Label Enhancement Hashing Induced by Class Prototypes for Domain Adaptive Retrieval

Domain adaptive retrieval (DAR) aims to perform effective cross-domain retrieval by transferring knowledge from the source domain to the target domain and reducing domain distribution discrepancy. However, the target domain often lacks annotations, and using pseudo-labels generated for target data may reduce retrieval accuracy due to their inaccuracy. Additionally, although features from the same class may differ across domains, they share global invariant information crucial for identifying samples in different domains. To address these challenges, this paper proposes a novel DAR method, Label Enhancement Hashing induced by Class Prototypes (LEHC). The approach first projects source and target domain features into a common subspace to reduce feature redundancy and domain discrepancy. Then, a label enhancement strategy is applied to convert discrete labels into continuous values, enriching semantic information. Orthogonal class prototypes in the common subspace capture global invariant information and induce relationships between enhanced labels and sample features. Finally, asymmetric similarity preserving is proposed, which retains both the pairwise similarity of the samples and the enhanced label information into hash codes. Experimental results on various benchmark datasets validate the effectiveness of LEHC, showing its superior performance in domain adaptive retrieval tasks.