Local Stretch Similarity Measure And Auto Query Expansion For Re-Ranking Of Image Retrieval

Using local neighbors (contextual information) to measure similarities between images is an effective re-ranking method in image retrieval, which has many advantages (e.g., low time complexity, easy promotion for new images). However, traditional methods usually cannot take neighbors into account comprehensively for the influence of noises (the more neighbors chosen, the more noises involved). To solve this problem, we propose Local Stretch Similarity Measure algorithm (LSSM). LSSM chooses multiple layers of neighbors to measure similarities, through which more contextual information can be considered and less noises will be introduced. Furthermore, we propose Auto Query Expansion (AQE) re-ranking method to transform the original single-query problem to a multi-query problem. By means of AQE, the robustness of LSSM can be enhanced. Extensive experiments are conducted on Corel-1K, Corel-10K and UK-bench datasets. Experimental results validate that our methods outperform other state-of-the-art methods.