Prototype-Aligned and Domain-Aware Zero-Shot Hashing
Zero-Shot Hashing(ZSH)methods typically rely on category attributes to transfer supervised knowledge from seen to unseen classes,and efficiently retrieve unseen category images.However,obtaining category attributes requires additional computing resources,and a heterogeneous gap exists between visual features and category attributes.Additionally,existing methods ignore the strong bias problem,leading the model to mistakenly identify seen class samples as unseen classes,thereby reducing retrieval accuracy.Moreover,ZSH encounters challenges in preserving semantic consistency between hash codes and the original data,as well as in achieving discrete optimization of hash codes.To address these challenges,a prototype-aligned and domain-aware ZSH method is proposed.It eliminates the dependence on special supervision knowledge,such as category attributes,thus reducing the cost of attribute annotation and avoiding the impact of cross-modal heterogeneous gaps.Specifically,the method first calculates prototypes for each class in the Hamming space and thereafter aligns the hash codes with the class prototypes to learn semantically consistent hash codes.To mitigate the quantization errors caused by relaxation strategies,a discrete optimization algorithm is proposed to address the discrete constraints of hash codes and achieve linear computational complexity.Moreover,a domain-aware strategy is used to separate the samples from the source and target domains,thereby alleviating the strong bias problem.The experimental results on the aPY,AWA2,and ImageNet datasets demonstrate that the retrieval accuracy of this method is improved by 2.6,9.4,and 14.9 percentage points compared to the optimal values in the comparison method,respectively,and the training time is significantly reduced compared to most baseline methods.
国家科技期刊平台
NSTL
万方数据
