Deep Neighborhood-Aware Hashing via Class-Center Guiding for Multi-Label Image Retrieval

Learning to hash effectively addresses the challenges posed by massive data due to its low storage cost and fast search speed. Under multi-label scenarios, pair-wise hashing is typically determined by roughly counting the number of shared labels. However, this approach overlooks the intrinsic semantic information, leading to significant intra-class variations and ambiguity in neighborhood relationships. In this paper, by learning semantic class centers, the novel Deep Neighborhood-aware Hashing framework, termed DNaH, is designed to generate a highly discriminative Hamming space, thereby preserving binary codes with rich neighborhood information. Specifically, by building cross-block non-local information transmission, the simplified and effective hierarchical transformer encoder is introduced into the deep hash stream to capture distinguishing semantic descriptors, which achieves the distinctive representation capability of sophisticated networks while effectively promoting convergence efficiency. The binary representations from the Label Network are viewed as the semantic centers of categories, in which the class-guided loss is proposed to encourage the learned hashing codes to be close to the corresponding class centers. Moreover, by maximizing the semantic similarity between the relative position of binary-like representations and corresponding centroids, the position quantization strategy is developed to decrease projection errors and significantly enhance the distinguishability of the discrete codes. Comprehensive experiments on three benchmark datasets reveal that the DHaN framework delivers better performance in visual similarity search compared to several leading deep hashing approaches. The source code is available at: https://github.com/QinLab-WFU/DNaH