Shared Latent Characteristic Anchored Hash Codes Generation for Efficient Fine-Grained Image Retrieval

Hashing-based fine-grained image retrieval (FGIR) is a promising solution for large-scale domain-specific data retrieval, yet it faces a fundamental contradiction between discriminative feature learning and compact binary codes generation. Existing methods often rely on complex feature extraction modules to improve fine-grained discriminability. While effective with long hash codes, they usually suffer from significant performance degradation under short code lengths due to the inherent limitations of Hamming space representation. To address this critical issue, this paper proposes a novel shared latent Characteristic Anchored Hash (CAH) codes generation method. Instead of relying on increasingly complex feature extraction networks, our approach directly addresses the feature contradiction inherent in the feature-to-hash mapping process. The core of the proposed method is the introduction of learnable Characteristic-vectors (C-vectors), which are explicitly defined as anchors in the feature space to represent the shared latent characteristics of semantically similar samples. On this basis, a characteristic matching loss and C-vector anchored feature refinement mechanism is designed for C-vectors and feature vectors optimization before hash codes generation. Furthermore, to enhance the discriminative power of the final hash codes, a cross-layer semantic information transfer module and an object-constrained multi-region augmentation strategy is designed, which improve fine-grained feature learning without introducing computational overhead at inference time. Comprehensive experiments demonstrate that our proposed method significantly outperforms state-of-the-art techniques in both retrieval accuracy and efficiency, establishing a new and promising direction for hashing-based FGIR.