CL-SRGAN: generative adversary network equipped with curriculum learning for image super-resolution

Abstract

Single image super-resolution is an approach to optimize the image stripe structure and improve the image quality. Recently, it developed rapidly based on convolution neural network, specially designed for this task, which becomes a hot topic of research and have shown remarkable result. Recently, many models have been developed based on Generative Adversarial Networks (GAN) and display enormous superiority compared with traditional deep learning methods. In GANs settings, adversarial loss pushes the generated image to natural image manifold with the help of a discriminator and at the same time trains discriminator to better discriminate the real image from those fake images generated by generator. In this course of confrontation, the generator is excellently trained and have achieved outstanding performance in the image super-resolution task. However, the traditional SRGAN image super-resolution reconstruction algorithm has slow training convergence speed. Moreover, excessive high-frequency texture sharpening leads to distortion of some details, which has a negative impact on the reconstructed image. In this work, curriculum learning algorithm is implemented to solve these problems and thus originally propose CL-SRGAN method, which is designed to help SRGAN achieve better performance on image resolution task. In the final experiment, CL-SRGAN has made an effective breakthrough in processing image reconstruction.