DeGAN: Mixed noise removal via generative adversarial networks

Abstract

Restoration of images corrupted by mixed noise (e.g., additive white Gaussian noise and impulse noise) is very difficult due to the complexity of the mixed noise distribution. Various mixed noise removal models involve the preprocessing based on outlier detection. However, the performance of these models largely depends on the accuracy of pixel location detection of outliers, and artifacts and missing image details are prone to occur when the mixture noise is strong. In this paper, a new denoising model based on generative adversarial network (DeGAN) is proposed to remove mixed noise in images. The proposed model combines generator, discriminator, and feature extractor networks. Through the mutual game between the generator and discriminator networks combined with additional training from the feature extractor network, the generator network implements a direct mapping from the noisy image domain to the noise-free image domain. In addition, we design a new joint loss function to incorporate information from image features and human visual perception into the mixed noise elimination task, which further improves the image quality and the visual effect. Abundant experiments show that the performance of our model is better than the state-of-the-art mixed noise removal methods in three different types of mixed noise scenarios, and the joint loss function does improve the denoising performance.