基于超分辨率重建的低质量视频人脸识别方法

Low-quality Video Face Recognition Method Based on Super-resolution Reconstruction

随着深度神经网络的兴起,人脸识别技术得到了飞速发展。但在光照条件差、低分辨率等情况下的低质量视频S2V(Still to Video)人脸识别由于存在低质量测试视频与样本库高清图像的异质匹配问题,仍然没有达到预期的效果。针对这个问题,提出一种基于超分辨率重建的低质量视频人脸识别方法。首先根据人脸姿态对低质量视频帧采用聚类算法和随机算法选取关键帧,然后建立一个面向低质量视频S2V人脸识别的超分辨率重建模型S2V-SR,对关键帧进行超分辨率重建,从而获得高分辨率且更多身份特征的超分辨率关键帧,最后使用视频人脸识别网络提取深度特征进行分类投票,得到最终的人脸识别结果。所提方法在COX视频人脸数据集上进行实验测试,在相对较高质量的cam1和cam3视频中获得了最好的识别准确率,即55.91%和70.85%,而在相对较低质量的cam2视频中获得了仅次于最好方法的识别准确率。实验结果证明,所提方法能够在一定程度上解决S2V人脸识别中异质匹配的问题,并且能够获得较高的识别准确性和稳定性。

With the rise of deep neural networks,face recognition technology has developed rapidly.However,S2V(Still to Video)face recognition for low-quality video that is poor lighting conditions and low resolution still does not achieve the expected results,because the heterogeneous matching problem between the test video of low-quality and the high-definition image of the sample library.To solve this problem,this paper proposes a face recognition method for low-quality video based on super-resolution reconstruction.First,it uses clustering algorithm and random algorithm to select key frames for low-quality video frames based on face pose.Then,it builds a super-resolution reconstruction model S2V-SR for low-quality video S2V face recognition,and performs super-resolution reconstruction on key frames to obtain super-resolution key frames with higher resolution and more identity features.Finally,it uses the video face recognition network to extract deep features for classification and voting to obtain the final result.The proposed method is experimentally tested on the COX video face data set,and the best recognition accuracy is 55.91%and 70.85%in the relatively high-quality cam1 and cam3 videos,while in the relatively low-quality cam2 video the recognition accuracy rate second only to the best method is obtained.Experiments show that the proposed method can solve the heterogeneous matching problem in S2V face recognition to a certain extent,and obtain higher recognition accuracy and stability.