Category Related BoW Model for Image Classification

摘要

Traditional Bag of Words model is constructed by clustering all categories of training images' key points. The huge amounts of points lead to low efficiency in model construction approach. It also ignores the content discrepancy of different category images in the model. In this paper, single image SIFT points are firstly processed by a noise and redundancy reduction algorithm to improve the model construction efficiency. Then the SIFT points from the same category images are clustered to produce the Category Related Visual Words (CRVW). To determine proper number of CRVW for certain category, SIFT points of each image is located in a regularly segmented block, and then the location distribution entropy is calculated. The category with more entropy is assigned with more visual words. Finally all CRVW of different categories are connected and the Category Related Bag of Words (CR-BoW) model is constructed. Experiments carried on Corel IK image set indicate that the proposed model can promote the classification power by 4-7 percentage points and reduce the model construction time cost to about 1/4 of the traditional method.