Image classification via learning dissimilarity measure in non-euclidean spaces

This paper presents a novel image classification scheme, named high order statistics based maximum a posterior (HOS-MAP). To bridge the gap between human judgment and machine intelligence, this framework first builds dissimilarity representations in a modified pseudo-Euclidean space. Then, the information of the dissimilarity increments distribution of each category is achieved based on high-order statistics (HOS) of triplets of neighbor points for each image data, as opposed to typical pair-wise measures, such as the Euclidean distance. Finally, a maximum a posteriori (MAP) algorithm with the information of Gaussian Mixture Model and triplet-dissimilarity increments distribution is adopted to estimate the relevance of each category in the database for each input new image. Experimental results on a general-purpose image database demonstrate that effectiveness and efficiency of the proposed HOS-MAP scheme.