Category-specific incremental visual codebook training for scene categorization

In this paper, we propose a category-specific incremental visual codebook training method for scene categorization. In this method, based on a preliminary codebook trained from a subset of training samples, we incrementally introduce the remaining training samples to enrich the content of the visual codebook. Then, the incremental learned codebook is used to encode the images for scene categorization. The advantages of the proposed method are (1) computationally efficient comparing with batch mode clustering method; (2) the number of visual words is determined automatically in the incremental learning procedure; (3) scene categorization performance is improved using the enriched codebook comparing with using the codebook trained from a subset of training samples. The experimental results show the effectiveness of the proposed method.