Bag-of-words image representation based on classified vector quantization

In recent years, bag-of-words image representation has been shown to be a good base for the representation of images. To build this representation, visual vocabulary is necessary, which is typically constructed by vector quantization of local image features. However, current approaches typically consider basic vector quantization for vocabulary construction, which limits the practical applicability of bag-of-words representation for many visual tasks. In this paper, we propose to build image representation based on classified vector quantization. A major effect of using classified vector quantization is that it can achieve a more compact visual vocabulary. The traditional visual vocabulary for image representation is compared against our method for scene categorization task on well-known dataset: 15 natural scenes. We demonstrate that our method can greatly reduce the complexity of vocabulary construction, and at the same time it can improve the categorization performance for scene categorization. Moreover, we show that our method reaps higher categorization performance when decreasing the vocabulary size.