A Novel Shape Descriptor for Rectangle-like Classification

In this paper, we propose a novel shape descriptor using the concept of projection and statistical moment to generate a new effcient descriptor (called K descriptor). The K descriptor can be used as a robust and invariant descriptor of global shape of an image. The K descriptor can also extract an intrinsic ”rectangularity” property from non smooth rectangular image. With the help of K descriptor, the rectangle-like and circle-like image are classified according to the three features: (1) the minimum, (2) the maximum and (3) the standard deviation of K-descriptor profile. SVM was used as the classifier. In the experiment on 1,000 binary image (500 testing, 500 training), our proposed method achieved 100% classification accuracy. These findings suggest that K descriptor is a very effcient shape descriptor for rectangle-like classification.