文献类型：期刊文献

英文题名：Cross-Domain Attention and Center Loss for Sketch Re-Identification

作者：Zhu, Fengyao[1];Zhu, Yu[1,2];Jiang, Xiaoben[1];Ye, Jiongyao[1]

机构：[1]East China Univ Sci & Technol, Sch Elect & Commun Engn, Shanghai 200237, Peoples R China;[2]Shanghai Engineer & Technol Res Ctr Internet Thin, Shanghai 200032, Peoples R China

年份：2022

卷号：17

起止页码：3421

外文期刊名：IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY

收录：;EI(收录号：20224112870754);WOS:【SCI-EXPANDED(收录号:WOS:000864326600004)】；

基金：This work was supported in part by the Natural Science Foundation of Shanghai under Grant 19ZR1413400, in part by the National Natural Science Foundation of China under Grant 82170110, and in part by the Science and Technology Commission of Shanghai Municipality under Grant 20DZ2254400.

语种：英文

外文关键词：Transformers; Task analysis; Face recognition; Feature extraction; Representation learning; Image color analysis; Cameras; Sketch re-identification; cross-domain attention; domain-invariant feature; center loss

摘要：Matching all RGB photos of the target person in the gallery database with the full-body sketch image drawn by the professional is defined as Sketch re-identification (Sketch Re-id). The big gap between the sketch domain and RGB domain makes Sketch Re-id challenging. This paper addresses the problem by proposing a new framework to obtain domain-invariant features, which uses CNN as the backbone. To make the model focus more on the regions related to the sketch image in the RGB photo, we propose a novel cross-domain attention (CDA) mechanism. It uses different ways of splitting feature maps in its two branches and calculates the relationship between different parts in the sketch images and RGB photos. Moreover, we designed the cross-domain center loss (CDC), which breaks through the limitations that datasets need to be in the same domain in the traditional center loss. It effectively reduces the gap between two domains and makes the features with the same ID closer. The experiment is performed on the Sketch Re-id dataset. Each person has one sketch image and two RGB photos. To evaluate the generalization, we also experimented on two popular sketch-photo face datasets. The result in the Sketch Re-id dataset shows the model performs 3.7% higher than the previous methods. And the result in the CUHK student dataset performs 0.38% higher than the state-of-the-art methods.