基于人脸区域特征相关性的视频流人脸识别研究

Research on face recognition of video stream based on face area feature correlation

目前,基于深度学习提取人脸特征进行人脸静态图片识别的方法,在Labeled Faces in the Wild(LFW)数据集等标准集上的正确识别率几乎接近人类。但是,在视频流中,由于人体的不停运动和姿态偏移等问题,导致检测到的部分人脸区域严重模糊和不完整,如监控系统中的人脸。这种情况下,单纯地采用基于图片的人脸识别方法,准确率会严重下降。在基于视频流的人脸区域提取时,本文提出采用单张人脸区域图像的特征自相关指标来衡量人脸的姿态以及模糊状况,针对连续多帧中人脸区域图像存在的信息冗余,提出利用连续多帧中人脸区域图像的特征互相关指标来衡量视频流中人脸区域的变化程度。基于提出的自相关指标与互相关指标,本文提出并实现了视频流中适用于识别的人脸区域图像的选取算法,以及加权投票的人脸识别算法。研究中收集并制作了基于视频流的人脸数据集,验证了本文提出算法的可行性。实验表明,本系统在有较高的识别率的同时,大幅度降低了人脸识别计算量,使得人脸识别可在视频流中实时稳定地进行。

At present, based on the depth of learning to extract facial features for face static image recognition method, the correct rate in the Labeled Faces in the Wild （LFW） data set and other standard sets is almost close to the artificial recognition rate. However, in the video stream, such as face in the monitoring system, due to the movement of the human body and attitude deviation and other issues, resulting in the detected part of the face area is seriously blurred and incomplete. In this case, simply using image-based face recognition method, the accuracy rate will be seriously reduced. In the face-based region extraction based on video stream, this paper proposes the characteristic autocorrelation index of the single face region image to measure the posture and the fuzzy condition of the face. For the information redundancy of the continuous multi-frame face area image, the cross-correlation index of the continuous face area image features is proposed to measure the degree of change in the face area of the video stream. Based on the proposed autocorrelation index and cross- correlation index, this paper presents and implements the selection algorithm of face region in video stream, and the face recognition algorithm of weighted voting. In this paper, the face data set based on video stream is collected and produced, and the feasibility of the proposed algorithm is verified. Experiments show that the system significantly reduces the amount of face recognition calculation, while having a high recognition rate, thus achieving the effect that face recognition can be performed in stable and real time in the video stream.