Scribd
Upload
129 views

Feature Extraction Gabor Filters

This document discusses using Gabor filters for extracting texture features to discriminate between defective and normal textures. It defines the Gabor filter as a Gaussian kernel function modulated by a sinusoidal plane wave. The Gabor filter has tunable parameters like orientation, spatial frequency, and bandwidth that allow it to be adapted for different defect types. When applied to a discrete texture image using 2D convolution, the Gabor filter acts as an orientation-selective bandpass filter that can extract discriminative features for defect detection. The goal is to optimize the Gabor filter parameters using an unsupervised cost function that maximizes separation between defective and normal textures.

Uploaded by

Zeeshan Aslam
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
129 views

Feature Extraction Gabor Filters

This document discusses using Gabor filters for extracting texture features to discriminate between defective and normal textures. It defines the Gabor filter as a Gaussian kernel function modulated by a sinusoidal plane wave. The Gabor filter has tunable parameters like orientation, spatial frequency, and bandwidth that allow it to be adapted for different defect types. When applied to a discrete texture image using 2D convolution, the Gabor filter acts as an orientation-selective bandpass filter that can extract discriminative features for defect detection. The goal is to optimize the Gabor filter parameters using an unsupervised cost function that maximizes separation between defective and normal textures.

Uploaded by

Zeeshan Aslam
Copyright
© Attribution Non-Commercial (BY-NC)
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 2

Problem Formulation

Let F represents local image feature extraction process dened by a set of


parameters :
F(
i
, S) = (x
1
, x
2
, ..., x
n
)

i
(1)
X

i
= (x
1
, x
2
, ..., x
n
)

i
N-dimensional local feature vector extracted for
a particular parameter set
i
of the feature extraction process to represent
faultless texture.
Extraction of defect discriminative robust texture features invloves optimiza-
tion of feature extraction process based on an unsupervised cost function C that
can represent appropriate measure of discrimination of defective and normal
texture features. Here unsupervised refers to ability of cost function to design
defect descriminative texture features without requiring any prior knowledge of
defects.
C(X

i
) = c

i
(2)
features that can descriminate Optimization of parameters of feature extrac-
tion requires that design of a cost function C that can represent unsupervised
measure of optimal separation between defects and normal texture is used for
the optimization of parameters of feature extraction process. Here unsupervised
refers to the fact that knowledge of defective texture is not required
g

(x, y) =
1
2
x

y
exp
_

1
2
_
x
r
(x, y, )
2

2
x
+
y
r
(x, y, )
2

2
y
__
exp (j2 (u
o
x +v
o
y))
(3)
Gabor envelope:
w

x
,
y
(x, y) =
1
2
x

y
exp
_

1
2
_
x
r
(x, y, )
2

2
x
+
y
r
(x, y, )
2

2
y
__
where = (u
o
, v
o
,
x
,
y
, ) represents parameter vector of Gabor lter.
Parameters u
o
and v
o
dene spatial frequency of sinusoidal carrier in cartesian
coordinates. Spatial extent and bandwidth of Gaussian envelope is determined
by (
x
,
y
) with being its rotation angle. (x
r
, y
r
) are the rotated (x, y) coor-
dinates:
x
r
(x, y, ) = xcos +y sin
y
r
(x, y, ) = xcos +y cos
The Gabor function forms a complete but nonorthogonal basis set.The real
(even symmetric) and imaginary (odd symmetric) parts of the Gabor function
can be separated out as:
g

e
(x, y) = w

x
,
y
(x, y) cos (2 (u
o
x +v
o
y))
(4)
g

o
(x, y) = w

x
,
y
(x, y) sin (2 (u
o
x +v
o
y))
Only even symmetric (real part) Gabor lters have been used in the proposed
defect defect scheme. This choice is more than justied as it causes huge savings
1
in computational resources with little or no loss in performance. Moreover,
Gabor lters with asymmetric (
x
=
y
) region of support are more suited for
real fabric defect detction than symmetric ones [Ajay].
Gabor function is shifted gaussian in the spatial-frequency domain and its
Fourier transform is given by:
G(u, v) = exp
_

1
2
_
(u u
o
)
2
r

2
u
+
(v v
o
)
2
r

2
v
__
(5)
Where

u
=
1
2
x
and
v
=
1
2
y
Gabor lter is orientation selective bandpass lter with tunable center fre-
quencies (u
o
x, v
o
y) and adjustable angular and radial frequency bandwidths.
This exibility in designing frequency characteristics and makes it possible to
design Gabor lters for great diversity of defect types.
In reality, a texture image T doesnt represent a continous function but
a matrix of discretely sampled set of intensity values T (x
i
, y
j
) = t
i,j
, i =
1, . . . , P, j = 1, . . . , Q, with points (x
i
, y
j
) forming a regular grid on the image
T. So, if Gabor lter g

o
(x, y) and image T represent discrete image data then
ltering is dened by the following 2-D discrete convolution operation:
_
T g

e
_
[x, y] =
n

j=n
m

i=m
T (i, j) g

e
(x i, y j) (6)
Where m =
M1
2
and n =
N1
2
with M N being the kernel size. In order
to have symmetric region of support M and N are preferred to be odd.
Add gures showing Gabor lter in spatial ( both real and imaginary) and
frequency domain as well if possible,. . . , if DC reponse of cosine phase Gabor
lter is eliminated then sine and cosine Gabor pair forms a quadrature pair hav-
ing same amplitude response and phase response shifted by 90

[Gabor Tutorial,
Energy ltering].
2

You might also like