Compressive image retrieval with modified images

Many existing image retrieval methods cannot deal with modified query images because modified parts may largely change the original color, shape and texture information. Since every pixel's information of an image has possibility to be changed by editors, local features become less credible. In our method, a very sparse measurement matrix is applied for the compressing procedure. The retrieval database can be compressed effectively by off-line. In order to ensure the retrieval efficiency, this method is not data-driven and does not depend on learning methods. The benchmark includes 5,000 images modified by people according to their preferences. The experimental results show that our method is more efficient than many descriptors on solving this visual similarity measure problem. Even for the images which are modified with complex effects, our method can perform favorably in contrast with other methods. The main contributions of this paper can be summarized as: 1) We propose a globally compressed image descriptor which is simple and effective for solving this visual similarity evaluation problem; 2) We build a benchmark for evaluating the descriptor's performance on solving this problem by comparing with other methods.