GSPH: Granularity-guided Saturation Proxy Hashing with Self-adaptive Feature Importance for image retrieval

Deep supervised hashing, learning compact binary codes under label supervision through deep neural networks, is mainstream for large-scale image retrieval. However, existing methods still face two limitations. First, existing methods typically adopt a category-agnostic optimization mechanism for intra-category compactness, which neglects the varying intra-category diversity caused by category granularity and thereby limits model generalization. Moreover, these methods assign equal importance to all extracted features when feeding them into the hash layer, neglecting the incorporated irrelevant background information during feature extraction, reducing hash codes discriminative power. To address these, we propose a novel Granularity-guided Saturation Proxy Hashing (GSPH) framework. First, we introduce a Granularity-guided Saturation Proxy (GSP) loss that employs category-specific Hamming balls to achieve an optimization saturation mechanism: samples within their Hamming balls are deemed saturated and thus terminated from optimization, while those outside continue to be optimized toward their proxy. Additionally, GSP establishes a negative boundary with fixed margin outside each category’s Hamming ball, effectively ensuring inter-category separability. Second, we develop a Self-adaptive Feature Importance (SFI) module that employs gating mechanism to regulate feature importance during feature extraction, ensuring more discriminative representations. Extensive experiments on four benchmark datasets demonstrate that our method consistently outperforms existing methods. • Granularity-guided Saturation Proxy loss for intra-category diversity preservation. • Self-adaptive Feature Importance module for background interference suppression. • The proposed method outperforms existing methods on four benchmark datasets.