Joint Deep Hashing and Object Detection for Image Retrieval

Abstract

In this paper, we propose a new hashing method combining deep hashing and object detection to address image retrieval. The most off-the-shelf hashing methods encode one image to one hashing code. While the images are multi-object in our natural world, it's difficult to express one image by one hashing code accurately if there exist many categories of objects in these images. Inspired by the rapid development of CNN and object detection methods, we propose a novel Deep Object Hashing (DOH) method to learn compact similarity preserving binary codes for the huge multi-object image dataset. Firstly, we detect the localization of objects in the feature map. Then we use the hashing network to extract binary code of the object with the highest confident score to represent the same category objects. Therefore, one new-coming image can be easily encoded to different binary representation depending on the number of categories in the image. Extensive experiments on two benchmarks PASCAL VOC 2007 and COCO show the promising performance of our method compared with the state-of-the-art methods. Codes and models will be released to promote deep hashing.