Abstract: Aiming at  the problems of blurred details, large computer resource occupation, and slow image processing for the existing image deblurring algorithms, we proposd a lightweight image blind deblurring network. Firstly, the main framework of the network used a multi-scale architecture to input images of different resolutions into the network, and gradually optimized the datails through cyclic processing.  Secondly, the asymmetric structure was designed to enhance the feature extraction ability of the encoder and the feature fusion ability of decoder. In the encoder, the mixed multi-scale convolutional layer and residual pyramid module were proposed to enhance feature extraction and  reduce the number of network parameters. In the decoder stage, deep semantics were introduced  by using jump linkage, and the multi-scale joint structure  loss function was proposed for optimization. Finally, we used two evaluation indicators to compare the performance of the method with the other classical methods on two widely used 
 GoPro and Kohler datasets. The experimental results show that the effect of the network  is better than that of the traditional methods and other classical deep learning mehtods. It not only improves the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), but also shortens the processing time.

Key words: image processing, blind deblurring, pyramid structure, multi-scale network