33]. PSNR is defined as:
 
M axi
P SN R = 20log10 √ (24)

us
M SE
where, M axi is the maximum possible pixel value of the image. Another
measurement approach of image quality is Structural Similarity Index Met-

an
275 ric (SSIM), which captures similarity between original image and compressed
image [34] defined as follows:

α β γ
M
SSIM (x, y) = [l(x, y)] · [c(x, y)] · [s(x, y)] (25)

where, α, β, γ > 0, are the parameters to adjust the relative importance of the
three components. A detail description can be found in [34].
d

We make a set of bit allocation on local frequency blocks using 8 coefficients,

te

280 16 coefficients and 32 coefficients on 256 × 256 TMT for luminance and chromi-
nance as listed in Table 2, Table 3 and Table 4, respectively. The average bit
p

lengths of Huffman code after assigning bit allocation on short, medium and
long frequency block respectively are given in Table 5 and Table 6. The average
ce

compression ratios from bit allocation technique for 40 color and 40 grayscale
285 images are listed in Table 7.
Experimental results show that the larger local frequency block-size produces
Ac

the minimum bit size of the compressed output. We observe that the assigned
bits allocation on local frequency block of 32 coefficients produce slightly better
compression ratio than local frequency blocks of 8 and 16 coefficients. Local
290 frequency blocks of 8, 16 and 32 coefficients produce almost similar average
compression ratio scores. Besides, the reconstruction errors from bit allocation
technique show significant improvement which is about 42% of JPEG compres-
sion and 39.5% of TMT image compression.

17

Page 17 of 32
 Table 2: A Set of TMT Bit Allocation for Local Frequency Blocks of 8 Coefficients on M ×
N − T MT
Luminancec Chrominance

t
ip
Block No Freq Order Bit Block No Freq Order Bit
1 1-3 10

cr
2-3 4-6 9
4-15 7-14 8 1 1-3 8

us
16-41 15-25 7 2-3 4-6 7
42-93 26-38 6 4-13 7-13 6
94-173 39-52 5 14-32 14-22 5

an
174-338 53-73 4 33-66 23-32 4
339-645 74-101 3 67-130 33-45 3
646-1288 102-143 2 131-227 46-59 2
M
1289-4478 144-267 1 228-446 60-83 1
4479-8191 268-510 0 447-8191 84-510 0
d

Table 3: A Set of TMT Bit Allocation for Local Frequency Blocks of 16 coefficients on
te

M × N − T MT
Luminancec Chrominance
Block No Freq Order Bit Block No Freq Order Bit
p

1 1-5 10
ce

2-3 6-9 9
4-10 10-17 8 1 1-5 8
11-27 18-28 7 2-3 6-9 7
Ac

28-54 29-41 6 4-8 10-15 6

55-107 42-58 5 9-20 16-24 5
108-198 59-79 4 21-44 25-37 4
199-383 80-110 3 45-77 38-49 3
384-830 111-162 2 78-138 50-65 2
831-2640 163-295 1 139-291 66-96 1
2641-4095 296-510 0 292-4095 97-510 0

18

Page 18 of 32
 t
Table 4: A Set of TMT Bit Allocation for Local Frequency Blocks of 32 Coefficients on

ip
M × N − T MT
Luminancec Chrominance
Block No Freq Order Bit Block No Freq Order Bit

cr
1 1-7 10
2 8-11 9

us
3-6 12-19 8 1 1-7 8
7-14 20-29 7 2 8-11 7

an
15-33 30-45 6 3-6 12-19 6
34-67 46-64 5 7-12 20-27 5
68-117 65-86 4 13-22 28-37 4
M
118-227 87-120 3 23-45 38-53 3
228-512 121-180 2 46-82 54-71 2
513-1641 181-350 1 83-178 72-106 1
d

1642-2047 351-510 0 179-2047 107-510 0

p te
ce

Table 5: Average Bit Length of Huffman Code for 40 Real and 40 Graphic Color Images
Real Graphic
Method ACY ACU ACV ACY ACU ACV
Ac

JPEG Compression 2.868 2.095 2.184 2.965 2.506 2.516

TMT Compression [35][36] 1.767 1.260 1.212 2.040 1.983 2.061
Bit Allocation on short block 2.872 2.407 2.519 2.921 2.668 2.679
Bit Allocation on medium block 2.714 2.488 2.541 2.707 2.592 2.570
Bit Allocation on long block 2.643 2.546 2.580 2.605 2.564 2.542

19

Page 19 of 32
 Table 6: Average Bit Length of Huffman Code for 40 Grayscale Images

t
Method AC

ip
JPEG Compression 2.934
TMT Compression [35] [36] 1.729

cr
Bit Allocation on short block 2.888
Bit Allocation on medium block 2.670

us
Bit Allocation on long block 2.520

an
Table 7: Average Compression Ratio Score for 40 Real Color Images, 40 Graphic Color Images
and 40 Grayscale Images
M
Method Real Graphic Grayscale
JPEG compression 3.324 2.972 2.685
TMT compression [35][36] 5.559 3.893 4.500
d

Bit allocation on short block 3.077 2.902 2.706

te

Bit allocation on medium block 3.099 3.049 2.995

Bit allocation on long block 3.088 3.112 3.173
p
ce

Table 8: Average Quality of Image Reconstruction for 40 Real and 40 Graphic Color Images
Real Graphic
Ac

Method ARE PSNR SSIM ARE PSNR SSIM

JPEG Compression 5.535 31.190 0.956 5.648 31.636 0.957
TMT Compression[35] [36] 5.258 31.372 0.947 4.840 32.274 0.952
Bit Allocation on short block 3.177 37.104 0.978 3.687 37.568 0.975
Bit Allocation on medium block 3.274 36.821 0.977 3.789 37.291 0.974
Bit Allocation on long block 3.226 36.634 0.977 3.794 36.745 0.974

20

Page 20 of 32
 Table 9: Average Quality of Image Reconstruction for 40 Grayscale Images
Method ARE MSE PSNR SSIM

t
JPEG Compression 4.635 45.425 32.380 0.980

ip
TMT Compression [35] [36] 4.474 38.133 32.696 0.971
Bit Allocation on short block 2.983 18.032 37.191 0.986

cr
Bit Allocation on medium block 3.062 18.975 36.931 0.985
Bit Allocation on long block 3.062 19.098 36.713 0.984

us
Table 10: Average Bit Size of Compressed Image from 40 Real Color Images, 40 Graphic
Color Images and 40 Grayscale Images

an
Method Real Graphic Grayscale
JPEG compression 230.997 Kb 258.394 Kb 95.331 Kb
TMT compression [35][36] 138.144 Kb 197.283 Kb 56.888 Kb
M
Bit allocation on short block 249.575 Kb 264.630 Kb 94.593 Kb
Bit allocation on medium block 247.824 Kb 251.851 Kb 85.458 Kb
Bit allocation on long block 248.636 Kb 246.774 Kb 80.672 Kb
d

In Table 8 and Table 9, we notice that TMT bit allocation provides statis-
te

295 tically higher image quality in image reconstruction. The large number of peak
frequency signals leads to improvement of the quality of image output. The
p

experimental results of compressed image size are shown in Table 10.

ce

The visual image outputs from TMT bit allocation on short, medium and
long frequency block of image signals are shown in Figure 12 and Figure 13. In
300 order to observe the visual quality of the output image, the right eye of Lena
Ac

image is zoomed in to 200%. The image output from a set of bit allocation
on image compression produces a rich texture of the original image. The bit
allocation on shorter local frequency blocks produces relatively similar image
output as the original one. Figure 13 presents that the large local frequency
305 blocks of 32 coefficients produce significant texture color pixels than the local
frequency blocks of 8 and 16 coefficients. It implies that the reconstructed
image pixels are relatively too far from the original pixels in Lena image. A

21

Page 21 of 32
 t
ip
cr
Figure 12: Original cropped baboon image (left) and TMT bit allocation on short block

us
(right) with zoom in 200%.

an
M
Figure 13: TMT bit allocation on medium block (left) and TMT bit allocation on long block
(right) with zoom in 200%.
d

large local frequency block produces less quality of the image output than a
te

short local frequency block. The local frequency block of 32 coefficients under
310 regular 256×256 discrete transform provides 2047 peak frequencies of 65535 AC
p

coefficients. The peak frequency contributes 3.12% of moment coefficients. For

ce

medium local block, the peak frequency contributes 6.24%. Otherwise, short
local block has more peak frequencies, which are 12.49%. The number of peak
signals provides significant contribution in the image quality. Compression of
Ac

315 the rate distortion of the proposed scheme is shown in Figure 14. Our proposed
TMT bit allocation exhibits higher PSNR values (bit rates 2-5) than JPEG
compression and TMT compression.
In this part, Baboon, Lena, Women and Zelda images are selected to analyse
and compare the image quality of the proposed TMT bit allocation technique to
320 JPEG compression and TMT compression. The samples color and grayscale im-
ages are zoomed in to 200% as depicted in Figure 15 and Figure 16, respectively.

22

Page 22 of 32
 t
ip
cr
us
Figure 14:

an
Comparison of the rate distortion among the TMT bit allocation, JPEG com-
pression and TMT compression of compressed Lena image.
M
The experimental results show that a set of bit allocations on short local im-
age signals produces better improvement on the quality of image reconstruction
d

than JPEG compression and TMT compression. In addition, the image output
te

325 from a set of TMT bit allocation is almost the same as the original image.
p

5. Conclusion
ce

Digital images require a significant amount of bits to store the image data.
Image data consists of redundant bits in the frequency image signals. The re-
dundant bits do not provide any contribution on the image quality. This work
Ac

330 has presented a new bit allocation technique using Tchebichef image compres-
sion. This technique can be used to replace the critical role of the quantization
process in image compression. Bit allocation technique is based on noticeable
distortion by assigning bits of the image signals. A set of bit allocations on
local frequency signals is considered to reduce the redundant number of bits on
335 frequency image signals. The goal of this study is to achieve an optimal bits-
budget on the frequency image signals. Experimental results reveal that this

23

Page 23 of 32
 technique performs better than the quantization process in image compression.
Moreover, the technique produces less distortion in the image pixels. The pro-
posed Tchebichef bit allocation provides a significant improvement on the image

t
ip
340 quality at relatively low bit rates. The technique shows the improvement of im-
age quality by approximately 42% of JPEG compression and 39.5% of TMT

cr
image compression.

6. Acknowledgments

us
The authors would like to express very special thanks to Universiti Malaysia
345 Pahang, Malaysia for providing financial support of this research project by

an
UMP Research Grant Scheme (RDU170399).

References
M
[1] H. Averbuch-Elor, Y. Wang, Y. Qian, M. Gong, J. Kopf, H. Zhang,
D. Cohen-Or, Distilled collections from textual image queries, Computer
d

350 Graphics Forum 34 (2) (2015) 131–142. doi:10.1111/cgf.12547.

URL http://dx.doi.org/10.1111/cgf.12547
te

[2] L. Va, G. Brunnett, Rate-distortion optimized compression of motion

p

capture data, Computer Graphics Forum 33 (2) (2014) 283–292. doi:

10.1111/cgf.12326.
ce

355 URL http://dx.doi.org/10.1111/cgf.12326

[3] I.-I. Telegraph, T. C. C. (CCITT), Information technology-digital compres-

Ac

sion and coding of continuous-tone still images-requirements and guidelines,

IEC 10918-1 T.81.

[4] M. Boliek, C. Christopoulos, E. Majani, Jpeg 2000 part i final committee

360 draft version 1.0. iso, IEC JTC1/SC29 WT1.

[5] S. Lee, S. J. Park, A new image quality assessment method to

detect and measure strength of blocking artifacts, Signal Process-
ing: Image Communication 27 (1) (2012) 31 – 38. doi:http:

24

Page 24 of 32
 //dx.doi.org/10.1016/j.image.2011.08.002.
365 URL http://www.sciencedirect.com/science/article/pii/
S092359651100097X

t
ip
[6] P. Bagheri Zadeh, A. Sheikh Akbari, T. Buggy, J. Soraghan, Multiresolu-
tion hvs and statistically based image coding scheme, Multimedia Tools and

cr
Applications 49 (2) (2010) 347–370. doi:10.1007/s11042-009-0371-z.
370 URL http://dx.doi.org/10.1007/s11042-009-0371-z

us
[7] K. Liu, Y. Zhou, Y. S. Li, J. F. Ma, A high performance {MQ} encoder
architecture in {JPEG2000}, Integration, the {VLSI} Journal 43 (3) (2010)

an
305 – 317. doi:http://dx.doi.org/10.1016/j.vlsi.2010.01.001.
URL http://www.sciencedirect.com/science/article/pii/
375 S0167926010000027
M
[8] J. M. Pascual, H. M. Mora, A. F. Guill, J. A. Lpez, Adjustable
compression method for still {JPEG} images, Signal Process-
d
ing: Image Communication 32 (2015) 16 – 32. doi:http:
//dx.doi.org/10.1016/j.image.2015.01.004.
te

380 URL http://www.sciencedirect.com/science/article/pii/

S0923596515000053
p

[9] P. Coltelli, G. Faconti, F. Marfori, On the application of quantization and

ce

dithering techniques to history of arts, Computer Graphics Forum 12 (3)

(1993) 351–362. doi:10.1111/1467-8659.1230351.
385 URL http://dx.doi.org/10.1111/1467-8659.1230351
Ac

[10] H. Lohscheller, A subjectively adapted image communication system, IEEE

Transactions on Communications 32 (12) (1984) 1316–1322. doi:10.1109/
TCOM.1984.1096017.

[11] K. Denecker, P. De Neve, S. Van Assche, R. Van de Walle, I. Lemahieu,

390 W. Philips, Psychovisual evaluation of lossy cmyk image compression for
printing applications, Computer Graphics Forum 21 (1) (2002) 5–18. doi:

25

Page 25 of 32
 10.1111/1467-8659.t01-1-00561.
URL http://dx.doi.org/10.1111/1467-8659.t01-1-00561

t
[12] X. Fei, L. Xiao, Y. Sun, Z. Wei, Perceptual image quality as-

ip
395 sessment based on structural similarity and visual masking, Sig-
nal Processing: Image Communication 27 (7) (2012) 772 – 783.

cr
doi:http://dx.doi.org/10.1016/j.image.2012.04.005.
URL http://www.sciencedirect.com/science/article/pii/

us
S092359651200094X

400 [13] K. Gu, G. Zhai, X. Yang, W. Zhang, A new psychovisual paradigm for

an
image quality assessment: from differentiating distortion types to discrimi-
nating quality conditions, Signal, Image and Video Processing 7 (3) (2013)
423–436. doi:10.1007/s11760-013-0445-2.
M
URL http://dx.doi.org/10.1007/s11760-013-0445-2

405 [14] F. Ernawan, N. A. Abu, N. Suryana, An adaptive jpeg image compression

d
using psychovisual model, Advanced Science Letters 20 (1) (2014) 26–31.
doi:doi:10.1166/asl.2014.5255.
te

URL http://www.ingentaconnect.com/content/asp/asl/2014/
00000020/00000001/art00006
p

410 [15] F. Ernawan, N. Abu, N. Suryana, Adaptive tchebichef moment transform

ce

image compression using psychovisual model, Journal of Computer Science

9 (6) (2013) 716–725, cited By 4. doi:10.3844/jcssp.2013.716.725.
URL https://www.scopus.com/inward/record.
Ac

uri?eid=2-s2.0-84880144758&partnerID=40&md5=
415 7014cb632362199c012136dd1275ec3b

[16] G. Zhai, X. Wu, X. Yang, W. Lin, W. Zhang, A psychovisual quality metric

in free-energy principle, IEEE Transactions on Image Processing 21 (1)
(2012) 41–52. doi:10.1109/TIP.2011.2161092.

[17] R. Crandall, R. van der Merwe, D. Mitchell, I. Hendry, B. Serlet, Psycho-

26

Page 26 of 32
420 visual image compression, uS Patent 8,891,894 (Nov. 18 2014).
URL https://www.google.com/patents/US8891894

t
[18] N. A. Abu, F. Ernawan, N. Suryana, A generic psychovisual error threshold

ip
for the quantization table generation on jpeg image compression, in: Sig-
nal Processing and its Applications (CSPA), 2013 IEEE 9th International

cr
425 Colloquium on, 2013, pp. 39–43. doi:10.1109/CSPA.2013.6530010.

[19] F. Ernawan, N. A. Abu, N. Suryana, An optimal tchebichef moment quan-

us
tization using psychovisual threshold for image compression, Advanced
Science Letters 20 (1) (2014) 70–74. doi:doi:10.1166/asl.2014.5316.

an
URL http://www.ingentaconnect.com/content/asp/asl/2014/
430 00000020/00000001/art00015

[20] F. Payan, M. Antonini, An efficient bit allocation for compressing

M
normal meshes with an error-driven quantization, Computer Aided
Geometric Design 22 (5) (2005) 466 – 486, geometry Processing.
d
doi:http://dx.doi.org/10.1016/j.cagd.2005.04.001.
435 URL http://www.sciencedirect.com/science/article/pii/
te

S0167839605000415

[21] K. Brandenburg, C. Faller, J. Herre, J. D. Johnston, W. B. Kleijn, Percep-

p

tual coding of high-quality digital audio, Proceedings of the IEEE 101 (9)
ce

(2013) 1905–1919. doi:10.1109/JPROC.2013.2263371.

440 [22] L. Hong-fu, Z. Cong, L. Rui-fan, Optimization of masking expansion al-

Ac

gorithm in psychoacoustic models, in: Intelligence Information Processing

and Trusted Computing (IPTC), 2011 2nd International Symposium on,
2011, pp. 161–164. doi:10.1109/IPTC.2011.48.

[23] T. Painter, A. Spanias, Perceptual coding of digital audio, Proceedings of

445 the IEEE 88 (4) (2000) 451–515. doi:10.1109/5.842996.

[24] H. Bao, I. M. S. Panahi, Psychoacoustic active noise control based on

delayless subband adaptive filtering, in: 2010 IEEE International Con-

27

Page 27 of 32
 ference on Acoustics, Speech and Signal Processing, 2010, pp. 341–344.
doi:10.1109/ICASSP.2010.5495863.

t
450 [25] V. Adzic, H. Kalva, B. Furht, Exploring visual temporal masking for video

ip
compression, in: 2013 IEEE International Conference on Consumer Elec-
tronics (ICCE), 2013, pp. 590–591. doi:10.1109/ICCE.2013.6487030.

cr
[26] R. Mukundan, Some computational aspects of discrete orthonormal mo-
ments, IEEE Transactions on Image Processing 13 (8) (2004) 1055–1059.

us
455 doi:10.1109/TIP.2004.828430.

[27] P. De Neve, K. Denecker, W. Philips, I. Lemahieu, An advanced color

an
representation for lossy compression of cmyk prepress images, Computer
Graphics Forum 19 (1) (2000) 3–12. doi:10.1111/1467-8659.00384.
URL http://dx.doi.org/10.1111/1467-8659.00384
M
460 [28] F. Douak, R. Benzid, N. Benoudjit, Color image compression algorithm
based on the {DCT} transform combined to an adaptive block scanning,
d

{AEU} - International Journal of Electronics and Communications 65 (1)

(2011) 16 – 26. doi:http://dx.doi.org/10.1016/j.aeue.2010.03.003.
te

URL http://www.sciencedirect.com/science/article/pii/
S1434841110000695
p

465

[29] R. Mantiuk, R. Mantiuk, A. Tomaszewska, W. Heidrich, Color correction

ce

for tone mapping, Computer Graphics Forum 28 (2) (2009) 193–202. doi:
10.1111/j.1467-8659.2009.01358.x.
Ac

URL http://dx.doi.org/10.1111/j.1467-8659.2009.01358.x

470 [30] C. V. Group, Dataset of standard 512x512 grayscale test images (Jun 2003).
URL http://decsai.ugr.es/cvg/cg/base.htm

[31] D. V. Ivanov, Y. P. Kuzmin, Color distribution - a new approach to texture

compression, Computer Graphics Forum 19 (3) (2000) 283–290. doi:10.
1111/1467-8659.00420.
475 URL http://dx.doi.org/10.1111/1467-8659.00420

28

Page 28 of 32
 [32] K. Vidyavathi, R. S. Sabeenian, Motion-compensated video compression
using variable length huffman coding, in: Machine Vision and Image
Processing (MVIP), 2012 International Conference on, 2012, pp. 77–80.

t
ip
doi:10.1109/MVIP.2012.6428765.

480 [33] Y.-K. Chen, F.-C. Cheng, P. Tsai, A gray-level clustering reduction

cr
algorithm with the least {PSNR}, Expert Systems with Applications
38 (8) (2011) 10183 – 10187. doi:http://dx.doi.org/10.1016/j.eswa.

us
2011.02.071.
URL http://www.sciencedirect.com/science/article/pii/
S0957417411002776

an
485

[34] C. Yim, A. C. Bovik, Quality assessment of deblocked images, IEEE Trans-

actions on Image Processing 20 (1) (2011) 88–98. doi:10.1109/TIP.2010.
M
2061859.

[35] H. Rahmalan, N. A. Abu, S. L. Wong, Using tchebichef moment for fast

d
490 and efficient image compression, Pattern Recognition and Image Analysis
20 (4) (2010) 505–512. doi:10.1134/S1054661810040115.
te

URL http://dx.doi.org/10.1134/S1054661810040115

[36] A. Begum, D. Manimegalai, A. Abudhahir, Optimum coefficients of discrete

p

orthogonal tchebichef moment transform to improve the performance of

ce

495 image compression, Malaysian Journal of Computer Science 26 (1) (2013)

060–075.
URL http://mjcs.fsktm.um.edu.my/detail.asp?AID=1347
Ac

29

Page 29 of 32
 Ac
ce
p
original lena croped JPEG compression TMT compression TMT Bit Allocation
te
d

30
M
an
original baboon croped JPEG compression TMT compression TMT Bit Allocation
us
Figure 15: Visual quality of Lena and Baboon color image, 512 × 512, 24 bits/pixel
cr
ip
t

Page 30 of 32
 Ac
ce
p
original Zelda croped JPEG compression TMT compression TMT Bit Allocation
te
d

31
M
an
original Women croped JPEG compression TMT compression TMT Bit Allocation
us
Figure 16: Visual quality of Zelda and Women grayscale image, 512 × 512, 8 bits/pixel
cr
ip
t

Page 31 of 32
 t
ip
cr
us
an
Dr. Ferda Ernawan is currently a Senior Lecturer at the Faculty of Computer
Systems & Software Engineering, Universiti Malaysia Pahang. He received his
Master in Software Engineering and Intelligence and Ph.D in image process-
M
ing from Faculty of Information and Communication Technology, Universiti
Teknikal Malaysia Melaka in 2011 and 2014 respectively. His research interests
include image compression, digital watermarking and steganography.
d
p te
ce
Ac

32

Page 32 of 32