A novel method for image classification based on bag of visual words

• We apply the salient region extraction to BOW model, it can produce more representive visual words and avoid the disturbance of complex background information.
• The visual words topological structure is able to integrate into global spatial information of visual words to improve BOW model.
• Delaunay triangulation method is employed to integrate into local spatial information of visual words to improve traditional BOW model.

Bag of words (BOW) model is widely applied in image classification. The traditional BOW neglects the spatial information and object shape information, which fails to distinguish between image features absolutely. In this paper, we combine the salient region with visual words topological structure. It not only can produce more representative visual words, but also can avoid the disturbance of complex background efficiently. Firstly, the salient regions are extracted and the BOW model is built on salient regions. Secondly, in order to describe the characteristics of the image more accurately and to resist the influence of background information, the visual words topological structure and Delaunay triangulation method are employed, which is able to integrate into the global and local information. The performance of the proposed algorithm is tested on several datasets, and compared with other models. The experiment results seem to demonstrate that the proposed method provide a higher classification accuracy.