Deep network-based image coding for simultaneous compression and retrieval
Qingyu Zhang, Dong Liu, Houqiang Li · 2017
Images on the Internet are usually in the form of compressed bitstream to save storage. To fulfill content-based image retrieval (CBIR), image features are also required to be stored in binary form. Can the bitstream of images and image features be unified and further condensed? Is it possible that the same binary code serves for compression and retrieval simultaneously? To address this problem, we make preliminary studies on a deep network-based image coding scheme in this paper. We first train a deep network for compressing images into bitstream, and then train another deep network for extracting image features as binary vector. We then combine the above two networks, and finetune the combined network using triplets of images for the task of CBIR. Our experimental results show that the proposed scheme achieves a compression ratio of 5.3 for 32×32 thumbnails, outperforms JPEG at similar compression ratios, and the resulting code is directly available for CBIR. Our work indicates a promising direction of simultaneous image compression and retrieval.