GNR602

Advanced Methods in Satellite Image

Processing
Instructor: Prof. B. Krishna Mohan
CSRE, IIT Bombay
[email protected]

Slot 13
Texture Segmentation Methods
Lectures 9-11 February 2023
IIT Bombay Slide 1
Lectures 9 – 11 February 2023 Texture Segmentation Methods

Contents of the Lecture

Concept of Texture
Co-occurrence matrix principle
• Texture features
Gabor Texture Feature Extraction Method

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 2

Concept of Texture
• Texture is an important visual cue
• What does texture mean? Formal approach or
precise definition of texture does not exist!
• Texture discrimination techniques are for the
part ad hoc.

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 2a

Concept of Texture
• Perception of texture is dependent on the spatial
organization of gray level or color variations.
• Manmade features have a repetitive pattern, where
a basic pattern or primitive is replicated over a region
• Large variation within the pattern leads to a textured
appearance, while flat regions lead to a smooth
appearance

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 3

Sample
Textures

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 4

Sample Textures

Source: www.pepfx.net

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 5

More Examples of Texture

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 6

From Remotely
Sensed Images

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 7

What is Texture?
• A feature used to partition images into regions of
interest and to classify those regions
• Spatial arrangement of colours or intensities in an
image
• Characterized by the spatial distribution of intensity
levels in a neighbourhood
• A repeating pattern of local variations in image
intensity
• An area attribute, not defined at a point

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 8

What is Texture?

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 9

Notion of Texture
• Suppose an image has a 50% black and 50% white
distribution of pixels.
• Three different images with the same intensity
distribution, but with different textures.

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 10

Composition of Texture
• Made up of texture primitives, called texels.
• Can be described as fine, coarse, grained, smooth,
etc.
• Tone is based on pixel intensity properties in the
texel, while structure represents the spatial
relationship between texels.
• If texels are small and tonal differences between
texels are large a fine texture results.
• If texels are large and consist of several pixels, a
coarse texture results.

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 11

Notion of Texture
• Statistical methods are particularly useful when
the texture primitives are small, resulting in
microtextures.
• When the size of the texture primitive is large,
first determine the shape and properties of the
basic primitive and the rules which govern the
placement of these primitives, forming
macrotextures.
GNR602 Lecture 9-11 B. Krishna Mohan
IIT Bombay Slide 12

Example of micro- and macro-texture

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 13

Description/Definition of Texture
• Non-local property, characteristic of region
more important than its size
• Repeating patterns of local variations in image
intensity which are too fine to be
distinguished as separated objects at the
observed resolution

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 13a

Definition of Texture
• There are three approaches to describing what
texture is:
• Structural : texture is a set of primitive texels in some
regular or repeated relationship.
• Statistical : texture is a quantitative measure of the
arrangement of intensities in a region.
This set of measurements is called a feature vector.
• Modeling : texture modeling techniques involve
constructing models to specify textures.

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 14

Texture Analysis
• Two primary issues in texture analysis:
- texture classification
- texture segmentation
• Texture classification is concerned with identifying a given textured
region from a given set of texture classes.
Each of these regions has unique texture characteristics.
Statistical methods are extensively used.

• Texture segmentation is concerned with automatically determining

the boundaries between various texture regions in an image.

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 15

Texture Classification
• Texture classification is concerned with
identifying a given textured region from a
given set of texture classes.
• Each of these regions has unique texture
characteristics.
• Statistical methods are extensively used.

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 16

Texture Segmentation
• Texture segmentation is concerned with
automatically determining the boundaries
between various texture regions in an image.
• Texture segmentation also results in regions
homogenous with respect to texture property

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 17

Example

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 18

Approaches to Measuring Texture

Edge per unit area
First Order Statistics
Mean / average, Standard deviation
Mean Deviation, Range, Median,
Skewness
Higher order statistics
Measuring energy in various frequency sub-bands
Fractal modeling
Geostatistical methods
Wavelet transform approaches
…
GNR602 Lecture 9-11 B. Krishna Mohan
IIT Bombay Slide 19

Edge per unit area

Textured areas are seen to be rough – spatial intensity
variations over small patches
Gradient operators produce moderate edge
magnitudes
Measuring the average edge magnitude over an area
(e.g., over 11x11 or 15x15) can help separate
textured areas from non-textured areas

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 20

Variance
In textured areas both high and low intensity
pixels can be found
Variance of the pixel intensities over an area
will be higher for textured areas compared to
non-textured areas
Variance image (e.g., as in ERDAS software) can
be used to represent texture

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 21

Directionality of Texture
Texture is a strong directional feature
e.g., horizontal stripes and vertical stripes are
clearly perceived separately
Some texture features can provide directional
information
Features like edge per unit area or variance
cannot handle texture orientation

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 22

Gray Level Co-occurrence Matrix

Approach
• GLCM is based on second order statistics (2-D histogram)
• It is conjectured (B. Jules, a psychophysicist) that textures
differing in second order statistics are indeed different.
(counter-examples provided later)
• Therefore numerical features were extracted from the image
in terms of the second-order statistics that were a measure of
the underlying texture.

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 23

Construction of GLCM
• A co-occurrence matrix is a two-dimensional array, P, in which
both the rows and the columns represent a set of possible
image values.

• A GLCM Pd[i,j] is defined by first specifying a displacement

vector d=(dx,dy) and counting all pairs of pixels separated by d
having gray levels i and j.

• (dx,dy) define the directionality of texture; dx=1,dy=0

represents horizontal direction;dx=1,dy=1 represents diagonal
direction

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 24

Definition of GLCM
• The GLCM is defined by:

Pd(i,j) = ni,j = #{f(m,n) = i, f(m+dx, n+dy) = j;

1≤m≤M; 1≤n ≤N}
– where nij is the number of occurrences of the pixel
values (i,j) lying at distance d in the image.

– The co-occurrence matrix Pd has dimension n× n,

where n is the number of gray levels in the image.

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 25

Example

There are 16 pairs of pixels in the image which satisfy this

spatial separation. Since there are only three gray levels –
0,1,2, P[i,j] is a 3×3 matrix.
GNR602 Lecture 9-11 B. Krishna Mohan
IIT Bombay Slide 26

Algorithm to construct GLCM

Count all pairs of pixels in which the first pixel has a value i, and
its matching pair displaced from the first pixel by d has a value
of j.

This count is entered in the ith row and jth column of the matrix
Pd[i,j]

Note that Pd[i,j] is not symmetric in this form of counting, since

the number of pairs of pixels having gray levels [i,j] does not
necessarily equal the number of pixel pairs having gray levels
[j,i].

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 27

Normalized GLCM
The elements of Pd[i,j] can be normalized by dividing
each entry by the total number of pixel pairs.

Normalized GLCM N[i,j], defined by:

P[i, j ]
N [i, j ] 
 P[i, j ]
i j

which normalizes the co-occurrence values to lie between 0

and 1, and allows them to be thought of as probabilities.

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 28

Numeric Features from GLCM

Gray level co-occurrence matrices capture properties of
a texture but they are not directly useful for further
analysis, such as the comparison of two textures.

Numeric features are computed from the co-

occurrence matrix that can be used to represent the
texture more compactly.

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 29

Haralick Texture Features

Haralick et al. suggested a set of 14 textural features
which can be extracted from the co-occurrence
matrix, and which contain information about image
textural characteristics such as homogeneity,
linearity, and contrast.

Haralick, R.M., K. Shanmugam, and I. Dinstein, "Textural

features for image classification” IEEE Transactions on
Systems, Man and Cybernetics: pp. 610-621. 1973.

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 30
Features from GLCM: Angular Second
Moment (ASM)
• Angular Second Moment ASM
K K
• ASM =  d (i, j ) / R
P 2

i 1 j 1

• R is a normalizing factor
• ASM is large when only very few gray level pairs are present in
the textured image
• K is the number of gray levels

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 31

Contrast (CON)
• Contrast CON
K K

• CON =  (i  j
i 1 j 1
) 2
Pd (i, j ) / R

• This feature highlights co-occurrence of very

different gray levels

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 32

Entropy (ENT)
• ENT = K K
 1 
 P[i, j ]ln 
P[i , j ]

i 1 j 1  
• ENT emphasises many different
co-occurrences
• P(i,j) is the normalized co-occurrence matrix,
each entry indicating probability of occurrence
of that gray level combination

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 33

Inverse Difference Moment (IDM)

• Inverse Difference Moment IDM
K K
Pd r [i, j ]
• IDM = 
i 1 j 1 | ( i  j ) m
|i  j
• IDM emphasises co-occurrence of close gray
levels compared to highly different graylevels.
m and r can user specified

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 34

Algorithm for image segmentation

• Specify a window of size wxw
• For the pixels in the window, compute the co-
occurrence matrix
• Derive the texture features from the co-occurrence
matrix
• Move the window by 1 pixel, and repeat the
procedure
• The procedure leads to texture images that may be
treated like additional bands, equal to the number of
features computed.

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 35

Input
Image

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 36

IDM
Feature

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 37

CLASSIFIED
IMAGE
(Mumbai)

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 38

CLASSIFIED IMAGE (Mumbai)

LEGEND

WATER

MARSHY LAND / SHALLOW WATER

HIGHLY BUILT-UP AREA

PARTIALLY BUILT-UP AREA

OPEN AREAS/ GROUNDS

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 39

Other Features from GLCM

• More features are defined from GLCM by Haralick et
al. and by other researchers
• e.g., Sahasrabudhe and Nageswara Rao used
eigenvalues of GLCM as texture features
• 1st and 2nd eigenvalues of GLCM were shown to be
capable of good texture discrimination
• Limited utility due to computational intensive nature
of features
• Haralick et al. defined 28 features of which ASM, ENT,
CON, IDM were most effective

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 40

Fast Computation of GLCM

• Fast Computation of GLCM useful for efficient
application
• Basis for fast computation – number of pixel pairs
that are common to computation of GLCM / features
at successive positions of the window
• Significant savings possible when window size is large,
and many features are computed

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 41

Redundant Computations
• When texture window moves by 1 pixel right,
– The first column moves out of the computation
– The last column enters the computation
– Many pixel pairs remain unchanged

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 42

Efficiency Considerations
• Incremental Adjustments
– Deduct the pairs formed with elements of first column
– Add pairs formed with elements of last column
– New matrix is ready

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 43

Efficiency Considerations
• Direct computation of features
– Examine each feature
– Make modifications to the feature directly instead of to the
GLCM
K K

• ASM = P
i 1 j 1
d
2
(i , j ) / R

• Deduct from Pd (i,j) for column moving out

• Add to Pd (i,j) for column coming in

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 44

Fast Computation
• ASM and CON can easily be adjusted for changes in
GLCM due to window shift
• ENT and IDM may be difficult

GNR602 Lecture 9-11 B. Krishna Mohan

 IIT Bombay Slide 45
Laws' texture energy measures
• Laws’ Texture description uses
– average gray level
– edges
– spots
– ripples
– waves

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 46

Laws Texture Energy Approach

• Some of the masks
• L5 = [ 1 4 6 4 1 ];
• E5 = [ -1 -2 0 2 1 ]
• S5 = [ -1 0 2 0 -1 ];
• W5 = [ -1 2 0 -2 1]
• R5 = [ 1 -4 6 -4 1]

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 47
Texture Energy
2-dimensional masks are created by cross-products
of the one-dimensional masks
The masks are used as filters to convolve with the
input image
Energy within specified window sizes WxW is
computed

GNR602 Lecture 9-11 B. Krishna Mohan

 IIT Bombay Slide 48
Energy and Texture
• E5L5 L5= [ 1 4 6 4 1]t; E5 = [-1 -2 0 2 1]t
E5L5 Cross product of the two
-1 -4 -6 -4 -1 vectors of size 5x1
-2 -8 -12 -8 –2 resulting in a mask of size
0 0 0 0 0 5x5
2 8 12 8 2
1 4 6 4 1

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 49
Laws Filter Masks
• Another filter mask:

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 50

Texture Based Analysis

Steps
– Convolve input image with Laws’ masks
– Nonlinearly transform Laws’ filter outputs
using sigmoid or similar function
– Using a moving window (or size 15x15 or
similar), find energy in each window position
– Generate texture energy images
– Classify using standard classifiers

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 50a
Input Image

Stage 1
Filter
Banks

Stage 2
Non-
Linearity

Stage 3
Energy
Compu-
tation

Gabor filter bank for texture analysis

GNR602 Lecture 9-11 B. Krishna Mohan
IIT Bombay Slide 51

Illustration

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 52
Transform Domain Texture Analysis
• Rapid or slow changes in local intensity or
color in the spatial domain corresponding to
high or low frequency components in the
transform domain
• Orientation in one domain is reflected in the
other domain in a perpendicular direction
• Texture therefore can also be analyzed
efficiently in transform domain as well

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 53

Transform Domain Texture Analysis

• Among the transform domain approaches,
Gabor functions and wavelets are popularly
used.
• Essentially the image is decomposed into
different frequency bands. Rapid local tonal
variations correspond to higher frequencies,
while slow variations correspond to lower
frequencies
GNR602 Lecture 9-11 B. Krishna Mohan
IIT Bombay Slide 54
Gabor Filters for texture analysis
• Multi-channel filtering approach to texture analysis
are:
• Functional characterization of the channels and the
number of channels
• Extraction of appropriate texture features from the
filtered images
• The relationship between channels (dependent vs
independent)
• Integration of texture features from different
channels to produce a segmentation
GNR602 Lecture 9-11 B. Krishna Mohan
IIT Bombay Slide 55
Gabor Textures
Different multi-channel filtering techniques
that are proposed in the literature differ in
their approach to one or more of the above
issues.

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 56

Gabor Filters
• The frequency spectrum of the image is divided into
radial and angular ranges such that there are Nr x No
windows in the frequency domain
• Retaining each such window, the inverse transform is
computed to generate a bandpass filtered input
image
• This process is based on the early processing of
visual information in the human visual cortex

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 57
Gabor Filter Bank
Impulse response and Fourier transform of a
real-symmetric Gabor filter

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 58
Gabor Filter Bank
Instead of using ideal low/high/bandpass filters,
Gaussian shaped filters are used for stability

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 59

Frequency
lobes

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 60
Stages in Gabor Filter Banks
Approach
-Bandpass filtering using radial and
angular sector basis
-Nonlinear transformation on the output of
bandpass filtering
-Energy computation in moving window of
size like 11x11 or 15x15 etc.
Nonlinear
Transformation -Classification of texture images
-Number of texture images = number of
bandpass filters

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 61
Input Image

Stage 1
Filter
Banks

Stage 2
Non-
Linearity

Stage 3
Energy
Compu-
tation

Gabor filter bank for texture analysis

GNR602 Lecture 9-11 B. Krishna Mohan
IIT Bombay Slide 62

Nonlinear transformation
• The inverse Fourier transformed images are
passed through a nonlinear function, similar
to the way the neural computation happens in
the human system
• Typical nonlinearity is a hyperbolic tangent
function
• tanh(α(t))=[(1-exp(-2 αt))/(1+exp(-2 αt))]

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 63
Energy Computation
• From the nonlinear transformed images, texture
energy is computed over local windows of user
specified size
• Jain and Farrokhnia, who first proposed this
approach, used 4 intervals in the radial direction, 4
angular sectors for orientation preference, one
lowpass and one highpass filter, in all 18 texture
images derived from the input image
• Ref: Pattern Recognition, vol. 24, pp.1167-1186,
1991

GNR602 Lecture 9-11 B. Krishna Mohan

IIT Bombay Slide 64

Illustration

Improved version can be found at "Integrating Region and Edge Information for
Texture Segmentation …", Image and Vision Computing, Vol. 26, No. 8, pp.
1106-1117, August 2008.

GNR602 Lecture 9-11 B. Krishna Mohan

Contd…