1. Image Texture Analysis Techniques- Survey
Anita Dixit, Dr.Nagaratna. P Hegde,
Asst.Prof, Dept of ISE, SDMCET Associate Prof., Dept of CSE
Dharwad. Vasavi College of Engineering, Hyderabad
Research Scholar, JNIAS ,Hyderabad nagaratnaph@gmail.com
anitadixitjoshi@gmail.com
Abstract blocks of pictorial data surrounds the area being
This paper discusses the various analyzed. Spectral features describe the average
methods used to analyze the texture property of tonal variation in various bands of visible and/or
an image. Texture analysis is broadly classified electromagnetic spectrum. Whereas Texture
into three categories: Pixel based, local feature feature describe the spatial distribution of tonal
based and Region based. Pixel based method variation with the band. Texture is concerned
uses grey level co occurrence matrices,
difference histogram and energy measures and with the spatial distribution of grey tones.
Local Binary Patterns(LBP) Local feature Texture can be classified into different types,
based method uses edges of local features and such as Fine, coarse or smooth, rippled,
generalization of co occurrence matrices. irregular or lineated. Texture is innate property
Region based method uses region growing and of virtually all surfaces-grain of wood, weave of
topographic models. a fabric, the pattern of crops in afield etc. it
Key Words: Co occurrence matrix,
contains the important information about the
Local Binary Pattern, Histogram
structural arrangement of surfaces and their
Motivation relationship with its surrounding environment.
Since textural properties are contain important
Texture analysis is one of the use full
information in discrimination purpose. Hence it
areas of study in machine vision. Human eyes
is important to build features for texture. [1]
are good judges of differentiating texture of
One common approach used to
natural surfaces. Successful vision system is the
characterize an image's spatial information is to
one which realizes this texture to the world extract features for classification which measure
surrounding it[2]. Major goals of texture the spatial arrangement of gray tones within a
research in computer vision are to understand, neighborhood of a pixel. This feature extraction
model and process texture, and ultimately to method is referred to as texture analysis and
simulate human visual learning process using includes a multitude of possible features that
computer technologies. have been developed to describe image texture.
Introduction
There are three fundamental features
with which a human being used to interpret
(a) (b) (c)
pictorial information; Spectral, Textural and
Contextual. Spectral information is nothing
but the average tonal variation in various bands.
Textual information gives the spatial
(d) (e) (f)
distribution of tonal variation with in a band. In Fig 1: Variety of textures (a) Tarmac (b) Brick (c)
contextual, information is derived from the wood (d) carpet (e) water (f) cloth

2. number of texture analysis techniques and some
Fig 1 shows the different types of texture which examples
are experienced my human vision system [4] in
general. The components of a texture, the
Texel(texture element), are notional uniform
Statistical method
Statistical methods analyze the spatial
micro-objects which are placed in an
distribution of gray values, by computing local
appropriate way to form any particular texture.
features at each point in the image, and deriving
If an intensity variation appears to be perfectly
a set of statistics from the distributions of the
periodic, it is called periodic pattern not texture.
local features. Depending on the number of
However, any completely random pattern would
pixels defining the local feature statistical
probably be called a ‘noise pattern’ rather than a
methods can be further classified into first-order
texture. If a pattern has both regularity and
(one pixel), second-order (two pixels) and
randomness then probably it would be called
higher-order (three or more pixels) statistics.
Texture.
The basic difference is that first-order statistics
Texture Analysis :
estimate properties (e.g. average and variance)
One common approach used to of individual pixel values, ignoring the spatial
characterize an image's spatial information is to interaction between image pixels, whereas
extract features for classification which measure second- and higher-order statistics estimate
the spatial arrangement of gray tones within a properties of two or more pixel values occurring
neighborhood of a pixel. This feature extraction at specific locations relative to each other. The
method is referred to as texture analysis[2] and most widely used statistical methods are co
occurrence features and gray level differences
includes a multitude of possible features that
[1], which have inspired a variety of
have been developed to describe image texture. modifications later on. These include signed
Texture analysis refers to a class of differences [7] and the LBP (Local Binary
mathematical procedures and models that Pattern) operator [6], which incorporate
characterize the spatial variations within imagery occurrence statistics of simple local
as a means of extracting information. Texture is microstructures, thus combining statistical and
an areal construct that defines local spatial structural approaches to texture analysis. Other
organization of spatially varying spectral values statistical approaches include autocorrelation
that is repeated in a region of larger spatial scale. function, which has been used for analyzing the
regularity and coarseness of texture and gray
Thus, the perception of texture is a function of
level run lengths but their performance has been
spatial and radiometric scales. Mathematical found to be relatively poor
procedures to characterize texture fall into four
general categories, statistical, geometrical, model- Model Based method
based methods and signal processing methods and Model-based methods hypothesize the
include Fourier transforms, convolution filters, co- underlying texture process, constructing a
occurrence matrix, spatial autocorrelation, parametric generative model, which could have
fractals, etc. [2] created the observed intensity distribution. The
Because texture has so many different intensity function is considered to be a
dimensions, there is no single method of texture combination of a function representing the
representation that is adequate for a variety of known structural information on the image
textures. Here, we provide a brief description of a
surface and an additive random noise sequence
Geometrical method

3. Geometrical methods consider texture difficult to apply to an image which is to be
to be composed of texture primitives, attempting segmented for texture analysis
to describe the primitives and the rules Autocorrelation shows the local intensity
governing their spatial organization. The variation as well as repeatability of the texture.
primitives may be extracted by edge detection It is use full for distinguishing short range and
with a Laplacian-of-Gaussian or difference-of- long range order in the texture. Auto correlation
Gaussian filter Once the primitives have been is not a good discriminator. in natural textures.
identified, the analysis is completed either by Hence Co occurrence matrix introduced by
computing statistics of the primitives (e.g. Harlick et al [1] became a large degree of
intensity, area, elongation, and orientation) or standard.
by deciphering the placement rule of the Pixel Based Models
elements The structure and organization of the In pixel based models texture is
primitives can also be presented using Voronoi described by statistics of distribution of grey
tessellations[2] Image edges are an often used levels or intensities in the texture. Most widely
primitive element. Harlick et al. [1] generalized used pixel based model is Grey Level Co
cooccurrence matrices, which describe second- occurrence model (GLCM). This is first
order statistics of edges. An alternative to introduced by Harlick et.al [1].
generalized cooccurrence matrices is to look for
pairs of edge pixels, which fulfill certain Grey Level Occurrence matrix (GLCM)
conditions regarding edge magnitude and The fundamental concept behind these
direction. Properties of the primitives (e.g. area matrices is spatial distribution of grey level
and average intensity) were used as texture elements. In this approach a set of matrices are
features created that show the probability that a pair of
Signal Processing Method brightness values (i,j) will occur at a certain
Signal processing methods analyze the separation from each other (Δx,Δy). The
frequency content of the image Spatial domain assumption is that the textural dependence will
filters, such as Law’s masks, local linear be at angles of 0°, 45°, 90° or 135° (with 0°
transforms proposed by Unser and Eden (1989), being to the right and 90° above) from the
and various masks designed for edge detection are original pixel that means four GLCM matrices
the most direct approach for capturing frequency would have to be created. Consider an image to
information Rosenfeld and Thurston (1971) be analyzed has Nx resolution horizontally and
introduced the concept of edge density per unit Ny resolution vertically. Grey tone appearing in
area: fine textures tend to have a higher density of each resolution cell is quantized to Ng levels.
edges than coarse textures. The set LxXLy is the set of resolution of
Texture Analysis methods: an image ordered in row and column. An image
I can be represented as function which assigns
Auto correlation and Fourier method. some grey tone in G. we assume that texture-
As we know that the texture is property context information in an image I is contained in
in which intensity of an image varies region to overall or average spatial relationship which the
region. This prompts us to calculate the variance grey tones in image I have to one another.
of intensity over the whole region of a texture. Texture-context information is more adequately
However most of the time this will not provide specified by matrix relative frequencies Pij with
enough description which is most of time which two neighboring pixels are separated by
needed. Especially when texels are well defined distance of d occur in an image such matrices of
or where there is high degree of periodicity in grey tone spatial dependency matrices are
texture. Then it is natural to consider the use of function of angular relationship between
Fourier analysis. Moreover Fourier method is

4. neighboring cells as well as the distance Instead of trying to explain texture formation on
between them. a pixel level, local patterns are formed. Each
Since all texture information is present in grey pixel is labeled with the code of the texture
tone spatial dependence matrices. Hence all primitive that best matches the local
texture features are extracted from these neighborhood. Thus each LBP code can be
matrices. There are total 14 set of features of regarded as a micro-texton. Local primitives
measures. But still it is difficult to say which detected by the LBP include spots, flat areas,
measure describes which feature of texture. edges, edge ends, curves and so on. Some
Following are 3 features out of 14 which define examples are shown in Fig. 3 with the LBP 8,R
the textural characteristics. They are Angular operator. In the figure, ones are represented as
Second Moment(ASM), Contrast(CON) and white circles, and zeros are black.
Correlation(COR)
These metrics are calculated for each pixel for
each using each of the four GCLMs and then a
final texture value is usually calculated as an Spot Spot/flat Line end Edge Corner
average of all four. It is obvious that these
Fig 3. Different Texture Primitives detected by LBP
measurements can be computationally
expensive especially as the quantization level The LBP distribution has both of the properties
becomes large. For many applications it may be of a structural analysis method: texture
beneficial to quantize the image into a smaller primitives and placement rules. On the other
number of gray levels prior to creating the hand, the distribution is just a statistic of a non-
GLCMs linearly filtered image, clearly making the
method a statistical one. For these reasons, it is
Local Binary Patterns to be assumed that the LBP distribution can be
The local binary pattern (LBP) texture successfully used in recognizing a wide variety
operator was first introduced as a of texture types, to which statistical and
complementary measure for local image structural methods have conventionally been
contrast. The first incarnation of the operator applied separately. Ojala et.al [7]
worked with the eight-neighbors of a pixel, .
using the value of the center pixel as a
threshold. An LBP code for a neighborhood was
produced by multiplying the threshold values Texture Classification
with weights given to the corresponding pixels, Texture classification refers to assigning
and summing up the result sample unknown image to one of a set of known
texture classes. Texture classification is one of
the four problem domains in the field of texture
analysis. The other three are texture
segmentation , texture synthesis, shape from
texture. Fig4 shows the general fame work used
for texture classification.
Fig 2.Calculating the original LBP code and a contrast
measure
Topi M¨aenp¨a¨a & Matti Pietik¨ainen[6] have
explained LBP as follows, The LBP method can
be regarded as a truly unifying approach.

5. unknown samples to be classified is
different from that of training data.
3. Rotation invariance; does the algorithm
cope, if the rotation of the images
changes with respect to the
viewpoint.[14]
4. Projection invariance (3-D texture
analysis); The algorithm may have to
cope with changes in tilt and slant
angles.
Fig-4. General frame work of texture classification 5. Robustness wrt. noise; how well the
algorithm tolerates noise in the input
Texture classification process involves two phases: images.
the learning phase and the recognition phase. In the 6. Robustness with respect to parameters;
learning phase, the target is to build a model for the the algorithm may have several built-in
texture content of each texture class present in the parameters; is it difficult to find the right
training data, which generally comprises of images values for them, and does a given set of
with known class labels. The texture content of the values apply for a large range of
training images is captured with the chosen texture textures.
analysis method, which yields a set of textural 7. Computational complexity;
features for each image. These features, which can be 8. Generativity; regenerating the texture
scalar numbers or discrete histograms or empirical that was captured using the algorithm.
distributions, characterize given textural properties of 9. Window/sample size; how large sample
the images, such as spatial structure, contrast, the algorithm requires being able to
roughness, orientation, etc. In the recognition phase produce a useful description of the
the texture content of the unknown sample is first texture content.
described with the same texture analysis method.
Then the textural features of the sample are compared Given a texture description method, the
to those of the training images with a classification performance of the method is often
algorithm, and the sample is assigned to the category demonstrated using a texture classification
with the best match. Optionally, if the best match is experiment, which typically comprises of
not sufficiently good according to some predefined following steps
criteria; the unknown sample can be rejected instead.
1. Selection of image data:
2. Partitioning of the image data into sub
Choosing an algorithm for Texture analysis images:.
When choosing a texture analysis 3. Preprocessing of the (sub)images:.
algorithm, a number of aspects should be 4. Partitioning of the (sub)images data into
considered [14] : training and testing sets.
5. Selection of the classification algorithm.
1. Illumination (gray scale) invariance; 6. Definition of the performance criterion:
how sensitive the algorithm is to changes two basic alternatives are available,
in gray scale. analysis of feature values and class
2. Spatial scale invariance; can the assignments,
algorithm cope, if the spatial scale of

6. It is obvious that the final outcome of a ellipses, and so on. Feature extraction tends to
texture classification experiment depends on identify the characteristic features that can form
numerous factors, both in terms of the possible a good representation of the object, so as to
built-in parameters in the texture description discriminate across the object category with
algorithm and the various choices in the tolerance of variations. [3] .
experimental setup. Results of texture Feature Extraction Methods
classification experiments have always been Serkan Hutipoglu, and Sunjit K.
suspect to dependence on individual choices in Mitra[5] suggested two different methods for
image acquisition, preprocessing, sampling etc., texture feature extraction, Quadratic teager
since no performance characterization has been filter and Singular value
established in the texture analysis literature decomposition(SVD). Quadratic teager filter
is used to find the local energy values. SVD
Markov random field models of texture values are used for feature extraction that
represents the low frequency property of an
Markov models have long been used for
image texture.
texture synthesis, to help with the generation of
realistic images. However, they have also
proved increasingly useful for texture analysis.
Applications
In essence a Markov model is a 1D construct in
Four major application domains related
which the intensity at any pixel depends only
to texture analysis are texture classification,
upon the intensity of the previous pixel in a
texture segmentation, shape from texture, and
chain and upon a transition probability matrix.
texture synthesis
Therefore, all experimental results should be
For texture analysis normally the image is
considered to be applicable only to the reported
converted to grey scale image. But the use of
setup. Fortunately, there is some recent work
joint color texture method using color histogram
aimed at improving the situation with
was proposed in [10].
standardized test benches, for example the
Using color and texture feature is an efficient
MeasTex framework for benchmarking texture
combination for content based image retrieval
classification algorithms. Additionally, an
[17]
increasing number of researchers are making the
Texture analysis is used majorly remote sensed
imagery and algorithms used in their work
images. Textural analysis techniques, namely
publicly available in the web
fractals and spatial autocorrelation methods,
were used to characterize these images in terms
Feature Extraction
of image complexity and roughness associated
Feature extraction (or detection) aims to
with forests. The effects of spatial and spectral
locate significant feature regions on images
characteristics of the data on the estimates of the
depending on their intrinsic characteristics and
textural indices were also examined.
applications. These regions can be defined in
Fractals are measures of the self-similarity and
global or local neighborhood and distinguished
thus ultimately measure the degree of
by shapes, textures, sizes, intensities, statistical
complexity of the imaged land surface
properties, and so on. Local feature extraction
Spatial autocorrelation is an assessment of the
methods are divided into intensity based and
correlation of a variable in reference to spatial
structure based. Intensity-based methods
location of the variable. Spatial autocorrelation
analyze local intensity patterns to find regions
measures the level of interdependence between
that satisfy desired uniqueness or stability
the variables, the nature and strength of the
criteria. Structure-based methods detect image
interdependence.
structures such as edges, lines, corners, circles,

7. Conclusion “Texture Feature Extraction Using Teager
Texture is one of the important feature Filters And Singular Value Decomposition”,
of recognizing an image. It is one such feature IEEE 1998
which cannot be defined properly in terms of
computer vision. [6] Topi M¨aenp¨a¨a & Matti Pietik¨ainen
Typically, a texture starts with a surface that ”Texture Analysis With Local Binary Patterns”,
exhibits local roughness or structure which is WSPC,2004
then projected to form a textured image. Such
an image exhibits both regularity and [7]. Timo Ojala, MattiPietikaÈinen, “Multi
randomness to varying degrees: directionality resolution Gray-Scale and Rotation Invariant
and orientation will also be relevant parameters Texture Classification with Local Binary
in a good many cases. However, the essential Patterns”, IEEE Transactions on Pattern
feature of randomness means that textures have Analysis and Machine intelligence, vol. 24, no.
to be characterized by statistical techniques, and 7, July 2002
recognized using statistical classification [8]. M. PIETIKÄINEN, T. OJALA and Z. XU,
procedures. Techniques that have been used for “Rotation-Invariant Texture Classification
this purpose have been seen to include Using Feature Distributions”, M.
autocorrelation, co-occurrence matrices, texture PIETIKÄINEN, T. OJALA and Z. XU, Pattern
energy measures, fractal-based measures, recognition, Vol 33,No1,pp43-52,2000
Markov random fields, and so on. [9]. Zhi Li, Guizhong Liu, Yang Yang, and
In this paper I have made an effort to Junyong You, “Scale- and Rotation-Invariant
present the various methods of texture analysis, Local Binary Pattern Using Scale-Adaptive
texture classification and their applications. Texton and Sub uniform-based Circular Shift
References [10]. Matti Pietik.inen, Topi M.enp.. and Jaakko
[1]. Harlick and Shanmugam, “Textural feature Viertola” Color Texture Classification with
for image classification”, IEEE 1973 Color Histograms and Local Binary Patterns”,
[2] Mihran Tuceryan, Anil K. Jain, “Texture [11].Zhenhua Guo, Lei Zhang, A Completed
Analysis”, The Handbook of Pattern Modeling of Local Binary Pattern Operator for
Texture Classification, submitted to IEEE
Recognition and Computer Vision (2nd
transaction 2010
Edition), by C. H. Chen, L. F. Pau, P. S. P.
Wang (eds.), pp. 207-248, World Scientific [12] Mehrdad J. Gangeh,_, Robert P.W. Duin,
Publishing Co., 1998. Bart M. ter Haar Romeny, Mohamed S. Kamel,
A Two-Stage Combined Classifier in Scale
[3] FRANK Y. SHIH, “Image processing and Space Texture, Classification Elsevier, July
pattern Recognition Fundamentals and 2012.
Techniques”, Image Processing and Pattern
Recognition, IEEE 2010. [13] K. Muneeswaran, L. Ganesan, S.
Arumugam, K. Ruba Soundar, Texture
[4] E. R. Davies, “Introduction to Texture classification with combined rotation and scale
Analysis”, HANDBOOK OF TEXTURE invariant wavelet features, Pattern Recognition
ANALYSIS ,Imperial College Press 38 (2005) pp 1495-1506.
[5] Serkan Hutipoglu, and Sunjit K. Mitra,

8. [14] T Ojala and M Pietikäinen, Texture
classification,1998 [18] Manik Varma and Andrew Zisserman, A
Statistical Approach to Texture Classification
[15] L.Li, C.S Tong, S.K Choy, Texture from Single Images, Kluwer Academic
Classification using refined histogram. IEEE Publishers,2004
Transaction vol 19,no 5,May 2010 [19]. C. Laymon and M. Al-Hamdan,
“Information Extraction from Remote Sensing
[16] Minh N. Do and Martin Vetterli, Rotation Imagery Using Textural Analysis “,National
Invariant Texture Characterization and Retrieval Consortium on Remote Sensing in
using Steerable Wavelet-domain HiddenMarkov Transportation, Dec2005
Models [20]. Timo Ojala, Matti Pietika¬inen,
[17] Young Deok Chun, Nam Chul Kim, “Unsupervised texture segmentation using
Content-Based Image Retrieval Using Multi feature distributions”, Pattern Recognition 32
resolution Color and Texture Features, IEEE (1999) 477-486
Transactions On Multimedia, Vol. 10, No. 6,
October 2008