Local feature matching using entropic graphs

We present a general framework for image discrimination based on identifying small, localized differences between images. Our novel matching scheme is based on an alternate information divergence criterion, the Renyi alpha-entropy. The minimum spanning tree (MST) is used to derive a direct estimate of alpha-entropy over a feature set defined by basis features extracted from images using independent componenet analysis (ICA). The MST provides a stable unbiased estimate of local entropy to identify sites of local mismatch between images. Sub-image blocks are ranked over a set of local deformations spanning small image regions. We demonstrate improved sensitivity to local changes for matching and registration and provide a framework for tracking features of interest in images.