Sie sind auf Seite 1von 9

Texture Classification using Fractal Analysis with Uniform Local Binary Patterns

(ULBP) and Gabor wavelets


Masood Ahmed, Arslan Shaukat, M. Usman Akram, M. Tahir Farooq
Department of Computer Engineering
College of Electrical and Mechanical Engineering
National University of Sciences and Technology (NUST), H-12 Islamabad, Pakistan
masoodjaffar@rocketmail.com, arslanshaukat@ceme.nust.edu.pk, usmakram@gmail.com, etfarooq@gmail.com
AbstractTexture-based image analysis has been an active
research area among research community of computer vision,
machine learning and image processing due its broader scope of
applications. Its applications ranging from image classification,
recognition, segmentation and retrieval of biomedical to industrial
imaging and to satellite imagery. This paper presents classification
of textures by proposing a new descriptor that is based on some
state of art texture descriptors but computationally efficient and
invariant against scale, illumination and orientation. State of art
descriptors include fractal analysis, a simple Uniform Local
binary pattern (ULBP) and Gabor wavelets (GW). Fractal
dimensions and boundary information of each texture image are
obtained using Segmentation based fractal analysis (SFTA). While
visual description and a scale-orientation invariance is achieved
using a ULBP and GW. Feature vectors are obtained separately
from each descriptor on given texture datasets and combined
together. Some of the benchmark texture databases are utilized to
perform classification experiments. While five standard classifiers
are trained and tested using both k-folds cross validation and
percentage split of data samples to get classification accuracy.
Obtained results show that our proposed approach outperforms
all state of art descriptors and their various extensions on the basis
of simplicity of computation, classification accuracy and size of
feature vectors.
KeywordsSFTA; ULBP; Gabor-Wavelets; k-fold cross
validation; normalization.

I.
INTRODUCTION
Textures are important properties to analyse several types of
images. In an image a textured area is characterized by varying
spatial or non-uniform intensity distributions and need for its
analysis comes into existence due to the fact that scenes from
different images do not exhibit smooth intensity variations.
Unlike these images textures are combination of random
intensity distributions, patterns and structures in images,
usually referred to as spatial variations, arrangement of
orientations and changes in size, shape and colors. These may
be broadly classified into several categories on the basis of
structure and human perception. Based on structure there are
natural texture which are observed naturally in objects and
cannot be designed by humans and artificial textures which are
designed by human beings. While tactile textures which are
perceptible by sense of touch and visual textures that refer to
visual perception, are categorized on the basis of human
perception. Textures also have several sub divisions based on
shape, size, colors, degree of randomness and spatial
homogeneity.
Analysis of textures has many potential areas of applications
covering image retrieval, segmentation and classification.
These areas have wide range of applications ranging from
biomedical imaging to industrial, aerial imagery to document
analysis and robotics to security surveillance. However due to
numerous applications textured based classification proved its
superiority over other areas. Industrial applications [1] of

textured images include surface inspection like defects location


and identification on lumber boards [2], on semiconductor
patterned wafers [3-4] and in textile fabrics [5], material testing,
remote sensing, quality control and robotics. In biomedical
imaging texture can be applicable for segmentation of
ultrasound kidney images [6], identification of cancer in
gastrointestinal images [7-8] and to classify several normal
images from abnormal images for several diseases.
Approaches analysis of texture must be quantitative not
qualitative because it has no precise definition to represent a
universal quality due to variety of natural and artificial images.
There are four broad approaches for texture analysis based on
feature vectors [9]. 1) Statistical approach which does not
operate directly on image pixels but on some matrices like
mean, variance, energy and entropy that are derived from
window pixels. Examples are gray level co-occurrence matrix
(GLCM) [10] and Gotlieb co-occurrence [11]. 2) Model based
approach that captures dependencies by assigning analytical
functions to textures. Model not only describes texture but can
also be used to synthesize it. Instances can be Gaussian Markow
Model [12], AR Model [13] and Gibbs Model [14]. 3) Filter
based approach for instance Gabor, Wavelet transform or
Fourier transform refers to decompose a single texture into
several images of specific amplitude, frequency or orientation.
Texture analysis best suits this approach. 4) Structural approach
which is based on structural primitives that are resolution cells
described by image attributes. Local Binary Pattern (LBP) [15]
is an example of this approach, however it can also be observed
as statistical due to placement rules.
In our previous work [16] we analysed and compared three
most common and significant descriptors for classification of
some publicly available datasets of textures. Based on that
experiments we perform classification of textures by exploiting
all approaches except model based techniques. We use
combination of SFTA (statistical), ULBP (structural) and GW
(filter based) descriptors. SFTA performs decomposition of
given texture image into user defined number of binary images
and then four measures are computed from each resulting
binary image as features. LBP gives statistical features in the
form of binary patterns obtained by setting central pixel a
threshold in a neighbourhood of 8 pixels. GW provides filter
based texture features that are based on localization properties
of both space and frequency domains. A scale and orientation
invariance is achieved using Gabor functions. The extracted
features from all three descriptors are then combined together
and classification performance is achieved using different
classifiers. For distribution of samples into test and train
subsets 10-folds cross validation and percentage split
approaches are applied. Five standard texture datasets Brodatz
album [17], UIUC [18], OUTEX [19] and KTH-TIPS [20]
whereas five classifiers that include K nearest neighbour
(KNN), Nave Bayes (NB), support vector Machine (SVM),

random decision forest (RDF) and multilayer perceptron (MLP)


are used.

Figure 1: Flow diagram of texture classification process


Sequence of the paper is as follows. Section II gives
literature survey of the relevant techniques, section III discusses
proposed method and feature extraction processes, while in
section IV we present classifiers, datasets and details of
experimental results and finally section V concludes the paper.
II.
LITERATURE SURVEY
Image segmentation, retrieval, classification and several
other applications using texture have an extensive literature due
to wide range of applications. Among research community most
of the researchers and authors of texture analysis focussed
features extraction which is key for evaluation of texture
classification, image retrieval and segmentation. Manjunath
[21] was among the early researchers who analysed texture for
image retrieval. His feature extraction method was based on
Gabor filters and wavelets. However due to some due to
limitations of computational complexity and capturing edge
discontinuities of Gabor and Wavelets respectively, Mumtaz et
al. [22] proposed dual tree complex wavelets transform to
overcome these limitations.
A framelet transform in
combination with GLCM were also used for retrieval by S.
Sulochana et al. [23] because framelet transform produces more
sub bands in decomposition due high frequency filter banks.
While authors of [24] also performed image retrieval by
combining features of shape and textures using Temuras six
properties [25] and Haralicks GLCM. Segmentation is a base

towards image analysis for higher purposes like scene analysis,


medical diagnoses, visualization and compression. However
texture based segmentation was first introduced for scene
imaging. During late 70s Sklansky [26] used textures and
boundary detection for feature extraction of segmentation while
researchers in [6] performed partitioning of natural scene
images and kidney ultrasound images using texture based
segmentations. From medical images extraction of organ of
interest was done using both texture and shape priors while
features vectors inside and outside regions was based on Gabor
filter bank. M. Ghiasi et al. [27] proposed fast semantic
segmentation based on texture and color for aerial images.
Where semantic segmentation according to authors performs
both segmentation and classification in single step. In [28] a
detailed survey was presented on texture and color
segmentation covering all covering categorisation of published
research in the area, publicly available databases and evaluation
of performance for several techniques. An extensive range of
applications has made texture classification to be most studied
than segmentation and retrieval.
Texture classification was implemented using fractal
analysis of A. F. Costa et al. [29] in which they segmented an
image and then determined several fractal dimensions and other
features for each segment and then classified textures on the
basis of these features. However features for each segmented
part can be increased or reduced and in our experiments we
added one extra feature. The same technique in combination
with wavelet transform was also used by Saraswathi et al. [30]
for classification of normal brain MRI images from abnormal
ones. LBP [15] became more common among researchers of
texture analysis and has made classification more accurate.
However for its improvement several extensions has been
introduced. [31] Added Fourier Histograms to simple LBP to
get rotation invariance while dominant local binary pattern
(DLBP) [32] gave most frequent patterns to reduce feature
vector size. In their work DLBP aims for capturing most recent
patterns of texture surfaces while Gabor features aim to add
some more texture attributes to feature vectors. Z. Guo et al.
[33] introduced a rotation and scale invariant texture
classification using local binary pattern variance (LBPV) while
in [34] introduce covariate shift in LBP. Most recently R. Mehta
and K. Egiazarian [35] proposed a novel dominant rotated local
binary pattern (DRLBP) and achieved promising accuracy
results. However their accuracy results increase with increase
in radius which leads to very large feature vector. They
achieved highest accuracy percentages for radius of 16 which
has very large vector size (about 65536 features ) and cannot be
run by WEKA or RapidMiner tools in order to analyse by
different classifiers. While in our experiments we have very
smaller vector size (about 153 features) with highest accuracy
percentages for about all datasets as compared other methods
and can easily be run by WEKA or RapidMiner studios to test
by multiple classifiers. Some of the researchers made
contribution towards texture classification using Gabor filters,
Wavelet transforms and combination of both techniques.
Authors of [8] and [9] used Gabor features, DS. Zaid et al. [36]
used discrete wavelets while S. Arivazhagan et al. [37] gave an
approach of combined Gabor and Wavelets techniques.
III. TEXTURE ANALYSIS USING PROPOSED TECHNIQUE
LBP with its multiple extensions, Gabor filters and GW are
most common texture descriptors for image classification. We
utilize fractal analysis in combination with both GW and simple
LBP of radius 1 and 8 neighbouring pixels. However further

extensions of LBP can also be used in this combination but


these extensions will increase dimensionality of feature vectors
and make our approach very complex even with the same or
lesser classification performance. Feature extraction procedures
using these descriptors are described below and after extraction
all features are combined for classification performance.
A. Segmentation based Fractal Texture Analysis (SFTA)
Fractal dimension in image processing and analysis estimate
and quantify the texture of objects or shape of images. Using
SFTA based feature extraction algorithm works by
decomposing a given gray scale image I(x, y) into a set of binary
images IB1, IB2 IBn by employing two thresholds binary
decomposition (TTBD) [26]. Then for each binary image mean,
size and standard deviation are computed whereas fractal
dimensions are extracted from borders of regions of binary
images. Binary image IB(x, y) is obtained by selecting a pair of
thresholds from T and perform decomposition as illustrated
mathematically by equation (1) are obtained on the basis of T
number of thresholds. Where T is user defined and set of
thresholds is achieved by utilizing multi-level Otsu algorithm
[38]. TTBD is shown by fig. 2 for an input texture where value
of Tis 8, thus a total of 16 binary images (8 for lower threshold
and 8 for upper one) are achieved.

1 < (, )
(, ) = {
0

(, ) =
{ 0,

1 ()2
2

(, ) = (2 [
where =

1
2

and =

+ 2 ])

(4)

. GF bank contains various

band pass filters of varying bandwidths, resolutions, and


orientations which are controlled by its parameters. Fig. 3
represents magnitudes, real parts and kernels of GF for an input
texture.
Gabor functions form complete and non-orthogonal basis
set of functions from which any given function can be expanded
in terms of these basis functions because this expansion gives a
localized frequency distribution which has been used for texture
analysis and image compression. A localized frequency is
however not suitable for features as it needs fixed window in
space and as results frequency bandwidth is constant on linear
scale.

(1)

where tl and tu represent lower and upper thresholds


respectively selected from T thresholds.
Features in four measurements mean, size, standard
deviation and fractal dimensions are extracted from each binary
image. Fractal dimensions are computed from borders of
regions of IB whereas boundaries (x, y) are extracted using (2).

1,

of complex sinusoid. A Fourier transform of the above (1) is


given by

( , ) 8 [(, )]:
(, ) = 0
(, ) = 1

(a)

(2)

where N8 is set of 8-connected pixels to (x, y). Values of image


(x, y) will give unity value if pixel of corresponding binary
imageIB(x, y) at position (x, y) has unity value and having at
least one neighboring pixel of zero value.
B. Gabor Wavelets
GF are linear filters commonly used by many researchers of
image classification, segmentation and registration and
particularly for texture analysis. Its optimal localization
properties in space and frequency has also highlighted its
significant role in several computer vision problems. It has
similarity with mammalian visual cortex cells due to two
characteristics like band pass nature and directionality. GF has
invariant response towards different scales and orientations like
band pass response of these mammalian cortex cells. A 2D
mathematical expression of GF is given by a product of
frequency of complex sinusoid that forms band pass filter in
frequency domain and Gaussian function.

(, ) =

1
1 2
(2 ) ( 2 (2

2
)+
2

2) (3)

Where standard deviation of Gaussian function controls


bandwidth while center frequency W is controlled by frequency

(b)
Figure 2: (a) Input gray-scale texture image (b) set of binary
images obtained after decomposition of input image using
TTBD algorithm using nt = 8

C. Uniform Local Binary Pattern


LBP is a simple local texture descriptor that is
computationally very efficient and discriminative and has made
significant progress in the areas of machine learning and
computer vision. It gives a binary pattern for a pixel of an input
textured image based on neighboring pixels after setting some
threshold value. LBPs simplicity and invariance property
against scale, rotation or illumination variations in real time
applications make it distinct from other state of art descriptors.
An original LBP descriptor works by default in a neighborhood
of 3X3 pixels where central pixel value is considered as
threshold. However neighboring pixels can be increased or
reduced for various textures datasets and applications. A 3X3
neighborhood contains 8 pixels resulting in 28 labels and
difference of average value above center value and average of
below center value gives contrast measure. Mathematical
expression for LBP is given below as

, = 1
=0 ( ). 2

Figure 3: Magnitudes of Gabor functions for a texture image


(1st row), Real parts of GF (2nd row) and kernel of any two GF
(3rd row)
For an optimal detection and localization of features on
different scales, filters with varying support are required. This
led to wavelets decomposition for which basic wavelet is Gabor
function. Gabor wavelet is a complex planer wave that is
restricted by a 2D Gaussian envelope. The only thing apart from
scale and orientation that makes two GW differ is ratio between
width and wavelength of Gaussian envelope. Each GW has a
certain orientation and wavelength and an input image is
convolved with GW to estimate magnitude of local frequencies
of that wavelength and orientation in image. GW can be
considered as a family of mutually similar functions. Suppose
g(x, y) is a mother GW then this mutually similar filter can be
created by dilation and rotations of mother wavelet.
An input image I(x, y) of size UxV gives a discrete wavelet
transform from its convolution with complex conjugate *of
mutual similar function created from mother wavelet.

(, ) = ( , )
(, ),

(5)

where p, q are scale and orientation values accordingly and s, t


are filter mask size variables. Self-similar function is defined
by equation below
1
1 2
) ( ( 2
2
2

(, ) = (

+ 2 )) 2 (6)

While the generating function through which wavelets are


obtained is defined by (7).
(, ) = (
, ),

(7)

where = ( + )and = ( +
), p=0, 1 P-1, q=0, 1Q-1. P and Q are total number
of scales and orientations respectively while a is scaling
factor.

(8)

1,
where ( ) = {
and P is total number
0,
>
of neighboring pixels, R represents radius, gc and gp are values
of central and neighboring pixels with coordinates (0, 0) and (x
+ Rcos (2 p/P), y Rsin (2 p/P)) respectively. Fig. 4 provides
illustration of LBP neighbors and radii whereas Fig. 5 gives
binary complete pattern and LBP value for radius 1 and 8
neighbors.
ULBP has at most two binary bitwise transitions between 0
and 1. Patterns 00000011 and 01111100 for instance have 1 and
2 transitions respectively are uniform while 100100000 and
11101101 are non-uniform patterns that contain 3 and 4
transitions respectively. Equation (9) gives mathematical
expression for ULBP.

(, ) = |(1 ) (0 )|
+ 1
(9)
=1 |( ) (1 )|

If U 2 in the above equation, LBP is uniform pattern


otherwise not uniform. An LBP for P number of neighborig
pixels has total of P (P-1) + 2 uniform patterns. Fig. 6
illustrates sequential ULBP patterns of (8, 1) neghborhood
where n represents number of ones increasing downward and
r from left to right gives bitwise circular shift. We like
majority of LBP researchers utilize only uniform patterns and
omit non-uniform due to some reasons. First, LBP patterns are
mostly uniform in texture images, about 90% patterns are
uniform in (8, 1) neighborhood while (16, 2) contains 70%
uniform patterns [15]. Secondly these patterns due to statistical
robustness have better recognition results for many applications
and thirdly LBP features dimensionality increases with increase
in neighboring pixels. This dimensionality can be reduced by
considering uniform patterns and discarding rest of features.
However according to authors of [34] non-uniform patterns
bear some more important information of textures and their
omission deteriorates performance thats why they utilized all
patterns. Although they got better performance than many other
researchers at the cost of very high dimensionality of features
with (16, 3) neighbourhood for some datasets. We unlike
DRLBP use a simple ULBP of default (8, 1) neighbourhood and
achieved higher accuracies than other LBP extensions or greater
number of neighbors.

regression etc. It is a Collection of computer vision, pattern


recognition and machine learning algorithms and well suited for
the development of new schemes of these areas. Before
classification results we describe classifiers and texture
databases.

(a)
(b)
(c)
Figure 4: LBP of (a) radius 1 and 8 neighbors (b) radius 2 and
8 neighbors and (c) radius 2 and 16 neighbors

Figure 5: Example of an LBP value for 3X3 neighborhood

Figure 6: A total of 58 Uniform Patterns in 8 neighborhood


IV. EXPERIMENTS AND RESULTS
Textures can be analyzed in natural scenes, biomedical,
facial or satellite images. We however utilized different
publicly available standard texture datasets for experiments.
We have used MATLAB for feature extraction while several
classifiers are used from WEKA toolkit for classification
accuracy. A WEKA contains multiple tools for data preprocessing, visualization, clustering, classification and

A. Datasets
Our experiments are tested on some of standard publicly
available texture datasets. These contain different number of
classes, images and sizes and a wide range of research has
already been conducted using these datasets. All datasets are
briefly discussed below
OUTEX dataset contains a large set of natural scenes and
other surface textures with different variations in rotation,
illuminations and spatial resolutions. There are 16 test suits
having different numbers of textures, sizes and variations in
OUTEX dataset. We have utilized OUTEX -12 and OUTEX 10 in our experiments. A total of 24 classes with 9 different
orientations in each suit, while there are 9120 and 4320 images
in suit-12 and suit-10 respectively.
KTH-TIPS has 10 classes that include texture surfaces of
aluminium foil, sand paper, linen, Styrofoam, brown bread,
orange peel, cracker, corduroy and sponge. Each has 9 scales of
9 different images taken in 3 poses (frontal, rotated 22 o right
and 22o right) and 3 illuminations thus making total of 810
images (109(33) = 810).
Original Brodatz textures are purely gray-scale and
containing 112 images of size 512512 each. Each image
represents a separate class of texture. We broke each image into
16 non-overlapping segments of size 128128 thus making a
total of 1792 textures.
UIUC dataset is also available publicly containing 25
classes and each class has 40 textures. The textures consist of
albedo variations like wood, 3D shapes like gravel and mixture
of both like carpet etc.
B. Classifiers
Classifiers play an important and significant role in the
fields of computer vision and pattern recognition to analyse
performances and produce results for any set of extracted
features. Researchers have proposed a large number of
classifiers for different applications. Some of the common and
most utilized classifiers are described here.
KNN is a simple and non-parametric machine learning
algorithm that belongs to lazy learning classifiers group and
used for both classification and regression. As a classifier it
assigns any object to a class that is most common in its K
number of nearest neighbors. In our experiments KNN
classifier is based on Euclidian distance and value of k is 1.
However a chi-square and other distance functions can also be
used.
Nave Bayes classifier is also a simple learning algorithm
based on Bayesian theorem. It is probabilistic and predictive
classification method having strong independent assumptions
between attributes of data. It calculates posterior probability
from likelihood and prior probabilities of class and predictor
with assumption that effects of value one predictor for a certain
class are independent of other predictors values.
SVM is a supervised algorithm that performs classification
by constructing hyper-plane which categorizes objects. A
hyper-plane minimizes margins between different classes and
the vectors that define it are support vectors. A linear SVM
simple separates classes by constructing a straight line between
objects having different class memberships. However for

complex and non-linear separability it constructs an optimal


hyper-plane by employing an iterative training algorithm that
minimizes error function. SVMs are also used for regression but
mostly for classification. Polynomial Kernel is one of the SVM
kernels used in this proposed work.
Multilayer perceptron (MLP) is a neural network used for
pattern recognition, classification, approximation and
prediction. It is used to solve complex problems where objects
are not linearly separable. An MLP works as feed-forward
neural network which means the flow of data in a direction from
input layer to output layer with having one or more hidden
layers between input and output and each layer is connected to
next one. It is a modification of standard linear perceptron and
also utilizes supervised learning to train network. Total hidden
layers for our experiments are set as sum of total classes and
attributes which is divided by two whereas sigmoid functions
acts as activation function.
Random forest is a tree based classifier that constructs a
multitude of decision trees at for training and gives a class as
mode of classes of individual trees. It can be used both for
classification and regression and efficient towards large
databases. A simple decision tree classifier has poor predictive
performance due to overfitting problem which occurs when too
many parameters are used to make data model during training
or criterion for training the model does not similar to criterion
used to test. Random forest thus reduces or overcomes the
chances of overfitting.
C. Results
Using the above classifiers all sets of extracted features
were distributed into train and test subsets using k-folds cross
validation and a percentage split. A k-folds cross validation

technique breakdowns original samples of features into training


and testing sub samples to train and test respectively the
classifiers. It randomly partitions the data into k equal size
samples and used to evaluate predictive models. Using a 10folds validation is common practice for analysing several
machine vision and pattern recognition problems thus we
follow the same technique. While in percentage split a
particular percentage of images is utilized to train classifiers
whereas remaining images represent test samples. In our
experiments total images of each database are equally
distributed into two sets. One set containing 50% images gives
training samples and other 50% are used as test set.
We obtained results on given datasets form some of recent
and most frequent descriptors or their extensions and compared
them with the proposed technique. Figure 7 shows results of
each descriptor for all datasets exploiting only KNN classifier.
The proposed descriptor gives maximum accuracy percentage
for four out of five datasets. The accuracy for Outex-12 is only
64.53% and minimum where Brodatz, KTH-TIPS, UIUC and
Outex-10 give 96.87%, 99.01%, 99.7% and 99.95%
respectively which are maximum among all. However DRLBP
has also better results and even maximum accuracy of 97.15%
for Outex-12 for which proposed method resulted with least
accuracy of 64.53%. Classification performance is also
compared using 10-folds validation and 50% split of data
samples as shown by figure 8, about all datasets have same
accuracy using 10-folds and percentage split whereas a minute
difference appears for some datasets. While table 1 contains
details of comparative classification between our proposed
descriptor and different texture classification techniques
discussed in literature review section.

Figure 7: Performance comparison on different datasets using KNN

Figure 8: Classification performance comparison using 10-folds and percentage split of data samples
Table 1: Comparison of Proposed Descriptor performance with different techniques in literature
Authors

Techniques

Farhan Riaz et. Al [7]

Gabor filters

Auto Correlation Gabor


Features (AGF)
H. Permuter et. Al [12]
GMM
Farhan Riaz et. Al [8]

AF Costa et. Al [29]

Fractal Analysis (SFTA)

D.Saraswathi et. Al [30]

WSFTA
Local Binary Pattern
Histogram Fourier
(LBP-HF)

T. Ahonen et. Al [31]

S. Liao et. Al [32]

Dominant Local Binary


Patterns (DLBP)

Z. Guo et. Al [33]

LBP Variance (LBPV)

A. Hassan et. Al [34]

Covariate Shift in ULBP

R. Mehta et. Al [35]

Dominant Rotated LBP


(DRLBP)

S. Arivazhagan et. Al [37]

Gabor Wavelets

Poposed

SFTA with GW andULBP

Year

Classifiers

Databases
Max Accuracy
Brodatz
98.00%
2013
SVM
USC-SIPI
70.75%
Chromoendoscopy (CHI)
82.30%
2012
SVM
Narrow Band Imaging (NBI)
88.40%
2005 Max Liklihood
VisTex
91.20%
Lungs CT RIOs
88.83%
2012
SVM
KTH-TIPS
95.15%
Textures Surfaces
93.11%
2014 Neural Network
Brain MRI
98.00%
Outex-12
92.50%
2009
KNN
KTH-TIPS
54.60%
CMU-PIE
74.80%
Some Brodatz Textures
99.40%
Outex-10
98.10%
2009
SVM
CUReT
92.77%
Meastex
85.82%
Outex-10
97.76%
2009
KNN
Outex-12
95.57%
Brodatz
99.00%
2014
SVM
USC-SIPI
87.00%
Outex-10
99.19%
2016
KNN
Outex-12
97.15%
KTH-TIPS
96.78%
Minimum
Basic Brodatz Textures
95.60%
2006
Distance
Brodatz
93.11%
Outex-12
85.48%
Outex-10
99.95%
KNN, MLP, NB,
2016
Brodatz
97.09%
RF and SVM
KTH-TIPS
99.38%
UIUC
99.70%

Table 2 enlists the accuracy percentage against each dataset


using multiple-classifiers where all classifiers have better
results but Nave Bayes classifier appears with least percentage
of accuracy for about all datasets while MLP outperforms all
the classifiers.
V. CONCLUSIONS
This paper has proposed an efficient and fast technique for
classification of textures from combination of three state of art
descriptors. We have addressed problems of rotation and
solution invariance and specifically dimensionality of features.
Features were extracted using ULBP, fractal dimensions and
GW which represent structural, statistical and filter based
approaches respectively. Extensive experiments have been
performed on five standard texture databases which include

images like non-overlapping parts of same textures and have


variations in scale, resolution and orientation. Obtained features
are trained and tested using 10 folds cross validation and
percentage split for five classifiers. Our experimental results
demonstrates that this proposed technique out performs all the
state of art methodologies on the basis of reduction in
dimensionality of feature vectors and accuracy percentage.
While testing with multiple classifiers majority of the features
gave same and very high performances but out of all datasets
Outex-12 could not have better performance which means that
there is still some problems regarding features extraction and
selection to be addressed to get more improvements in accuracy
of both datasets. In order to broaden the range of applications
in future the same combination of features for other industrial
and biomedical textures can be used.

Table 1: Accuracy Percentage for all Datasets using Multiple Classifiers with 10-Folds Cross Validation
Datasets
KTH-TIPS
Brodatz
OUTEX-12
UIUC
OUTEX-10

Nave Bayes
89.75
96.21
48.44
98.2
99.23

SVM
99.38
98.04
81.15
99.5
99.95

RandomForest
98.14
96.87
81.75
99.4
99.88

References

[10]
[11]

[1]

[2]

[3]

[4]
[5]
[6]

[7]

[8]

[9]

M. Pietikainen and T. Ojala, Texture Analysis in


Industrial Applications, J. L. C. Sanz (ed.), Image
Technology, pp. 337-359, Springer-Verlag Berlin
Heidelberg, 1996.
R. W. conners, C. W. McMillin, K. Lin and R. E. Vasquez,
Identifying and Locating surface Defects in Wood: Part of
an Automated Lumber Processing System, IEEE
transaction on Pattern aAnalysis and Machine
Interlligence, Vol. Pami-5, No. 6, pp. 573-583, Nov 1983.
V. Brecher, New techniques for patterned wafer
inspection based on a model of human preattentive vision,
Vol. 1708 Applications of Artificial Intelligence, Machine
Vision and Robotics, SPIE, pp. 452-459, 1992.
A. Ravishankar Rao, A Taxonomy for Texture
Description and
Identification,
Springer-Verlag,
NewYork, 1992.
C. Neubauer, Segmentation of defects in textile fabric
Proc. 11th International Conference on Pattern
Recognition, Vol. I, pp. 688-691, 1992.
J. Xie, Y. Jiang and H. T Tsui, Segmentation of Kidney
from Ultrasound Images Based on Texture and Shape
Priors, IEEE TRANSACTIONS ON MEDICAL
IMAGING, VOL. 24, NO. 1, JANUARY 2005.
Farhan Riaz, Ali Hassan, Saad Rehman, and Usman
Qamar, Texture Classification Using Rotation- and ScaleInvariant Gabor Texture Features, IEEE SIGNAL
PROCESSING LETTERS, VOL. 20, NO. 6, JUNE 2013.
Farhan Riaz, Francisco Baldaque Silva, Mario Dinis
Ribeiro, and Miguel Tavares Coimbra, Invariant Gabor
Texture Descriptors for Classification of Gastroenterology
Images, IEEE TRANSACTIONS ON BIOMEDICAL
ENGINEERING, VOL. 59, NO. 10, OCTOBER 2012.
U. Marmol, Use of Gabor Filters for Texture
Classification of Airborne Images and LIDAR data,
Archives of Photogrammetry, Cartography and Remote
Sensing, Vol. 22, pp. 325-336, 2011.

[12]

[13]
[14]

[15]
[16]

[17]

[18]

[19]

[20]

[21]

KNN
99.01
96.87
64.53
99.7
99.95

Multilayer Perceptron
99.38
97.09
85.48
99.7
99.95

Robert M. Haralick, Statistical and Structural approaches


to Textures, Proceedings of IEEE, Vol. 67, No. 5, pp. 786808, May 1979.
Gotlieb, C. C. and Kreyszig, Texture descriptors based on
co-occurrence matrices, Computer Vision, Graphics, and
Image Processing. 51(1), 70-86, 1990.
H. Permuter, J. Francos and I. Jermyn, A study of
Gaussian mixture models of color and texture features
for image classification and segmentation, Jaournal of
Pattern Recognition, No. 39, pp. 695-706, 2006.
D.Vaishali, 2D Autoreggressive Model for Texture
Analysis and Synthesis, International Conference on
Communication and Signal Processing, 2014.
Dongxiao Zhou, Texture Analysis and Synthesis using a
Generic
Markov-Gibbs
Image
Model,
https://www.cs.auckland.ac.nz/~georgy/research/texture/t
hesis-html/thesis.html. Accessed August 3, 2016.
M. Pietikainen, A. Hadid, G.Zhao and T.Ahonen:
Computer vision using local binary patterns (Springer,
2011).
M. Ahmed, A. Shaukat and M. U. Akram, Comparative
Analysis of texture Descriptors for Classification, IEEE
International Conference on Imaging Systems and
Techniques, IST, Oct 2016.
D-Chen and A. Safia, "Multi-Band Texture Batabase,"
[Online].
Available:
http://multibandtexture.recherche.usherbrooke.ca/index.ht
ml. [Accessed 2nd November 2016].
J. Ponce, "The Ponce Group Computer Vision and
Robotics," Beckman Institute, University of Illinois at
Urbana-Champaign,
[Online].
Available:
http://wwwcvr.ai.uiuc.edu/ponce_grp/. [Accessed 2 nd
November 2016].
"University of OULU Finland," [Online]. Available:
http://www.outex.oulu.fi/index.php?page=outex_home.
[Accessed 2nd November 2016].
"The KTH_TIPS Database," [Online]. Available:
http://www.nada.kth.se/cvap/databases/kthtips/index.html. [Accessed 2nd November 2016].
B.S.Manjunath and W.Y. Ma, Texture Features for
Browsing and Retrieval of Image Data, IEEE

[22]

[23]

[24]

[25]
[26]
[27]

[28]
[29]

TRANSACTIONS ON PATTERNANALYSISAND
MACHINE INTELLIGENCE, VOL. 18, NO. 8, AUGUST
1996.
A. Mumtaz, S. A. M. Gilani and T. Jameel, A Novel
Texture
Image
Retrieval
System
Based
On
Dual Tree Complex Wavelet Transform and Support
Vector Machines, IEEE--ICET 2006 2nd International
Conference on Emerging Technologies, November 2006.
S.Sulochana, R.Vidhya, Texture Based Image Retrieval
Using Framelet Transform: Gray Level Co-occurrence
Matrix (GLCM), (IJARAI) International Journal of
Advanced Research in Artificial Intelligence, Vol. 2, No.
2, 2013.
N.Puviarasan Dr.R.Bhavani and A.Vasanthi, Image
Retrieval Using Combination of Texture and Shape
Features, International Journal of Advanced Research in
Computer and Communication Engineering Vol. 3, Issue
3, March 2014.
Tamura, H., Mori, S. and Yamawaki, Textural features
corresponding to visual perception, IEEE Trans. On
Systems, Man and Cybernetics. 8(6), pp. 460-473, 1978.
Jack Sklansky, Image Segmentation and Feature
Extraction, IEEE Transactions on Systems, Man and
Cyberetics, Vol. Smc-8, No. 4, pp. 237-247, April 1978.
M. Ghiasi and R. Amirfattahi, Fast semantic segmentation
of
aerial
images
based
on color and texture, 8th Iranian Conference on Machine
Vision and Image Processing (MVIP), 2013.
Dana E. Ilea and Paul F. Whelan, Image segmentation
based on the integration of colourtexture descriptorsA
review, Pattern Recognition 44, pp. 24792501, 2011.
A. F. Costa, G. H-Mamani and A. J. M. Traina, An
Efficient Algorithm for Fractal Analysis of Textures,
XXV IEEE SIBGRAPI Conference on Graphics, Patterns
and Images, pp 39-46, 2012.

[30]
[31]

[32]

[33]

[34]

[35]

[36]

[37]

[38]

D. Saraswathi, G. Sharmila and E. Srinivasan, An


Automated Diagnoses System Using Wavelet Based SFTA
Texture Features, IEEE, ICICES, 2014.
T. Ahonen, J. Mathas, Chohe and M. Pietikinen,
Rotation Invariant Inmage Description with local binary
Patterns Histogram Fourier Features, SCIA, pp. 61-70,
Springer 2009.
S. Liao, MW. K. Law and Albert CS. Chung, Dominant
Local Binary Patterns for Texture Classification, IEEE
Transaction on Image Processing, Vol. 18, No. 5, May
2009.
Z. Guo, L. Zhang and D. Zhang, Rotation invariant texture
classificationnusing LBP variance (LBPV) with global
matching,Pattern Recognition 43, pp. 706719,
ELSEVIER, 2010.
A. Hassan, F. Riaz and S. Rahman, Rotation and scale
invariant texture classification by compensating for
distribution changes using covariate shift in uniform local
binary patterns, Electronics Letters, Vol. 50, pp. 27-29,
Jan 2014.
Rakesh Mehta and Karen Egiazarian, Dominant Rotated
Local BinaryPatterns (DRLBP) for texture classification,
Pattern Recognition Letters 71, pp. 1622, ELSEVIER
2016.
S. Zaid, J. Jadav and, J. Doere, An Efficient Wavelet
based approach for texture classification with feature
analysis, 3rd IEEE international Advance Computing
Conference (IACC), pp. 1149-1153, 2013.
S. Arivazhagan, L. Ganesan and S. P. Priyal, Texture
Classification using Gabor Wavelets based rotation
invariant features, Pattern Recognition Letters, No. 27,
pp. 1976-1982, 2006.
P. S. Liao, T. S. Chen and P. C. Chung, A Fast Algorithm
for Multilevel Thresholding, Journal of Information
Science and Engineering, No. 17, pp. 713-727, 2001.

Das könnte Ihnen auch gefallen