Self-Organized Neighborhood Preserving Projections Using Information

One of the main problem for object recognition in computer vision is the transformation of an object (e.g. by shift, rotate, scale or depth deformation) within several views of that object. Object recognition in human vision might be explained by special temporary mappings where an unknown object is compared to stored object prototypes of standardized size and view. How does this mapping work? One solution to this problem is to use dynamic links for a mapping between the input object and the target object. In this paper, we propose a new neighborhood preserving self-organizing map (SOM) algorithm, based on information theory. We present a new similarity measure based on the distance of local pattern entropies. Experimental results illustrate the robustness and efficiency of our new algorithm.