Research On Blurred Image Restoration Based On Conditional Generative Adversarial Networks

Research on digital image restoration has always been one of the hot research directions in the field of image processing.Due to its superior visual intuitiveness and strong information inclusion,it has an irreplaceable position in social industry or daily life.Of course,degraded images often affect the integrity and correctness of information expression.The public security system is difficult to distinguish the blurred portraits,traffic camera shots blurred license plate,the unstable image transmission process causes image damage and degradation.All the above shows the urgency and necessity of the research on blurred image restoration.The traditional blurred image restoration method based on PSF can be divided into two categories: non-blind deblurring and blind deblurring.Non-blind deblurring is based on the known point spread function and has limitations in the requirements of real-world scenarios.Blind deblurring mainly uses a priori gradient regular model iterative optimization to estimate the point spread function.Due to the particularity of the features in the model,the restoration effect is also difficult to satisfy.With the development of neural networks,digital image processing has entered a new stage of development,especially the applications of the generative adversarial networks(GAN).However,due to the freedom and difficulty of convergence of GAN,it is still not effective for specific problems.In this paper,the conditional generative adversarial networks(CGAN)network model is used to study the blurred image restoration.CGAN can constrain the image generation direction,but poor convergence stability of the model loss will still result in a decrease in the quality of the restoration.Therefore,this paper adds the residual module(ResNet)to ensure that the model is stable and the common details are not lost;adding Perceptual Loss to improve the recovery of the distribution details;adding Gradient Penalty to enhance the stability of convergence,which benefits the reverse to update.According to the experiment,the test results are in line with expectations,which can meet the clear requirements of blurred images.The research content of this paper mainly includes the following points:1.By studying CGAN,this paper proposes an improved blurred image restoration model based on the conditional generative adversarial networks for digital image restoration.Compared with GAN,the model can constrain the generation of data distribution,and the specification generation distribution meets the requirements of specific scenarios.2.Data set preprocessing and model initialization.Use Gaussian Blur preprocess CelebA,motion blur algorithm preprocess GOPRO.The joint distribution of the CelebA and GOPRO training data sets and the corresponding blurred images is used as a model input to supervise the image generation direction.Initialize the design principles of the network referring deep convolutional neural network DCGAN.3.Enhance the depth performance of the generated model.For the network depth and inversely proportion,this paper is designed to join ResNet,which can solve the impact of depth on the degradation of the network,and also ensure the preservation of common information.4.Loss function robustness design.For the problem that generator loss can not restore the underlying information of the image,this paper joins the pre-training VGG-19 network to optimize the image detail reduction ability;For the difficulty to converge discriminator loss,this paper adds gradient penalty to solve the difficult convergence problem.Then loss has a good guiding effect on image generation.5.Blurred image restoration test and evaluation.The test set comes from the non-training set of CelebA and GOPRO,and the test results are evaluated from both the subjective and objective perspectives.Through the subjective and objective evaluation of the test results,the superiority of the research on the restoration effect is shown.