The Angular Radial Partitioning(ARP) [7, 8], is a novel approach for image representation based on geometric distribution of edge pixels. Not like shape context, input image may consist of several complex objects. It starts by converting the color image to a gray intensity image by eliminating
the hue and saturation while retaining the luminance. In the
next step the algorithm normalizes the gray image to the size
201X201, followed by applying the Canny Edge operator for
edge detection. The normalized edge image I, is employed
for feature extraction where edge pixels are considered as
'1' and '0' otherwise. In the following, we consider pixels
I(; ) to be either equal to T1T
for edge pixels or T0T
for
non-edge pixels. The algorithm uses the surrounding circle
of I for partitioning it to MXN sectors, where M is the
number of radial partitions and N is the number of angular
partitions. The number of edge points in each sector of I is
chosen to represent the sector feature. The scale invariant
image feature f(k; i) is then defined to be: