Image Classification Using Sparse Coding and Spatial Pyramid Matching

Recently, the Support Vector Machine (SVM) using Spatial Pyramid Matching (SPM) kernel has achieved remarkable successful in image classification. The classification accuracy can be improved further when combining the sparse coding with SPM. However, the existing methods give the same weight of patches of SPM at different levels. Clearly the discriminative powers of SPM at different levels are distinct and there are correlation relationships among the sparse coding bases vectors, which usually have negative influence on the classification accuracy. This paper assigns different weights to the patches at different levels of SPM, and then proposes a new spatial pyramid matching kernel. Furthermore, the Principle Component Analysis (PCA) is employed to reduce the dimension of the feature vectors in order to decrease correlation among vectors and speed up the SVM training process. The preprocessing can enhance the discriminative ability of the new kernel as well. Experiments carried out on Caltech101 and Caltech256 datasets show that the new SPM kernel outperforms the existing methods in terms of the classification accuracy.