Research On Blind Deblurring Algorithm For Motion Blurred Image Based On Deep Learning

Image deblurring technology is not only one of the research hotspots in the field of computer vision,but also the most basic and significant research topic.The relative movement between the handheld device and the photographed object will cause the image to have a blur effect,and the main goal of image deblurring is to restore a clear image from a complex blurred image through an algorithm model.Since the generation of blur kernels does not satisfy uniqueness,blind deblurring of images is not only an ill-posed problem,but also an ill-conditioned problem.With the continuous application of deep learning technology in the field of image processing,image deblurring technology has improved in terms of evaluation indicators and visual effects.The deep neural network is used to extract the most representative features in the blurred image,a neural network model applicable to the structure of the blurred image is constructed in an end-to-end manner,and additional auxiliary information is added to train the network structure.Then make it optimal so that the blurred image can be recovered without the need to estimate the blur kernel.However,the process of generating blurred images is more complicated,and the information such as multi-scale features,dark channel,and image edges in the deblurring process cannot be used more effectively.Therefore,this thesis studies the multi-scale features of images,edge features and dark channel information,and proposes three effective image blind motion blur network models.The main research results are as follows:In the case of inaccurate blur kernel estimation,there will still be blurred areas in the restored image.To solve this problem,a multi-scale end-to-end encoder-decoder network model is proposed.The network model takes advantage of the multi-scale structure of the image and gradually refines the image features from the "coarse layer" to the "fine layer".So that only one blurred image can be input to obtain the corresponding recovered clear image without the need to estimate the blur kernel.Meanwhile,the low-dimensional information is fused with the high-dimensional information using improved residual networks and skip connections,which facilitate the generation of clear images.Experimental results show that compared with the traditional SFD method,the PSNR of the proposed model is increased by about 6.61 d B and 2.13 d B,respectively.Compared with CNN-based methods such as MRF CNN,Multi-scale CNN and Deblur GAN,the PSNR is increased by about 5.61 d B and 1.59 d B,the experimental results are better than the other four methods,the restored image is closer to the real image,and the texture details are clearer.In order to make full use of the edge features in the image and use it as auxiliary information to help the image deblurring operation in the deblurring process,an edge refinement network model is proposed.After the blurred image passes through the first stage of the model,the refined image strong edge feature will be obtained,which is obtained after the fusion operation is performed on the features extracted at every two layers in the codec network.The extracted features are input into the second stage of the model together with the blurred image as auxiliary information to help the model remove the blurred area.In the experimental part,an ablation comparison experiment with or without edge refinement subnets is set up.The model with edge refinement subnets improves the PSNR/SSIM by about 1.51 d B and 0.072,respectively.Compared with the model without edge refinement subnets,which further illustrates the important role of the edge features in image deblurring.The dark channel information describes the smallest intensity value in the image.Combining the dark channel features with the deep neural network can facilitate image deblurring tasks.For blurred images taken in dark environments,the difference from ordinary images is that the former contains rich dark channel features.In order to effectively use this feature,a multi-scale deblurring network model combining dark channel and edge features is proposed.Firstly,the model uses GAN to continuously reconstruct the fused images containing dark channel and edge information,and then the reconstructed fused images and blurred images are used as the input of the multi-scale deblurring model,which finally generates clear images with the combination of lowdimensional features and high-dimensional features of different scales.The model well solves the problem of difficult image recovery in low-light environment.Extensive experiments on Go Pro,K(?)hler and Li datasets show that the proposed model can not only recover texture details in images better,but also handle blurred images in dark environment well,and the deblurring effects are better than other models.