真实非受控场景下的人脸图像超分辨率重建技术研究

近年来，公共场所的智能监控系统建设在社会管理、服务民生、社会治安等方面发挥了积极作用。在真实场景下，受到摄像采集设备与拍摄目标间距离等因素的影响，通常只能获取低分辨率的人脸图像。因此，人脸超分辨率重建是应用于智能监控的重要技术之一。现有大部分人脸超分辨率算法通常针对已预设降质模型的正面低分辨率人脸图像开展重建研究，难以适应实际场景面临的复杂情况。真实场景下监控摄像头拍摄到的人脸图像具有分辨率极低、采集环境复杂多变，人脸姿态难以控制三个主要特点，本项目将针对上述三个问题重点开展以下研究：1）极低分辨率图像的大倍数超分辨率重建技术；2）面向异质成像环境的人脸超分辨率重建技术；3）基于多角度低分辨率图像的正面高分辨率人脸生成技术，旨在提高真实复杂场景人脸超分辨率重建技术的效率、准确性和鲁棒性，助力智能监控系统和社会安防平台的建设。

In recent years, the construction of intelligent monitoring systems in public places has played an active role in social management, serving people's livelihood, and social security. In real scenes, it usually captures low-resolution face images due to the factors such as the far distance between the camera acquisition device and the human target. Thus, face hallucination is one of the important technologies used in intelligent monitoring. Most of the existing face hallucination algorithms usually conduct research on the frontal low-resolution face images with predefined degradation model, which is difficult to adapt the complex situations in real scenes. The face images captured by the surveillance camera have three main characteristics: extremely low resolution, complex and variable collection environment, and unfixed facial pose. This project will focus on the following researches to solve the above three issues: 1) reconstruct extremely low-resolution face images by large magnification factors; 2) reconstruct face images under heterogeneous imaging environments; 3) generate frontal high-resolution face images based on multiple low-resolution images with unfixed pose. This project aims to improve the efficiency, accuracy and robustness of face hallucination technology under real and complex scenes, which will further support the construction of intelligent monitoring systems and social security platforms.