Semantic Consistency-Enhanced Refined Hashing for Fine-Grained Image Retrieval

Fine-grained hashing for image retrieval aims to convert images into semantic-preserving binary hash codes for efficient retrieval from large-scale databases with many similar subcategories. Fine-grained images exhibit large intra-class variability and small inter-class differences, making fine-grained hashing particularly challenging. Although recent methods have made progress, their retrieval accuracy is still limited by two main issues: 1) learning feature representations with strong discriminability, and 2) learning hash codes with high semantic consistency. To address these problems, we propose a novel hashing method called Semantic Consistency-Enhanced Refined Hashing (SCERH), which includes two key components: Classification-Guided Feature Refinement (CGFR) and Semantic Consistency Constraint (SCC). Specifically, to learn more discriminative feature representations, CGFR constructs a specialized feature mask for each category and adaptively selects the appropriate feature mask to refine the feature representation. CGFR not only preserves beneficial semantic information but also suppresses irrelevant noise. Moreover, to enhance the semantic consistency of the hash codes, SCC minimizes the semantic discrepancy between the feature representation and the corresponding hash code, ensuring that the hash code more comprehensively captures and preserves the semantic information in the feature representation. Experimental results demonstrate that SCERH significantly outperforms several existing competitive fine-grained hashing methods in terms of retrieval accuracy, confirming its effectiveness and superiority.