Image denoising is one of the most classic problems in computer vision for restoring corrupted images. It has been approached by using various traditional state of the art architectures in convolutional neural network (CNN), which has demonstrated considerably better results than the prior methods. There has been recent advancements in approaching the problem using generative adversarial networks (GAN), which has shown considerable promise. In this paper, we propose a novel denoising adversarial architecture to generate denoised image samples from a noisy distribution. A denoising autoencoder has been employed as the Generator to learn image distributions and generate denoised images while the discriminator penalizes the generated output. We employ an additive loss comprising of root mean square and mean absolute error for the Generator function. The model is trained adversarially followed by extensive experiments. We achieved PSNR and SSIM values comparable to the state-of-the-art for a range of blind and non-blind Gaussian noise.

中文翻译：

---

DAE-GAN：一种基于自动编码器的高斯去噪对抗网络

图像去噪是计算机视觉中用于恢复损坏图像的最经典问题之一。已经通过在卷积神经网络 (CNN) 中使用各种传统的最先进架构来实现它，这已经证明了比之前的方法更好的结果。最近在使用生成对抗网络 (GAN) 解决问题方面取得了进展，这已显示出相当大的前景。在本文中，我们提出了一种新颖的去噪对抗架构，以从噪声分布中生成去噪图像样本。已使用去噪自动编码器作为生成器来学习图像分布并生成去噪图像，而鉴别器则惩罚生成的输出。我们对生成器函数采用了一个附加损失，包括均方根和平均绝对误差。该模型经过对抗性训练，然后进行大量实验。对于一系列盲和非盲高斯噪声，我们实现了与最新技术相当的 PSNR 和 SSIM 值。