Research On Image Repair And Reconstruction Based On Generative Adversarial Networks

In the image field,the requirements for image clarity and resolution are getting higher and higher.Studying the problems of blurred image repair and image super-resolution reconstruction not only helps to improve the visual quality of the image,but also improves the robustness of visual applications.It has important practical application value.Based on the research of Generative Adversarial Networks,this paper attempts to improve the Generative Adversarial Networks and apply them to image processing to solve the problems of image blur and low resolution.The following research was done specifically:(1)Analyzing and learning the basic models and algorithms of Generative Adversarial Networks.Using Generative Adversarial Networks to generate images has become a hot research issue,but the problems of gradient disappearance and model collapse still appear when it is working.To solve this problem,a Wasserstein Generative Adversarial Network(WGAN)was proposed and improved to enable it to solve the problems of image blur and low resolution.(2)A fuzzy insulator image restoration algorithm is proposed based on WGAN.First,the algorithm uses a fuzzy algorithm to generate fuzzy insulator dataset and train the model;then a residual network is introduced into the generator to improve the clarity of the insulator picture generated,and Wasserstein distance is introduced into the loss function to avoid the model appearing the problem of collapse during training;Finally,experiments are designed and compared with other deblurring algorithms to effectively verify that the proposed method can recover fuzzy insulator pictures and improve the insulator target detection rate of the pictures of insulator.(3)A super-resolution reconstruction algorithm image is proposed based on WGAN.First,the algorithm introduces a residual network module in the generator to improve the utilization of feature information of low-resolution pictures;secondly,a discriminator based on a multi-layer convolutional neural network is designed to improve the discriminative stability;then,by constructing a loss function combining perceptual loss,Wasserstein distance,and mean square error loss,the quality of the generated image is improved;finally,a comparative experiment of super-resolution reconstruction of pictures in the data sets Set5,Set14 and BSD100 is designed to verify the effectiveness of the proposed method.