可判别性标签语义指导的域适应检索

Discriminative Label Semantic Guidance Learning for Domain Adaptive Retrieval

从不同领域准确检索相似的对象,域适应检索解决了信息检索中的域偏移问题.然而,现有的方法仍然存在两个问题:a)忽略了类结构差异造成的域偏移(跨域和域内不同类的距离较近);b)忽略了特征与标签之间的语义差异.为了解决上述两个问题,本文提出了一种高效的可判别性标签语义指导学习(DLSG)方法.该方法探索源域和目标域的类结构,通过拉大不同类的距离使得类别更具有判别性.然后通过标签语义指导学习(LSG)来增强特征的标签语义,以提高学习的有效性.此外,动态对齐边缘分布和条件分布,以减少域差异.最后,采用两步哈希策略生成高质量的哈希码.在多个跨域检索数据集上的实验表明,DLSG的性能得到了提高.

To accurately retrieve similar objects from different domains,domain adaptive retrieval solves the domain shift of information retrieval.However,existing methods still have two problems:a)they ignored the domain shift caused by class structure discrepancy(in cross-domain and intra-domain,the distance between different classes is close);b)they ignore the semantic discrepancy between features and labels.To solve the above two issues,we propose an efficient Discriminative Label Semantic Guidance Learning(DLSG).This method explores the class structure of the source domain and target domain,and makes the class more discriminative by maximizing the distance mean of one class from means of other classes.Second,Label Semantic Guidance learning(LSG)is used to enhance the label semantics of features during transfer learning.In addition,dynamically align margin distribution and condition distribution to reduce domain discrepancy.At last,a two-step hashing strategy is used to generate high-quality hash codes.Experiments on multiple cross-domain retrieval datasets demonstrate that the advanced performance is gained and the retrieval speed of DLSG is quick.