Image Classification with a Novel Semantic Linear-Time Graph Kernel

Image classification is currently a vital and challenging topic in computer vision. Although it has been achieved many classification algorithms so far, the classification of natural images still remains great difficulties in image processing. In this paper, we propose a semantic linear-time graph kernel for image classification. Each image is represented by a graph and the vertex of each graph corresponds to a segmented region. The semantic graph kernel can be used to compute similarities of graphs efficiently. The performance of the proposed graph kernel is demonstrated via a comparative performance assessment which is carried out on two large natural image databases. Experimental results show both the efficiency and precision of the proposed algorithm are better than those of histogram-based image classification which are carried out on Support Vector Machines (SVMs).