Asymmetric Similarity-Preserving Discrete Hashing for Image Retrieval

Hashing methods have been widely studied in the image research community due to their low storage and fast computation. However, generating compact hash codes is still a challenging task. In this paper, we propose a novel Asymmetric Similarity-Preserving Discrete Hashing (ASPDH) method to learn compact binary codes for image retrieval. Specifically, the pairwise similarity matrix is approximated in the asymmetric learning manner with two different real-valued embeddings. In addition, ASPDH constructs two distinct hash functions from the kernel feature and label consistency embeddings. Therefore, similarity preservation and hash code learning can be simultaneously achieved and interactively optimized, which further improves the discriminative capability of the learned binary codes. Then, a well-designed iterative algorithm is developed to efficiently solve the optimization problem, resulting in high-quality binary codes with reduced quantization errors. Extensive experiments on three public datasets show the rationality and effectiveness of our proposed method.