基于低秩先验和对抗学习的低质量视频人脸表征学习研究

Learning robust representation for human face is the core of face recognition problem. However, it is challenging to achieve this goal for the low-quality video face recognition under complex surveillance environment. One has to overcome the corruptions caused by uncontrolled factors in surveillance environment, and meanwhile, the cross-modality difference between real-time captured low-quality faces and registered high-resolution faces should be alleviated. In this proposal, based on the low rank property of video faces, we are trying to develop the robust face representation model for low-quality video face recognition by means of low rank theoretical framework and the spirit of adversarial learning. We focus on the relationship analysis model between low-quality video face sequences and high-resolution faces, and the shared face representation across modalities. First, we build the relationship model between low-quality video face sequences and high-resolution faces by establishing low rank representation learning architecture for video faces. Further more, we try to build a video face analysis model using adversarial network, attempting to mine the intrinsic structure of video face distribution and learn the shared robust representation of multi-modality video faces. Our research in this proposal will contribute the development of low rank representation, adversarial learning and video face representation learning.

人脸识别技术的核心是实现人脸的鲁棒表征。在复杂的视频监控环境下，低质量人脸的鲁棒表征学习极具挑战性，一方面需要克服光照、姿势等非可控因素的干扰，另一方面还需要消除低质量视频人脸与高清注册人脸间的跨模态差异。本研究基于视频图像的低秩先验，在低秩表达的理论框架下结合对抗学习的思想，重点研究低质量视频人脸识别的鲁棒表征问题。本研究拟从低质量人脸与高清人脸的关联关系分析和跨模态共享表征两方面入手，首先通过建立视频人脸的低秩表达学习体系，研究低质量视频人脸序列与高清人脸间的关联关系模型；在此基础上，通过构建基于生成式对抗网络的视频人脸分析模型，充分挖掘无类标视频人脸数据分布的内蕴结构，学习跨模态共享的低质量视频人脸鲁棒表征模式。本研究将进一步丰富低秩表达和对抗学习的理论和方法，同时也促进视频人脸表征学习技术的发展。

人脸鲁棒表征学习是视频低质量人脸识别的关键。本项目重点研究基于数据驱动的低质量视频人脸预处理方法，建立视频场景下的非可控人脸生成模型以及面向低清监控人脸与高清人脸的关联关系分析模型，解决非可控视频人脸识别数据量不足、图像质量不佳等难题；在此基础上，针对现实应用中广泛存在的跨模态人脸匹配问题，尤其是涉及多光照变化和跨年龄匹配的应用场景，研究基于生成式对抗网络的跨模态人脸内蕴表征方法，发展并建立适用于视频人脸的鲁棒表征模型，极大克服了光照变化、证件照及实拍照年龄差异过大引起的多模态类内差异过大的问题；继而进一步针对视频弱小目标检测、动态视频人脸活体检测难等问题提出系列优化方法，为构建面向低质量视频人脸的识别系统提供了坚实保障。