New image distance and its application in object recognition

This paper presents a new distance measure for image matching based on local Kullback-Leibler divergence, which we call Image Kullback-Leibler Distance (IKLD). Unlike traditional methods, IKLD takes account into not only the spatial relationships of pixels, but also the structure information around pixels. Therefore, it is robust enough to small changes in viewpoint. In order to illustrate its performance, we imbed it into support vector machines for view-based object recognition. Experimental results based on the COIL-100 show that it outperforms most existing techniques, such as traditional PCA+LDA (principal component analysis, linear discriminant analysis), non-linear SVM, Discriminant Tensor Rank-One Decomposition (DTROD) and Sparse Network of Winnows (SNoW).