An Integrative Codebook for Natural Scene Categorization

Natural scene categorization (NSC) is an important and challenging task. Several state-of-the-art NSC systems use a codebook of visual terms to characterize images with the statistic of visual word counts. However, some kind of codebook generally tends to be more favorable for characterizing a special scene category, which takes either flat property or salient one. To obtain the good tradeoff performance between the block-based codebook and salience- based one, an integrative codebook for image representation is proposed in this paper. Moreover, we exploit a modified Latent Dirichlet Allocation (LDA) model for learning the code words and themes of natural scene categories. The categorization accuracy of the model based on our integrative codebook achieves about 73.67% on a large set of 15 categories of complex scenes.