Image classification based on local spatial pyramid Kernel

This paper develops a new algorithm based on Bag-of-Word to reflect spatial relationship of objects for visual object categorization. Beyond existing spatial pyramid for image representation, our contributions are the following: 1) we propose a combinational detector based on Maximally Stable Extremal Regions detector and Hessian-Laplacian detector to extract more discriminative features; 2) for object categorization, we propose local spatial pyramid kernel which encodes spatial information of objects and is robust to spatial transformation of objects. The proposed approach is evaluated on two popular image databases: Xerox7 and ImageNet. Experimental results demonstrate that our method gives better recognition rates in comparison with spatial pyramid.