面向低质量人脸识别的特征表达增强

As one of the widely used biometric technologies, face recognition has gradually played an important role in the fields of finance, education, security, medical care, etc., and has greatly improved the quality and efficiency of related services. In recent years, face recognition has made great progress, and a variety of representative face recognition models have been proposed. However, the latest research methods show that face image quality is essential for extracting high discriminative facial features. Besides, most of the existing methods perform well on high-quality face images but often perform poorly on low-quality faces (occlusion, the side face, blur, small face, etc.). Typically, it is often impossible to ensure the acquisition of high-quality face images in the unconstrained environment. In order to meet the high recognition rate and strong robustness required by various related applications and starting from the urgent needs of current low-quality face recognition reality applications, this proposal aims to enhance the feature representation of low-quality faces by enhancing the face details in the pixel space, spatial relationship of face parts, and temporal relationship of face sequences, thereby improving the accuracy and robustness of face recognition and face verification. The research results of this proposal will greatly promote the wide application of face recognition in various industries and support the industrial development of related technology applications.

人脸识别作为当前广泛应用的生物识别技术之一，逐渐在金融、教育、安防、医疗等领域发挥着重要作用，大幅度地提高了相关服务的质量和效率。近年来人脸识别研究取得了巨大进展，提出了多种经典的人脸识别模型。然而，最新的研究成果表明，人脸图像质量对于提取高区分性人脸特征至关重要，现有的大多数方法在高质量人脸上性能良好，而在低质量人脸（遮挡、侧脸、模糊、远距离等）上往往表现不尽人意，低质量人脸识别是一个极其重要又极具挑战性的问题。针对无约束环境下往往无法确保获取到高质量人脸图像的现状，同时为了满足各项相关应用所需的高识别率和强鲁棒性要求，从当前低质量人脸识别现实应用的迫切需求出发，本项目旨在从人脸的像素空间域、人脸部位的空间关系、人脸序列的时间关系上进行增强处理，以实现低质量人脸的特征表达增强，从而提高人脸识别与人脸验证的精度与鲁棒性，推动人脸识别在各行各业中广泛应用，极大地支撑相关技术应用的产业化发展。

本课题组自承担项目以来，经过三年的努力，课题工作进展顺利，各项工作已完成，取得了较为丰硕的科研成果。具体来讲，针对人脸感知、低质增强等，提出了面向文本指导图像修复的多模态融合学习算法，低质修复质量得到显著提升；提出了多源信息融合算法以实现多参考的自动人眼修复，实现人脸人眼修复质量的显著提升；基于立体图视差关联，提出了视差注意的域自适应算法以提升立体图像（人脸）重构质量；提出了多模态非对称对偶学习算法以实现无监督的人脸图像中眼镜去除，主客观性能达到了先进效果。提出轻量级全局分析模块并将其引入图像去模糊任务，计算成本获得了指数级降低且去模糊性能得到提升。.相关科研成果陆续得到了国际著名期刊和学术会议的认可。共有14篇相关学术论文发表在国际著名的IEEE Transactions等国际期刊和学术会议上。在所发表文章中，4篇已被SCI收录（包括IEEE TBC、IEEE TMM），10篇已被CCF A类会议收录。