Image Denoising via Generative Adversarial Networks with Detail Loss

Abstract

Image denoising is a challenging task which aims to remove additional noise and preserve all useful information. Many existing image denoising algorithms focus on improving the typical object measure, peak signal-to-noise ratio (PSNR), and take the mean square error (MSE) as their loss function to train their networks. Although these algorithms can effectively improve the PSNR on the benchmark dataset, their denoised images often lose some important details or become over-smooth in some texture-rich regions. In order to solve this problem, we introduce Generative Adversarial Networks (GAN) and perceptual loss from single image super-resolution (SISR) field into our image denoising work. The GAN and perceptual loss can help our network to better focus on the recovering of details during denoising. To understand easily, we use the term Detail Loss to represent the whole loss which includes the MSE and the perceptual loss. Besides, we propose a new convolutional neural network which achieves state-of-the-art result on PSNR. Our experimental results show that our method outperforms the current state-of-the-art methods on preserving the details during denoising. Compared with the current state-of-the-art methods, the denoised images by our method are clearer, sharper and more realistic on details.