Evolutionary Approach to GAN Hyperparameter Tuning: Minimizing Discriminator and Generator Loss Functions

Image Reconstruction has always been among machine vision's challenging topics. One of image restoration's most challenging is to fill the damaged area after deleting, it in a visually acceptable way. The beginning of image reconstruction goes back to the last five decades, but due to the ineffectiveness of the basic methods, new methods have been offered. In the field of image restoration, GAN or Generative Adversarial Networks can be very useful due to the high similarity between the generated data and the training data. In this paper, by presenting an algorithm based on these networks, we try to increase the accuracy of the image restoration process. The GAN's accuracy is related to the correct selection of parameters. Using trial and error methods to find parameters is time-consuming. In this paper, the optimal parameters of the GANs have been used by providing suitable coding for the genetic algorithm. The simulation results represent that the proposed GA has notable performance. The best results give a minimum value of about 0.16 for the discriminator loss function and 0 for the generator loss function.