Multi-granularity feature learning network for deep hashing

Abstract With the ever-increasing growth of massive high-dimensional data, deep learning to hash technology has been widely used for approximate nearest neighbor search on large-scale datasets, due to its remarkable efficiency and retrieval performance. In this paper, we propose a novel supervised deep hashing method, named Multi-granularity Feature Learning Hashing (MFLH), to learn compact binary descriptors. Specifically, we design an end-to-end trainable network to jointly learn feature representations and hash codes, in which a global stream and a local stream are responsible for learning feature representations with different granularities, and a hashing stream is devoted to encoding multi-granularity features into binary codes. In the local stream, a Cyclic Shift Mechanism (CSM) strategy is developed to assist mining more discriminative local fine information. In the meantime, an approximate sign activation function, which can be used for continuous optimization, is introduced to reduce quantization error. Furthermore, an improved variant of the triplet loss is presented to enhance the representation of pair-wise similarity for hash codes. Extensive experiments demonstrate that our proposed method significantly outperforms state-of-the-art hashing methods on the benchmark datasets, thereby verifying the effectiveness of our approach. Source code is provided for reproducibility.