Variable-Weight and Multi-index Similarity Measure for Feature Fusion

The BoW model is a mainstream in image retrieval, and is employed in many state-of-the-art retrieval process. However, there are two main problems in this model: semantic gap caused by image quantification, and insufficient feature discriminative power. Aiming to handle the problems of semantic gap and feature fusion. This paper propose to choose multiple features to describe interest points and make evaluations for them so that the better judgement result is obtained. And then, we will generate dictionaries for each features and their cartesian product. As a result, a coupled index for all images will be built. Finally, we utilize variable-weight for the scores from those three dictionaries to obtain the total score for each image in datasets. By this means, we narrow semantic gap and implement feature fusion. At the same time, we also prove that previous fusion approaches are the special cases of our proposed approach by changing the weights. The experimental results on two public datasets show the benefit of our approach, specially at the aspect of recognition accuracy.