视频监控中跨域图像重构与识别

Person re-identification and face recognition across cameras in video surveillance is critical for the construction of safety city. However, due to limited camera monitoring range of cameras, how to identify the same person across different cameras? Cameras have different imaging modes for day and night because of illumination variations, how to identify the same person across visible images and infrared images? After obtaining face image from different views and resolutions of the same person, how to reconstruct one high-resolution frontal face image? In the light of the aforementioned problems, based on the techniques of cross-domain image reconstruction, we implement this project in four phases: (1) we would design deep cooperation networks for universal cross-domain image reconstruction; (2) we would further study the problem of image reconstruction-based cross-domain person re-identification; (3) moreover, we would propose a cross-domain deep mapping networks for fast cross-domain person re-identification based on image reconstruction techniques. From this we can obtain the moving trajectories of the target person and as well as face images across different views and resolutions; (4) finally, we would employ frontal view face image super-resolution techniques to generate high-resolution frontal view face image for face recognition and further identify the target. This project has great significance for public security and national defense affairs.

视频监控中行人的跨摄像头重识别以及身份的识别是平安城市建设的重要内容。然而，由于每个摄像机的监控范围有限，如何在不同摄像机之间识别同一行人？由于日间与夜晚光照不同，摄像头成像模式不同，如何在不同时段得到的可见光与红外图像间识别同一行人？在得到同一行人不同角度、不同分辨率的人脸图像后，如何重构出正面清晰的人脸以进行人脸识别确认行人身份？针对以上问题，本课题拟分四步开展研究：（1）以跨域图像重构技术为基石，设计深度协同网络实现通用型的跨域图像重构算法；（2）进一步基于图像重构技术实现跨域图像行人重识别；（3）更进一步设计跨域深度映射网络实现基于图像重构的快速跨域图像行人重识别，以得到目标行人的移动轨迹，同时得到行人不同角度不同分辨率的人脸图像；（4）最终通过正面视角人脸图像超分辨重建得到目标行人的清晰正面人脸图像，以进行人脸识别确定目标身份。本项目的研究对公共安全和国防事业等具有重要意义。

为全面推进“天网覆盖”工程建设，助力监控系统的智能化改造升级，本项目以视频监控中的图像和视频数据为研究对象，重点研究跨域图像重构与识别的基础理论和关键技术。具体上，本项目研究了跨域行人重识别技术，实现了目标人物在跨模态、跨摄像头下的跟踪；其次，重点探索了跨域人脸图像重构和图像超分辨率重建技术，实现了低质量人脸图像的清晰化，为准确的身份识别提供了保障；最后，开展了可信机器学习技术的研究，提升了模型的可信任性和可靠性。相关成果已经实现技术转化。依托本项目，获2019年国家自然科学基金优秀青年基金资助，荣获2020年教育部自然科学一等奖，2020年中国图象图形学学会自然科学二等奖；发表与课题相关并标注本基金资助的学术论文共计113篇，其中国际知名期刊（IEEE T-PAMI, T-IP, T-NNLS, IJCV和Pattern Recognition等）论文57篇，中文期刊论文5篇，中国计算机学会推荐的A类国际会议（CVPR、ICML、NeurIPS、ICCV、AAAI、IJCAI和ACM MM等）和计算机视觉顶级会议ECCV论文28篇；合著新一代人工智能系列教材《人脸图像合成与识别》；获得国家发明专利授权16项。培养博士生6人，硕士生6人。