面向安防监控场景的低分辨率人脸识别算法研究  

Research on low-resolution face recognition algorithm for security surveillance scene

作　　者：卢峰[1] 周琳 蔡小辉 Lu Feng;Zhou Lin;Cai Xiaohui(78090 Troops of PLA,Chengdu 610054,China;State Key Laboratory of Integrated Services Networks,Xidian University,Xi’an 710071,China)

机构地区：[1]中国人民解放军78090部队,成都610054 [2]西安电子科技大学综合业务网理论及关键技术国家重点实验室,西安710071

出　　处：《计算机应用研究》2021年第4期1230-1234,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目;教育部—中国移动科研基金资助项目。

摘　　要：针对安防监控场景中获取的人脸图像质量不佳、细节信息丢失导致的人脸识别准确率低下的问题,提出一种基于超分辨率重建的低分辨率人脸识别算法。该算法包括超分辨率重建和人脸识别两个子网络,分别实现低分辨率人脸图像的超分辨率重建和人脸特征的提取。首先通过增加超分辨率重建子网络激活函数前的特征图数量实现广泛激活,保证信息流的有效传递,重建出包含更多细节信息的高分辨率人脸图像;然后在训练时结合图像内容损失和身份损失,在重建图像的同时保留更多身份信息,使得提取到的人脸特征具有更强的辨别性。实验结果表明,该算法提升了低分辨率人脸识别的准确率,在监控人脸数据集QMUL-SurFace上的性能优于传统算法。Aiming at the problem of low face recognition accuracy caused by poor image quality and loss of detailed information of face pictures obtained in security surveillance scene,this paper proposed a low-resolution face recognition algorithm based on super-resolution reconstruction.The algorithm included two sub-networks:super-resolution reconstruction and face recognition,which could respectively realize super-resolution reconstruction of low-resolution face image and extraction of face features.Firstly,the algorithm increased the number of feature maps before the activation function of super-resolution reconstruction sub-network to achieve wide activation and ensure effective transfer of information flow,so as to reconstruct high-resolution images containing more effective detailed information.Then,it combined image content loss and identity loss during training to retain more identity information while reconstructing image,which could make extracted face features more discriminative.Experimental results show that the proposed algorithm improves accuracy of low-resolution face recognition and has better performance than traditional algorithms on surveillance face dataset QMUL-SurFace.

关 键 词：安防监控 超分辨率重建 广泛激活 身份损失 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]