结合对齐度准则的视频人脸快速配准算法

Fast registration algorithm for face in video based on alignment metric

在视频人脸替换过程中,针对人脸检测和特征点定位不考虑前后帧的时域联系,导致处理后的视频人脸出现抖动、错位现象的问题,提出结合对齐度准则的视频序列人脸配准方法。通过人脸特征点定位确定待匹配人脸的候选区域和参考人脸五官图像,并根据梯度信息得到参考人脸搜索子图像。构造图像尺度金字塔,在尺度金字塔中快速搜索对齐度极值,极值坐标对应搜索子图像相似度最高的配准点,构造匹配点对。采用Procrustes analysis算法计算对应匹配点的变换关系矩阵,完成待配准图像的坐标变换。根据实验数据,提出的算法提高了优化搜索的速度,处理后的视频与其他算法相比,前后帧对齐的人脸位置和姿态稳定,经处理的视频播放过程中人脸位置自然流畅。算法消除了肤色区域的影响,采用前后帧人脸的有效信息,保持了灰度变化大的人脸区域的位置和姿态,实现视频人脸的精确配准。

Aiming at the problem that in the process of video face replacement,face detection and feature points localization do not consider the contact of consecutive frames,in time domain which causes jitter and misplacement of the processed face in video,and the method of face registration in video based on alignment metric is put forward.The candidate region of the unregistrated face image and the facial features of the reference face image are determined by the feature point positioning,and the reference face search kernel is obtained according to facial gradient information.The image scale Pyramid is constructed,and search the extreme value of the alignment metric in the scaled image Pyramid at a high speed;the coordinates of the extreme value is the corresponding matching point that get a highest similarity compared with the search kernel and the matching pair of points are generated.Procrustes analysis algorithm is applied to calculate the transformation matrix of the corresponding matching points and the affine transformation of the image to be registered is completed.According to the experimental data,the speed of the optimization search is improved,and compared with other algorithms,the aligned face position and pose of the frame is much more stable in the processed face video by the proposed algorithm.Meanwhile the face location in the processed videos is naturally smooth when playing.The algorithm eliminates the influence of the skin color region,uses the valid face information of consecutive frames,maintains the position and poses of the face area with large gray scale change,and realizes the precise registration of the video face.