An Integration of Shape Context and Semigroup Kernel in Image Classification

Shape context is a rich descriptor for shapes and can be exploited to find pointwise correspondences between shapes, and thereby to obtain shape alignment by thin plate spline (TPS). It is invariant under scaling and translation and robust under small geometrical distortions and presence of outliers. These features will supply a gap of the defect of semigroup kernel for its weakness in dealing with the deformation of the image. This paper integrates these two methods by defining a new kernel on shapes and images which is the combination of the shape distance from shape context and image similarity from semigroup kernel. Experiments of SVM classification on handwritten digits showed that it outperforms other existing kernels and the result of the data visualization exhibited another successful application of this new kernel.