Research On Image Deblurring Method Based On Dual Discriminator Conditional Generative Adversarial Networks

The image may be blurred by many factors during the shooting process,and the sharpness of the image is an important premise for people to evaluate the image.Therefore,image deblurring is also an important research topic.The primary task of the traditional image deblurring method is the estimation of the image blur kernel.However,due to the many types of natural image blur,the estimation of the blur kernel is very difficult,which affects the deblurring effect.Therefore,this paper adopted the end-to-end idea and used the Dual Discriminator Conditional Generative Adversarial Networks to improve the sharpness of the blurred image.The main research work is summarized as follows:(1)The structure of the original Conditional Generative Adversarial Network consists of a generator and a discriminator.The generator captures the potential distribution of real data samples and generates new data samples for the purpose of deceiving the discriminator;the discriminator is a binary classifier that discriminates whether the input image from the sample distribution of real data or the sample distribution generated by analyzing each pixel in the image.(2)Based on the idea of original Conditional Generative Adversarial Network,this paper adopted the design of the dual discriminator.The network structure of the generator and the two discriminators are constructed separately,and the two discriminators are given different loss functions.Therefore,the two discriminators have different optimization goals.The first discriminator focuses more on the distribution of real data samples.If the sample from the real data distribution,gave a higher "reward" to this discriminator.The second discriminator focuses more on the distribution of generated data samples.If the sample from the generated data distribution,gave a higher "reward" to this discriminator.In the process of network learning,the two discriminators are trained against the generator at the same time.The data samples generated by the generator should deceive the two discriminators,which requires higher quality of the generated samples.This method not only improve the image clarity,but also improve the generalization performance of the network model.(3)After the network is constructed,the network would be trained.During the training process,the relevant parameters,network structure and optimizer type would be adjusted according to the specific conditions,and then retrained to achieve the purpose of improving the training effect.(4)After the model training converges,the deblurring effect is tested and compared with the deblurring effect of similar methods.In addition to testing the data samples in the test,which also added the actual captured image and the network image.The test results show that compared with the deblurring effect of the similar method,the effect of the method used in this paper is relatively better from the test data of PSNR and SSIM in the objective evaluation method from or the subjective level,which confirms the feasibility of the method used in this paper.