rx X me. ze
|
its 710, TEXTURE ANALYSIS
(data
lution Although texture is one of the more
important characteristics used by a human
ar images to identify objects of interest in an
igital image, machine classification of digital
,, each images is frequently based solely upon
lar image spectral features which describe the
dar overall tonal variations within an image.
des Textural features, however, describe the
jf the local spatial distribution of tonal
) for values within an image (Ref.9). In spite
of tremendous advances in the field of
nt set computer technology during the last de-
1 and Figure 5. "Texture" Image From cade, the cost of N EEE digital
the ir Data Set 709 (X) image texture on an general-purpose
1e computer is still very high. The appli-
1 order cation of classification procedures in-
NS volving image texture to remote sensing agricultural surveys has remained pri-
could marily a research topic due to the magnitude of the computational resources
required to achieve a reasonable throughput. With the advent of SAR, which can
igures achieve a much finer resolution than a real aperture radar, as well as recent
of advances in digital computation devices, there is hope to exploit texture in-
formation in support of agricultural inventories.
same Reviewing the literature on texture models (Refs.9-11) there is strong
r, supportive evidence that the Spatial Gray Level Dependence Method (SGLDM) is
center representative of the best texture algorithms currently available. Consequently
Mta in I selected the SGLDM-Approach as a starting point for my research.
| and Initially, it is assumed that the texture context information in an image is
rical contained in the overall or "average" spatial relationship which the gray tones
st in the image have to one another. This relationship can be characterized by the
i- south set of co-occurrence matrices P (i,j) d,0 whose i,j*^ element is the relative
on down frequency with which two neighboring picture elements separated by distance d
Irns in a direction 9 occur in the image, one with gray tone i and the other with
gray tone j.
|ternate : à É
The computation of the texture measurement values involved three steps. First,
the "Texture Image" (Fig.4) was level-sliced into a smaller number of intervals
in order to ease the computational burden. Since the majority of the data values
lay between 100 and. 160 with a mean of y - 128 and a standard deviation of
c- l0, I chose to quantize into 8 intervals of one standard deviation around
tively the mean. The next step consisted of calculating the co-occurrence matrices over
appear a moving circular window of radius 5 pixels using a displacement distance of
e same d = 2 and four different angles, 8 = 0°, 45°, 90°, and 135°. Textural features
of these (scalar functions) were then computed from these co-occurrence matrices at
ition each location in the 1024 x 1024 image. In order to obtain texture measures
gery which are insensitive to the orientation of the sensor, I chose to use the
ly average value of a texture measure over all four angles (i.e., from each of the
jing four 8 by 8 co-occurrence matrices developed at each possible location of the
ize moving window). The scalar function which was employed in this investigation
ent is called "Contrast" (Refs.9, 11) or "Inertia" (Ref.12). Its mathematical de-
I finition is described in Eq. 3.
5).
235
CES E a e acil ————
nd M ei N NE FE SS AR