A Deep Learning Method for Joint Processing of Video Super-Resolution and Motion Deblurring
Aiming at the problems of motion blur and low res-olution caused by various reasons in video frames,this paper proposes a convolutional neural network for joint processing video super-resolution and motion deblurring which is sup-posed to be able to cope with severe motion blur and restore sharp high-resolution video frames. The proposed network first introduces a pyramid optical flow network to extract sharp latent images in blurred frames and generate high resolu-tion optical flow maps for motion estimation. The motion compensation is then completed through frame-to-frame recur-sion,making full use of temporal's a priori information. Next,the super resolution features and deblurring features are extracted separately in a parallel manner. Finally,the channel attention mechanism is used to enhance the fusion efficiency of the two branch features. The experimental results on public benchmarks show that the results of our method are superior to the existing super resolution methods and deblurring methods.
万方数据
维普
