基于光流的视频缺陷检测及修复方法

为了修复视频中的划痕和斑点，提出一种基于光流的视频缺陷检测及修复方法。首先，根据光流场得到相邻帧对应像素之间的位置关系，利用对应点灰度差确定像素点所在位置是否为缺陷。其次，修正缺陷区域的光流，以修正光流指向的相邻帧修补点填补对应的缺陷点。最后，针对已修复的视频帧重新计算光流场并重复修复步骤，直到该帧满足迭代修复的收敛条件。针对DAVIS视频数据集的不同场景，模拟产生数量为单帧像素点总数1%左右的缺陷后进行检测修复实验，给出查全率与误识别率的关系曲线，其中，误识别率为0.1%时，查全率可达80%以上；修复后的SSIM大于0.991，LPIPS小于0.037。针对老旧视频的修复实验表明，算法能够有效去除细小划痕和大小斑块。

Video defect detection and repair method based on optical flow

In order to repair scratches and spots in video, this paper proposed a video defect detection and repair method based on optical flow field. Firstly, it obtained the positional relationship between the corresponding pixels of adjacent frames according to the optical flow field, and used the intensity difference of the corresponding points to determine whether the pixel position was a defect point. Then, it corrected the optical flow of the defect area and filled the corresponding defect points with the repair pixels in the adjacent frames pointed by the corrected optical flow. Finally, for the repaired video frame, it recalculated the optical flow field and repeated the above repair steps until the frame satisfy the convergence condition of iterative repair. For different scenes in the DAVIS dataset, after simulating the defects that account for about 1% of the total number of pixels in a single frame, this paper conducted a detection and repair experiment, and gave the curve between the recall rate and the false recognition rate. When the false recognition rate was 0.1%, the recall rate could reach more than 80%; the SSIM was greater than 0.991 and the LPIPS was less than 0.037. Experiments on old videos show that scratches and spots can be effectively removed.