Abstract:In recent years, the video deblurring method based on double-branch structures has caught extensive attention due to its restoration performance in scenarios with low computing resources, but its insufficient exploration and utilization of the complementarity of the front and back frames of the video result in the inability of the model to fully capture the immediate information of the video, thus limiting the processing ability of long-term blurring. To this end, an adaptive double-branch structure deblurring network is proposed. In the stages of feature extraction and feature reconstruction, five stacked U-Net networks are adopted for feature extraction and reconstruction. In the feature fusion stage, firstly, the current frame and front and back frames that have passed through the feature extraction stage are input into the multi-frame branch to obtain the multi-frame features. Then, the same current frame is input into a single-frame branch to obtain the single-frame feature. Finally, the multi-frame features and single-frame features are input into the feature fusion module, and the weights generated by the intermediate information extraction module are employed to guide the feature fusion module to generate the fusion features. The network can leverage not only the time information between consecutive frames, but also the spatial information of the current frame. Meanwhile, under the guidance of weights, it can adaptively generate fusion features, and the adaptive mode can improve the accuracy of the restored features. A large number of experiments show that the method can visually restore clearer images and more accurate details compared with the results of the mainstream video deblurring networks on the public datasets GOPRO, DVD and BSD, and the method is the best in terms of PSNR and SSIM adopted as the evaluation indicators.