Abstract
The multiple-assignments approach alleviates the quantization error and enhances the accuracy of the Content-Based Image Retrieval (CBIR) systems. It aims to hard assign each feature vector to k-nearest visual words. However, during the matching step, the k-nearest visual words are used independently and ignore the significant of the best visual word. In this paper, we present our CBIR system which encapsulates several approaches such Hamming embedding, soft-assignment, multiple-assignments and graph fusion. We particularly focus on the multiple-assignments strategy. We propose an efficient soft multiple-assignments strategy to highlight the best k-nearest visual word. To this end, we explore the SOM topology which proved its performance in so doing. Moreover, we use graph fusion approach to fuse multi-features ranking lists. Extensive experiments are conducted on Holiday and Ukbench public datasets. The experimental results are promising and outperform the state-of-the-art CBIR systems. In fact, we have reached a mAP = 85.6 on Holidays dataset and a KS score of 3.87 on Ukbench dataset.