Research On Motion Blurred Image Restoration Method Based On Multi-scale Recurrent Network

The increasing complexity of computer vision tasks is based on high-quality images.In process of image acquisition,the motion blur of images will be caused by the motion of the object,the motion of the camera or both simultaneously.In view of the key problem that current deep learning image deblurring methods cannot take into account both the effect and efficiency of deblurring,the following studies are carried out as follows:(1)Study and reproduce motion blurred image restoration methods in deep learning such as Deblur GAN,Deblur GAN-v2,SRN-Deblur Net.SRN-Deblur Net was selected as the basic network of this paper for its best performance.By analyzing the theory of the generation of motion blurred images,this paper embedded CBAM attention mechanism into the residual convolution block of SRN-Deblur Net to reasonably allocate operational resources to improve non-uniform blur removal effect.In this paper,the connection mode of CBAM submodule is improved,emphasizing the expression of spatial submodule,so that the SSIM and PSNR of the network on Go Pro test dataset reach 0.9396 and 30.7246 d B respectively.(2)This paper studies the performance difference between Res Block and Dense Block in SRN-Deblur Net,and designs a kind of dense connected convolution with limited channel dimension.On the premise of ensuring the network performance,the convolution depth of multi-scale network is greatly reduced.In order to further compress the network,this paper replace the standard convolution with the depthwise separable convolution,which improves the computing speed of the network to approximately 2.7 times of the original network.(3)A multi-scale recurrent attention network is proposed for motion blurred image restoration.In order to solve the problem that the edge of synthetic fuzzy images of Go Pro dataset have superposition trace,we use DAIN interpolation method to increase the frame rate of the images of the dataset to 32 times of the original,making the generated fuzzy image more realistic.In order to measure the deblurring effect of the deep learning algorithm on real blurred datasets,500 real blurred images were taken to expand Lai testing set,and evaluation without reference was conducted.In this paper,the edge loss function is introduced in network training to improve the generated edge details.Comparative experiments and ablation experiments were carried out for the proposed network model.Experiments prove that the proposed method has great generalization performance and robustness.Compared with the best performance of Deblur GAN,Deblur GAN-v2 and SRN-Deblur Net,the SSIM and PSNR have increased by about 1.64%,1.36% and 1.74%,1.30% on the Lai dataset and K(?)hler dataset,respectively.The average single frame running speed on the Go Pro dataset is nearly 2.5times faster than SRN-Deblur Net.(4)The function of the above-mentioned motion deblurred image restoration method based on multi-scale recurrent network in the restoration of image motion blur caused by multi-axis low-frequency vibration and unidirectional high-frequency vibration is studied.The average Sobel gradient value of the restored image is increased larger than 6% and 15%,respectively.Using the proposed network to restore the motion blur images collected by the UAV platform,the average Sobel gradient value of the video increased by about 9.97%.