Image classification using tensor representation

Ziming Zhang, Syin Chan, Liang-Tien Chia

Abstract: We propose a new approach to exploit the different discriminability of image features at different scales simultaneously. By modifying the Bag-of-words model, we represent an image as a matrix whose elements are the occurrences of a set of codewords within different scale ranges. In this way, we can represent an image collection using a 3rd-order tensor. Then a new classification method, tensor-pLSA, which is an extension of Probabilistic Latent Semantic Analysis (pLSA), is introduced to classify these images based on this tensor representation. Finally, we compare the tensor representation with the original matrix representation to show the effectiveness of our approach.