Improved bags-of-words algorithm for scene recognition

This paper proposes an effective method to scene recognition based on bags-of-words (BoW) algorithm. Current scene classification methods usually treat all the codewords equally important when using BoW histogram to represent an image. This assumption, however, does not comply with many real-world conditions as different codewords usually have different discriminating power when representing different scene categories. Considering this, this paper proposes an effective technique to perform scene recognition. It first uses k-means algorithm to construct a codebook, in addition with an occurrence matrix. The importance of each codeword for each scene category is then estimated based on the above cooccurrence matrix. Finally this discrimination information is incorporated into the original BoW histogram of the image and produces a new BoW histogram. Support vector machine (SVM) is used to train these BoW histograms. Experimental results on the 15 scene dataset show that the proposed method is very effective compared with state-of-art works.