Scene Classification Using Candidate Classes Selection with Particle Filter and Criterion Mining for Final Decision with AdaBoost

This paper presents a scene classification method using criterion mining and adaptive integration. Since scene classification requires scene composition and shift-invariant similarity of fine parts, the latter two are represented by global and local Kernel Principal Component Analysis (KPCA), respectively. In addition, the reconstruction errors obtained with either KPCA are integrated adaptively with a particle filter since it is not known which KPCA is effective for which test sample. Although adaptive integration can be used to select several candidate classes, it can not provide a satisfactory criterion for the final decision. Therefore, criterion mining is performed with AdaBoost by using ten measures obtained by adaptive integration. The use of criterion mining entails an improvement of about $10 %$ in comparison with the adaptive integration method. The proposed method achieves a score of 91.06 % and provides state-of-the-art performance in an evaluation experiment using a 13-scene database.