基于样本旋转的生成困难样本的深度度量学习方法OA北大核心CSTPCD

Existing deep metric learning methods guide efficient training of the model by constructing hard sample generation meth-ods.The hard sample generation methods based on algebraic computation have the advantages of simplicity and efficiency.Howev-er,such methods lack consideration of the overall data distribution,resulting in strong randomness of the generated samples and slow convergence of the model.To address this problem,we propose a hard sample generation method based on sample rotation by rotating positive samples in a triad to the reverse extension of the line connecting the anchor point and the class center on the axis of the class to which they belong,and give a new loss function to construct a deep metric learning model(RHS-DML)for generating hard samples based on sample rotation,effectively improving the training efficiency of the model.Experiments on image retrieval were conducted on the Cars196,CUB200-2011,and Stanford Online Products datasets,and compared with the symmetric sample generation method in algebraic computing.The results showed that the retrieval performance of the algorithm proposed was 2.4%,0.7%,and 1.4%higher than the symmetric sample generation cost method on the three datasets,respectively.