Salient feature point detection for image matching

摘要

A saliency based feature point detector is proposed, based on a decision-theoretic formulation of saliency. The saliency of an image region is defined to be the Kullback-Leibler (K-L) divergence between the conditional probability density function (pdf) for the matching regions and a background pdf. These pdfs are modeled by elliptically symmetric distributions (ESDs). We improve the ESD models by reducing the number of parameters without any significant degradation in the modeling of image regions. Experimental results from the Middlebury stereo dataset show that 1) the accuracy of estimates of saliency is increased and 2) fewer computations are required. It is also verified that the saliency of a region can be viewed as a measurement of how suitable the region is for image matching. In the Middlebury stereo dataset, salient regions are dense, and a promising matching rate is achieved.