Kernel Methods for Extracting Local Image Semantics

This paper describes an investigation into using kernel methods for extracting semantic information from images. The specific problem addressed is the local extraction of ‘man-made’ vs ‘natural’ information. Kernel linear discriminant and support vector methods are compared to the standard linear discriminant using a multi-level hierarchy. The two kernel methods are found to perform similarly and significantly better than the linear method. An advantage of the kernel linear discriminant over the SVM method is that accurate class-conditional density estimates can be determined at each level allowing posterior estimates of class membership to be evaluated. These probabilistic outputs give a principled framework for combining results from a number of semantic labels.