Video deblurring method and system

Abstract

The invention provides a video deblurring method and system, and the method comprises the steps: firstly extracting image features in a fuzzy video frame sequence through employing a feature extraction network, and obtaining a first image feature; then, static context information and dynamic context information in the first image feature are extracted by adopting a spatio-temporal context self-attention network, and the static context information and the dynamic context information are fused to obtain a second image feature; carrying out optical flow and correction on the second image feature by adopting the depth feature propagation network according to an optical flow value of the fuzzy video frame sequence to obtain a third image feature; and finally, performing image reconstruction on the third image feature by adopting a video reconstruction network to obtain a clear video frame sequence. According to the video deblurring method provided by the invention, the video deblurring model can better pay attention to global information in adjacent frames through the spatio-temporal context self-attention network, and longer-term information in all video frames can be paid attention to by utilizing the depth feature propagation network, so that the video deblurring effect is improved.

Description

technical field [0001] The present invention relates to the technical field of video processing, in particular to a video deblurring method and system. Background technique [0002] With the wide application of many high-definition shooting devices in daily life, people urgently need to apply deblurring technology on these devices to get clear video. The video deblurring method requires the input of multiple frames of video to obtain the restored multi-frame video. The video may contain unknown motion blur and luminance noise and other factors, so the hardware of the device and the robustness of the algorithm have higher requirements. . Since video deblurring is a highly ill-conditioned problem, traditional methods are mainly based on a Bayesian framework, using various priors to solve this type of problem. However, such degenerate methods based on motion blur assumptions and latent frames usually do not match, in addition, the priors constructed by such methods usually le...

AI Technical Summary

Problems solved by technology

However, this kind of motion blur-based assumption and the degradation method of the potential frame usually do not match. In addition, the priors constructed by such methods usually bring about very complicated optimization problems, which are often difficult to solve.

At present, with the development of convolutional neural networks, people usually directly stack adjacent frames of video to deal with the problem of video deblurring, but this does not fully pay attention to the information of time latitude

Then 3D convolution, RNN and other methods have been proposed to solve the problem of insufficient utilization of time dimension information, but these methods cannot effectively use global information

Images