COMBINED STATISTICAL AND MODEL BASED TEXTURE FEATURES FOR

IMPROVED IMAGE CLASSIFICATION

Omar Al-Kadi*
*University of Sussex, Brighton, United Kingdom ([email protected])

Keywords: Image classification, texture analysis,

Bayesian classifier

Abstract
This paper aims to improve the accuracy of texture
classification based on extracting texture features using five
different texture methods and classifying the patterns using
a nave Bayesian classifier. Three statistical-based and two
model-based methods are used to extract texture features
from eight different texture images, then their accuracy is
ranked after using each method individually and in pairs.
The accuracy improved up to 97.01% when model based
Gaussian Markov random field (GMRF) and fractional
Brownian motion (fBm) were used together for
classification as compared to the highest achieved using
each of the five different methods alone; and proved to be
better in classifying as compared to statistical methods.
Also, using GMRF with statistical based methods, such as
Gray level co-occurrence (GLCM) and run-length (RLM)
matrices, improved the overall accuracy to 96.94% and
96.55%; respectively.

1. Introduction

MAGE texture represents the appearance of the surface

and how its elements are distributed. It is considered an
important concept in machine vision, in a sense it assists in
predicting the feeling of the surface (e.g. smoothness,
coarseness etc) from image.
Various texture analysis approaches tend to represent
views of the examined textures form different perspectives,
and due to multi-dimensionality of perceived texture, there
is not an individual method that can be sufficient for all
textures [1]. Therefore, this work is mainly concerned with
texture classification accuracy improvement using textures
features derived from model and statistical based methods.
In the model-based approach, a set of parameters which
are driven from the variation of pixel elements of texture are
used to define an image model. The two models methods
used in this work are Gaussian Markov random field
(GMRF) and fractional Brownian motion (fBm), where the
former sets the conditional probability of the intensity of a
certain pixel depending on the values of the neighbouring

pixels while the latter exploits the self-similarity of texture

at varying scales. For statistical-based methods, first and
second order statistics is derived after analyzing the spatial
distributions of pixel grey level values. Gray level cooccurrence, run-length and autocovariance function methods
were selected for feature extraction.
The obtained texture features by different methods are
used individually and in combination with each other for
classification. A supervised learning approach was adopted
for training and testing the extracted features from samples
of image segments obtained from each image class using a
nave Bayesian classifier.

2.Methodology
2.1 Data set preparation
Different type of texture images ranging from fine to coarse
for the purpose of texture classification were used in this
paper as shown in Fig.1 . Eight different texture images
having size of 256x256 with 8-bit grey levels were selected
from the Brodatz album [2]. Each image which defines a
separate class was divided into size of 32 x 32 image
segments with 50% overlapping. Nearly one third of the
images segments (64 samples) referring to each class was
used for training, while the rest (192 samples) was used for
testing the classifier.
2.2 Texture features extraction
Five different methods 2 model and 3 statistical based
were used to extract different texture features from 2048
image segments samples referring to 8 different texture
classes, as follows:2.2.1 Model-based features methods
Random fields
Based upon the Markovian property, which is simply the
dependence of each pixel in the image on its neighbors only,
a Gaussian Markov random field model (GMRF) for third
order Markov neighbors was used [3]. The 7 GMRF
parameters are estimated using least square error estimation
method.

Fig. 1. Eight different Brodatz texture images showing from up to

bottom and from left to right: herringbone cloth (D16), canvas
(D20), coffee beans (D74), calf leather (D24), fur (D93), quartz
(D98), cheese cloth (D106) and plastic bubbles (D112).

Fractals

The GMRF model is defined by the following formula:

p ( I ij | I kl , ( k , l ) N ij ) =

Fractals are used to describe non-Euclidean structures that

show self-similarity at different scales [4]. There are several
fractal models used to estimate the fractal dimension; the
fBm which is the mean absolute difference of pixel pairs as
a function of scale as shown in (4) was adopted [5].

2
n

I ij l skl ;l
1

l =1

exp

2
2

2
2

(1)
Where the right hand side of (1) represents the probability of
a pixel (i,j) having a specific grey value I ij , given the values
of its neighbors, n is the total number of pixels in the
neighborhood Nij of pixel Iij, which influence its value, l is
the parameter with which a neighbor influences the value of
(i,j), and skl ;l is the value of the pixel at the corresponding
position (see Fig.2) where,
sij ;1 = I i 1, j + I i +1; j

sij ;3 = I i 2, j + I i + 2; j

sij ;5 = I i 1, j 1 + I i +1; j +1

sij ;2 = I i , j 1 + I i ; j +1

sij ;4 = I i , j 2 + I i ; j + 2

sij ;6 = I i 1, j +1 + I i +1; j 1

For an image segment of size M and N the GMRF

parameters and are estimated using least square error
estimation method, as follows:
1 sij ;1 sij ;1 L sij ;1 sij ; n

O
M
M = M

ij s s
n ij ; n ij ;1 L sij ; n sij ; n

sij ;1

I
ij ij M

sij ; n

ij

2 =

ij

l =1

l ij ;l

( M 2)( N 2)

(3)

E ( I ) = K r H

(4)

Where I = |I(x2,y2) I(x1,y1)| is the mean absolute

difference of pixel pairs; r = [(x2 x1) + (y2 y1)] is the
pixel pair distances; H is called the Hurst coefficient; and K
is a constant.
The fractal dimension (FD) can be then estimated by
plotting both sides of (4) on a log-log scale and H will
represent the slope of the curve that is used to estimate the
FD as: FD = 3 H

By operating pixel by pixel, an FD image was generated for

each sample image segment where each pixel has its own
FD value. Then first order statistical features were derived,
which are: mean variance, lacunarity (i.e., variance divided
by mean), skewness and kurtosis.
2.2.2 Statistical-based features methods
Co-occurrence matrices

(2)

I s

Fig. 2. Third order Markov neighbourhood for each

sample image pixel (Iij)

The Grey level co-occurrence matrix (GLCM) Pg (i, j | , d )

represents the joint probability of certain sets of pixels
having certain grey-level values. It calculates how many
times a pixel with grey-level i occurs jointly with another
pixel having a grey value j. By varying the displacement
vector d between each pair of pixels many GLCMs with
different directions can be generated. For each sample
image segment and with distance set to one, four GLCMs

GLCM
Training Testing
set
set
100%
82.81%
100%
100%
100%
100%
100%
88.54%
100%
99.48%
100%
97.40%
100%
97.92%
100%
100%
100%
95.77%

Texture
type
D16
D20
D74
D24
D93
D98
D106
D112
accuracy

Texture feature extraction method

GMRF
RLM
Training Testing Training Testing
set
set
set
set
100%
99.48%
100%
77.08%
100%
98.96%
100%
100%
98.44%
83.85%
98.44%
98.44%
95.31%
91.67%
100%
85.94%
100%
98.44%
98.44%
88.02%
100%
100%
100%
92.19%
100%
99.48%
100%
98.44%
89.06%
91.67%
100%
89.06%
97.85%
99.61%
95.44%
91.15%

fBm
Training Testing
set
set
98.44%
83.33%
100%
80.73%
100%
97.40%
96.88%
76.56%
90.63%
83.85%
93.75%
96.35%
92.19%
76.04%
95.31%
91.67%
95.90%
85.74%

ACF
Training Testing
set
set
7.81%
1.56%
95.31%
86.46%
21.88%
25.52%
21.88%
12.50%
56.25%
54.69%
4.69%
0.52%
45.31%
41.15%
65.63%
45.31%
39.84%
33.46%

TableI: Overall accuracy of classification using each texture feature extraction method individually

having directions (0,45,90 &135) were generated.

Having the GLCM normalized, we can then derived eight
second order statistic features which are also known as
haralick features [6] for each sample, which are: contrast,
correlation, energy, entropy, homogeneity, dissimilarity,
inverse difference momentum, maximum probability.
Run-length matrices
The grey level run-length matrix (RLM) Pr (i, j | ) is defined
as the numbers of runs with pixels of gray level i and run
length j for a given direction [7]. RLMs was generated for
each sample image segment having directions (0,45,90
&135), then the following five statistical features were
derived: short run emphasis, long run emphasis, gray level
non-uniformity, run length non-uniformity and run
percentage.
Autocovariance function
The Autocovariance function (ACF) is the autocorrelation
function after subtracting the mean. It is a way to investigate
non-randomness by looking for replication of certain
patterns in an image. The ACF is defined as:
M x N y

( x, y ) =

( I (i, j ) )( I (i + x, j + y) )
i =1

j =1

( M x)( N y )

(5)

Where I (i, j ) is the grey value of a M x N image, is the

mean of the image before processing and x, y are the amount
of shifts.
After calculating the ACF for each sample image segment
the peaks of the horizontal and vertical margins were fitted
using least squares by an exponential function. Therefore,
each sample is represented by four different parameters,

which are the horizontal and vertical margins values

referring to the ACF and exponential fittings.

2.3 Classification algorithm

The nave Bayesian classifier (nBC) is a simple probabilistic
classifier which assumes attributes are independent. Yet, it
is a robust method with on average has a good classification
accuracy performance, and even with possible presence of
dependent attributes [8]. From Bayes theorem,
Pi (Ci X ) =

P( X Ci ) P(Ci )
P( X )
(6)

Given a data sample X which represent the extracted

texture features vector ( f 1, f 2, f 3 K K fj ) having a probability
density function (PDF) P ( X C i ) , we tend to maximize the
posterior probability P (C i X ) (i.e., assign sample X to the
class Ci that yields the highest probability value).
Where P (C i X ) is the probability of assigning class i given
feature vector X ; and P ( X C i ) is the probability; P (Ci ) is
the probability that class i occurs in all data set; P( X ) is the
probability of occurrence of feature vector X in the data
set.
P (Ci ) and P( X ) can be ignored since we assume that all are
equally probable for all samples. This yields the maximum
of P (C i X ) is equal to the maximum of P ( X C i ) and can be
estimated using maximum likelihood after assuming a
Gaussian PDF [9] as follows:
P ( X Ci ) =

1
(2 ) n / 2 i

exp 12 ( X i )T ( X i )
i

(7)

Combined
methods
GMRF & fBm
GMRF & RLM
GMRF & GLCM
RLM & fBm
GLCM & ACF
fBm & GLCM
GMRF & ACF
GLCM & RLM
RLM & ACF
fBm & ACF

No. of
features
12
27
39
25
36
37
11
52
24
9

Train set
accuracy
99.80%
100%
100%
100%
100%
100%
97.66%
100%
99.61%
96.48%

Test set
accuracy
97.01%
96.94%
96.55%
95.70%
95.05%
94.66%
92.84%
92.12%
89.58%
85.48%

Table II
Overall accuracy of texture feature extraction methods
combined in pairs
Where i and i are the covariance matrix and mean vector
of feature vector X of class Ci; ii and i-1i1 are the
determinant and inverse of the covariance matrix; and
( X i )T is the transpose of ( X i ).

3.Experiment Results and Discussion

Initially, each method was applied individually to each of
the eight different texture images to show which performs
better classification accuracy. The overall classification
accuracies for each of the eight Brodatz images are shown in
Table 1. The GLCM and GMRF achieved the highest
classification rate with 95.77% and 95.44%, while RLM and
fBm scored 91.15% and 85.74%. All methods achieved
relative good accuracy except for the ACF.
The method that achieved the lowest misclassification in all
of the eight texture images was GMRF. The least accuracy
was in texture 3 with 83.85% and that was due to large
structure of the image texture which was beyond the size of
the used third order neighborhood box to capture.
Then the highest classification accuracy which was achieved
by the GLCM method was set as the accuracy improvement
criteria to compare the performance of the next part, where
texture features from different methods are combined
together to investigate if they may assist in increasing
accuracy rate.
It was found that using the two model-based texture features
(GMRF and fBm) together improved the overall accuracy
up to 97.01%. Also, when combining the statistical-based
RLM and GLCM with the GMRF texture features it
increased the overall accuracy to 96.94% and 96.55%;
respectively. RLM with fBm gave nearly the same accuracy
using GLCM alone, while the rest of combinations scored
less than the predefined accuracy improvement criteria.
Table 2 summarizes all classification accuracies and number
of used texture features for all possible paired combinations.

It was also noticed that the GMRF texture features appear in

all paired combinations which improved the overall
accuracy, and with only 12 features, the GMRF and fBm
combination overcame the accuracy achieved when using
the GLCM individually which needs 32 features. Combining
more than two methods with each other did not improve
much the accuracy (e.g. GMRF with fBm and RLM
achieved the highest with 97.07% classification accuracy)
and will increase the time for computation as well. Accuracy
could be further improved if a feature selection method was
used to remove possibly highly correlated features; this
needs to be further investigated.

4. Conclusion
As texture feature extraction methods tend to capture
different image texture characteristics, using different
combinations could assist in improving the classifier
accuracy. Using a nBC, it was shown that combined modelbased texture feature extraction methods (GMRF with fBm)
proved to be better in classifying as compared to statistical
methods. The model-based combined features improved the
overall classification accuracy above the highest achieved
using each of five different methods individually. Moreover,
using GMRF features with statistical methods (RLM and
GLCM) improved the overall accuracy as well.

5. References
[1] M. Tuceryan and A. Jain, The Handbook of pattern
Recognition and Computer Vision, 2 ed: World Scientific
Publishing Co., 1998.
[2] P. Brodatz, A Photographic Album for Artists and
Designers. New York: Dover, 1996.
[3] M. Petrou and P. Gacia Sevilla, Image processing:
Dealing with texture: Wiley, 2006.
[4] B. B. Mandelbrot, Fractal Geometry of Nature. San
Francisco, CA: Freeman, 1982.
[5] C. C. Chen, J. S. Daponte, and M. D. Fox, "Fractal
Feature Analysis and Classification in Medical Imaging,"
IEEE Transactions on Medical Imaging, vol. 8, pp. 133142, 1989.
[6] R. M. Haralick, Shanmuga.K, and I. Dinstein, "Textural
Features for Image Classification," IEEE Transactions on
Systems, Man and Cybernetics, vol. SMC3, pp. 610-621,
1973.
[7] M. M. Galloway, "Texture analysis using gray level run
lengths," Computer Graphics Image Processing, vol. 4, pp.
172-179, 1975.
[8] F. Demichelis, P. Magni, P. Piergiorgi, M. A. Rubin, and
R. Bellazzi, "A hierarchical Naive Bayes Model for
handling sample heterogeneity in classification problems: an
application to tissue microarrays," Bmc Bioinformatics, vol.
7, 2006.
[9] R. C. Gonzales and R. E. Woods, Digital Image
Processing, 2 ed: Prentice Hall, 2002.