Research On Motion Image Deblurring Algorithm Based On Generative Adversarial Networks

As an important information carrier,image contains abundant information and becomes the main way for people to record their life and communicate their emotions.As a kind of image restoration technique,the application value of motion image deblurring is increasing.With the advent of the era of artificial intelligence,deep learning plays an important role in the image field.As an excellent generation model,GAN is widely used in the fields of image generation,repair,classification and style transfer.Image deblurring has always been a research focus in the field of image processing.How to improve network performance,speed up training and improve image quality have become the focus of attention.In order to make the model can be used in the real scene and further improve the quality of the generated images,this paper improves and optimizes the model based on the generative adversarial networks,and proposes a new way to solve the problem of deblurring.The advantages of the algorithm model in this paper are as follows: 1.Using the real pictures as the training set,it can be trained in both paired and unpaired data sets,so it has stronger generalization ability in the real scene.2.Using the method of image translation,two "dual form" conditional generative adversarial networks are used to transform image deblurring into a problem of mutual conversion between blurry and clear fields.3.In the model design,global residual connection was adopted,WGAN was used as the discriminator,ResNext and Res Netv2 were used to replace the ResNet residual network module,and SFTGAN super-resolution was added to reconstruct the structure to improve the image edge and texture effect,so as to restore a clear and natural image.4.Yolo-v3 object recognition method was used in the evaluation of results,and the image quality was comprehensively measured by combining with conventional indicators.The method in this paper has carried out experiments on Gopro data sets and Lai data sets and achieved good fuzzy removal effect.Especially on Gopro data sets,SSIM and PSNR values have improved by 15.97% and 1.51% respectively compared with the benchmark model CycleGAN.At the same time,the results are similar to those algorithms of paired data sets.After the optimization of model structure and loss function through comparison experiments,SSIM,PSNR,Yolo recognition rate and other indicators are closer to or even better than other algorithms of paired data sets,which provides a powerful help for solving the problem of motion blur in real scenarios.