Scene classification using adaptive integration of reconstruction errors

This paper proposes an adaptive integration method of reconstruction errors obtained from different point of view. There are some methods for integrating local and global processing. However, integration parameters are fixed for all test samples though effective parameters are different for every sample. Therefore, we select adaptively the parameters from only a test image. In static image recognition, the information of an input image is not changed. However, the posterior probability in weight space is changed for every test image. Thus, the posterior probability in weight space is estimated by a particle filter, and effective weight with high probability for the test sample is selected. Experimental results demonstrate the effectiveness of our approach.