融合人脸跟踪和聚类的人脸图像去重方法

Face Image Deduplication Based on Fusion of Face Tracking and Clustering

人脸图像去重处理对智能监控系统中的人脸识别有着重要意义。针对视频中人脸检测环节会产生大量重复的人脸图像的问题,提出了一种融合人脸跟踪和聚类的人脸图像去重方法。在视频中,利用Multi-task Convolutional Neural Network中的人脸检测算法提取人脸框及其对应的坐标。根据人脸跟踪算法构建人脸轨迹和约束矩阵,同时引入人脸质量评估算法,从人脸轨迹中挑选人脸姿态和图像清晰度较好的人脸图像作为人脸轨迹的代表图片;再由约束矩阵和无监督聚类算法对人脸代表图像进行聚类以获取同一个人的人脸图像;最后对每一个人的人脸图像再次进行评估,得到去重结果。实验结果表明,通过人脸跟踪与无监督聚类融合的方法获取同一个人的人脸图像,再结合人脸质量评估算法,能够快速有效地从一段视频中获取每个人不重复的高质量人脸图片。

Face image deduplication is of great significance to face recognition in intelligent surveillance systems,since face detection in videos will produce a large number of repeated face images.In this paper,a method of face image deduplication in videosby integration of face tracking and clustering is proposed.In a video,use the face detection algorithm in the Multi-task Convolutional Neural Network to extract the face frame and its corresponding coordinates.Face tracking is used to construct the face trajectory and the constraint matrix,and the face quality evaluation algorithm is introduced to select the face pose and image clarity from the face trajectory.An optimal face image is used as a representative of the face trajectory.Combined with the constraint matrix and unsupervised clustering algorithm,the representative images of the faces are clustered to obtain the face image of the same person.Finally,the face image of each person is evaluated again to obtain the deduplication.Experimental results show that,through face tracking and unsupervised clustering,the face image deduplication method in videos can quickly and efficiently obtain high-quality face images that are not repeated for each person from a video.