Improving Image Classification Quality Via Dissimilarity Measure In Non-Euclidean Spaces

Abstract

This paper proposes an image classification scheme by learning the dissimilarity measure in non-Euclidean spaces. Specifically, the dissimilarity representations of samples from a pseudo-Euclidean space are first constructed; then, the dissimilarity increment distribution information of each category is achieved with respect to the high-order statistics of triplet-neighbor points for each image; finally, a maximum a posteriori algorithm fused with the Gaussian Mixture Model and triplet-dissimilarity increments distribution is utilized to estimate the relevance of each image category with each input image. Experimental results conducted on a general image database demonstrate the effectiveness of the proposed scheme.