Research And Implementation Of Image Demotion Blurring Based On Deep Learning

Improved quality images have a significant influence on people’s daily lives,guiding them in various ways.However,blurry images such as out-of-focus or motion blur can hinder the acquisition of crucial information from images.In this article,we focus on the issue of motion blur in traffic monitoring images.For example,in road traffic,motion blur is easy to occur in the image because of the fast motion of the object photographed.Addressing the problem of blurry images can effectively aid the road traffic system.This article explores the current image datasets and motion blur removal algorithms available and presents a novel dataset focused on road traffic and a deep learning-based motion blur removal algorithm.The main content is as follows:Using Markov random process to generate random motion trajectories,using sub-pixel interpolation to discretize the trajectories and generate blur kernels,to construct a blurry dataset.This model uses the generative adversarial network as the infrastructure,which incorporates the feature pyramid network,so that the detailed textures and features of small objects in the image can be restored.Improved Dense Net is used as the most important feature extraction part of this study,and ordinary convolution is replaced by deformable convolution,so that the whole feature information of the object can be completely extracted and clear images can be better restored.In this paper,Patch GAN network is improved to improve the discriminant ability of discriminator by expanding its field of view.Due to the shortcomings of the algorithm based on deep learning,this study introduced the pruning network.Compared with the basic algorithm,the improved algorithm reduced the number of parameters by 32.6% without significantly reducing performance.This article used three different datasets to test and compare this model.Through comparative studies,it was found that in the dataset constructed in this study,the PSNR and SSIM obtained by this algorithm increased by 6.8% and 3.0% compared to the SRN algorithm.Finally,this article also designs a Py QT5-based image deblurring visualization system,which can read the image directly through the camera,but also through the local upload to obtain the image,and then through the system to deblur the image,the image has been significantly improved,the operation is simple,easy to use.