Deep Top Similarity Preserving Hashing for Image Retrieval

As a kind of approximate nearest neighbor search method, hashing is widely used in large scale image retrieval. Compared to traditional hashing methods, which first encode each image through hand-crafted features and then learn hash functions, deep hashing methods have shown superior performance for image retrieval due to its learning image representations and hash functions simultaneously. However, most existing deep hashing methods mainly consider the semantic similarities among images. The information of images' positions in the ranking list to the query image has not yet been well explored, which is crucial in image retrieval. In this paper, we propose a Deep Top Similarity Preserving Hashing (DTSPH) method to improve the quality of hash codes for image retrieval. In our approach, when training the convolutional neural network, a top similarity preserving hashing loss function is designed to preserve similarities of images at the top of the ranking list. Experiments on two benchmark datasets show that our proposed method outperforms several state-of-the-art deep hashing methods and traditional hashing methods.