Abstract: Non-uniform blind deblurring in dynamic scenes is a challenging computer vision problem.Although deblurring algorithms based on deep learning have made great progress,there are still problems such as incomplete deblurring and loss of details.To solve these problems,a deblurring network based on residual attention and multi-feature fusion is proposed.Unlike the existing single-branch network structure,the proposed network consists of two independent feature extraction subnets.The backbone network uses an encoder-decoder network based on U-Net to obtain image features at different scales,and uses the residual attention module to filter the features,so as to adaptively learn the contour features and spatial structure features of the image.In addition,in order to compensate for the information loss caused by the down-sampling operation and up-sampling operation in the backbone network,a deep weighted residual dense subnet with a large receptive field is further used to extract rich detailed information of the feature map.Finally,the multi-feature fusion module is used to gradually fuse the original resolution blurred image and the feature information generated by the backbone network and the weighted residual dense subnet,so that the network can adaptively learn more effective features in an overall manner to restore the blurred image.In order to evaluate the deblurring performance of the network,tests are conducted on the benchmark data sets GoPro and HIDE,and the results show that the blurred image can be effectively restored.Compared with the existing methods,the proposed deblurring algorithm has achieved excellent deblurring performances in terms of visual effects and objective evaluation indicators.

Key words: Image deblurring, Attention mechanism, Encoding-Decoding structure, Dense residual network, Feature fusion