Optical flow correction network for repairing video

Some areas in the video may be defective due to storage medium damage or data transmission loss. The area can repaired utilizing optical flow since represents motion information of pixel and contains redundancy properties video. However, it is a challenge for current algorithms directly calculate correct area. Therefore, an correction network (OFC-Net) proposed solve this problem. Firstly, as basic framework, OFC-Net employs lightweight U-shaped structure gradually “repair” broken map based on known neighborhood. And then, activation function that conforms property used accelerate model’s convergence during training, while simplified loss adopted evaluate similarity from global perspective improve prediction accuracy. As result, corrected output at masked position obtained. Secondly, by guide locating pixels nearby frames repair region. Optical experiments were conducted MPI Sintel, DAVIS datasets, real movie clips. experimental results show accurate without bringing artifacts under wide range scenarios arbitrary irregular mask shapes, which demonstrate network’s strong generalization capabilities. average PSNR SSIM are greater than 33.67 0.986, respectively, outperforming traditional method PatchMatch deep learning-based synthetic methods.