摘 要：红外图像与可见光图像的跨模态行人重识别是一个备受关注的任务问题。与传统可见光条件下的单一模态行人重识别问题相比，跨模态行人重识别还包含成像原理导致的模态差异的难点。现有方法仅从特征级别增加约束，忽略了模态差异的独特性。为此，文章将分别处理模态差异和外观差异，使用循环生成对抗网络减小模态差异，使用度量学习降低外观差异。同时，还提出一个跨模态的图像融合模块。在 RegDB 和 SYSU-MM01 数据集上的实验结果证明了该方法的有效性。
关键词：生成对抗网络；行人重识别；跨模态
中图分类号：TP391.41 文献标识码：A 文章编号：2096-4706（2020）04-0107-03
Research on Cross Modal Pedestrian Recognition Based on Generated Countermeasure Network
FENG Min，ZHANG Zhicheng，LYU Jin，YU Lei，HAN Bin
（Nanjing Nari information and Communication Technology Co.，Ltd.，Nanjing 210003，China）
Abstract：The problem of cross modal pedestrian recognition of infrared image and visible image is a task of great concern. Compared with the single modal pedestrian recognition problem under traditional visible light，cross modal pedestrian recognition also involves the difficulty of modal difference caused by imaging principle. The existing methods only add constraints from the feature level，ignoring the uniqueness of modal differences. For this reason，the paper deals with the modal difference and appearance difference respectively，uses the cyclic generation countermeasure network to reduce the modal difference，and uses the metric learning to reduce the appearance difference. At the same time，a cross modal image fusion module is proposed. Experimental results on RegDB and SYSUMM01 datasets show the effectiveness of the proposed method.
Keywords：generate countermeasure network；pedestrian recognition；cross modal
参考文献：
[1] WANG Z，HU R M，CHEN C，et al.Person Reidentification via Discrepancy Matrix and Matrix Metric [J].IEEE Transactions on Cybernetics，2017：1-15.
[2] WANG Z，HU R M，LIANG C，et al.Zero-Shot Person Reidentification via Cross-View Consistency [J].IEEE Transactions on Multimedia，2015，18（2）：260-272.
[3] WEI L H，ZHANG S L，GAO W，et al.Person Transfer GAN to Bridge Domain Gap for Person Re-Identification [C]//Proceedings of Conference on Computer Vision and Pattern Recognition，2018.
[4] WU A C，ZHENG W S，YU H X，et al.RGB-Infrared Cross-Modality Person Re-Identification [C]//2017 IEEE International Conference on Computer Vision （ICCV）.IEEE，2017.
作者简介：冯敏（1972-），男，汉族，江苏如皋人，工程师，硕士研究生，主要研究方向：视频图像处理。