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

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.