An improved feature fusion algorithm for image retrieval
Shilin Che, Dongfang Chen, Xiaofeng Wang · 2016
Single feature extract method for image retrieval usually cannot get good results and it is hard to find a stable fusion method when across image datasets because the image descriptors usually describe image from single aspect. For this pragmatic issue, an improved FCTH-BoVW feature fusion algorithm called robust FCTH-BoVW is proposed in this paper. Traditional feature fusion methods almost use liner weight method, but this method will be easily affected by the weights, and not easy to find the optimal weights. In this paper, FCTH feature is used as the global feature and BoVW feature as the local feature, and use k-reciprocal nearest neighbors method to fuse them. FCTH is an efficient descriptor and contains both color and texture information, in order to emphasize the center area as main region of interest, an annular weight method is proposed to improve the global feature. A soft assignment method is adopted to quantize the visual words for the improvement of the local feature. This method enhances the description ability of both global and local features. We find those improvements are quite practical, the results show that this fused method gets a higher retrieval precision, and we also find it is a stable image retrieval method for varieties of image datasets.