The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B4. Beijing 2008
1372
CON=TTP(i~j) 2
‘ j
HOMO=IZ-^Ar
ijHHf
DIS=X Z | i~j\P{ i J )
i j
ENT-X Z P{ i »y )[— log P{ij)~\
i j
' ASM=YTP 2 {i,j)
i j
MEAN=Y'Li*P(i,j)
i j
SD= lYXP(i-MEAN) Z
(i-MEAN)( j-MEAN)
2T
cor=xx
i j
SD<
(1)
Where Z and j are two different grey levels of the image, P is
the number of the co-appearance of grey levels Z and j .
Edge density {ED) is usually computed through the number of
edge pixels in a given window divided by the window size
(Equation 2). The detection of edge pixels is the key issue in
ED computing. There are a variety of algorithms have been
proposed for edge detection. Among them, Canny edge
detection operator (Canny, 1986) is one of effective method. It
formulated edge detection as an optimization problem and
defines an optimal filter, which can be efficiently approximated
by the first derivative of Gaussian function in the one-
dimension case. In this study, edge detection by the Canny edge
detection operator was performed and a binary edge image
(edge pixel is coded as ‘1’, and non-edge pixel is coded as ‘0’)
was produced. Then, ED can be computed based on the binary
edge image.
i w/2 w/2
ED(x,y)=— X Z g(x+k,y+l)
w k=-w/2l--w/2
(2)
Where W is the window size. g(x,y) is the pixel value in the
given window.
the textural features whose differences will be used to
characterize various class types need to be extracted over a
local area of unknown size and shape. If this information is
gathered over areas that are not large enough with respect to the
texture elements or variations, then one cannot expect these
local analyses to provide feature values that are invariant across
the textured region. Consequently, it is desirable to extract the
textural information over as larger an area as possible. If this is
the case (i.e. texture features are not calculated from a single
texture class), the features would be representing a hybrid
values. This problem is similar to mixed-pixel problem and may
be termed as mixed-texel problem (Shaban and Dikshit, 2001).
Therefore, the need for a large window size results in a trade
off between large window sizes that give stable texture
measures and the increasing proportion of between-class
variance texture pixels such large windows produce.
2.2.2 Quantization level
The dimension of a GLCM is determined by the maximum gray
value of the pixel. The more levels included in the computation,
the more accurate the extracted textural information, of course,
a subsequent increased computation cost (Soh and Tsatsoulis,
1999). Some of the major quantization schemes are uniform
quantization, Gaussian quantization and equal probability
quantization. The uniform quantization scheme is the simplest,
in which gray levels are quantized into separate bins with
uniform tolerance limits with no regard to the gray level
distribution of the image. This technique is not always
preferable. The Gaussian quantization technique is one such
scheme. The grey level distribution of the original image is
assumed to behave normally. Each quantization bin has the
same area under the curve and thus different space smaller
spaces in the middle of the distribution and larger spaces at the
tails of the distribution. In the equal probability quantization
scheme, each bin has similar probability and it has been shown
to represent accurate representation of the original image in
terms of textural based on GLCM (Conners and Harlow
1978).The Guassian quantization scheme assume a Gaussian
grey level distribution, which is not always true for high-
resolution imagery. Equal probability quantization normalizes
different image samples so that a bright feature and a dark
feature, given the same texture, would have the same co
occurrence matrix, which is undesirable since grey value is
important in residential analysis. Thus, in our experiment, we
have focused on the uniform quantization scheme.
2.2.3 Displacement
The displacement parameter^ is important in computation of
GLCM. Applying large displacement value to a fine texture
would yield a GLCM that does not capture detailed textural
information, and vice versa (Soh Tsatsoulis, 1999).
2.2 Parameters effect on texture
Although absolute values of texture features had little meaning,
it was worthwhile to understand how each feature varied with
varying parameters given an image acquisition configuration.
This should provide users with the knowledge with which to
make a good selection of parameter values instead of testing all
possible combinations. The following four parameters need to
be pre-set for designing the texture features introduced above.
2.2.1 Window size
The moving window size used to calculate texture is a key
parameter. In texture analysis, one of the main problems is that
2.2.4 Orientation
Every pixel has eight neighboring pixels allowing eight choices
for 9 , which are 0°, 45°, 90°, 135°, 180°,225°, 270°or 315°.
However, taking into consideration the definition of GLCM, the
co-occurring pairs obtained by choosing 6 equal to 0°would be
similar to those obtained by choose 0 equal to 180°. This
concept extends to 45°, 90° and 135° as well. Hence, one has
four choices to select the value of 9 . Sometimes, when the
image is isotropic, or directional information is not required,
one can obtain isotropic GLCM by integration over all angles.