Integration of Reconstruction Error Obtained by Local and Global Kernel PCA with Different Role

Abstract

This paper presents a scene classification method using the integration of the reconstruction errors by local Kernel Principal Component Analysis (KPCA) and global KPCA. There are some methods for integrating local and global features. However, it is important to give obvious different role to each feature. In the proposed method, global feature with topological information represents the rough composition of scenes and local feature without position information represents fine part of scenes. Experimental results show that accuracy is improved by using the reconstruction errors obtained from the different point of views. The proposed method is much better than only local KPCA, global KPCA and linear Support Vector Machine (SVM) of bag-of-visual words with the same basic feature. Our method is also comparable to conventional methods using the same database.