With the rising popularity and wide applications of image processing technologies in recent years, various deep learning methods have been proposed for image denoising. However, since most of such methods focus mainly on synthetic noises, their denoising effects on the spatially variant real-world noises could be further improved with more sophisticated network and training schemes. In this paper, a multi-scale generative adversarial network (MSGAN) that employs a novel network architecture and a well-designed training scheme is proposed. Specifically, a cascade multi-scale module is proposed as a basic building block of MSGAN to make use of the multi-scale context and increase the network learning capacity first, and then, a spatial attention mechanism is applied onto MSGAN to refine the denoising results. Finally, a sophisticated training scheme, which combines the pixel-level loss with the adversarial loss, is designed to suppress the real-world noises while restore both high-frequency and low-frequency image details simultaneously. Extensive experiments are conducted with several typical datasets to verify the effectiveness of MSGAN. Results demonstrate that MSGAN is promising for real-world image denoising in terms of both quantitative metrics (PSNR, SSIM) and visual quality.

近年来，随着图像处理技术的日益普及和广泛应​​用，各种深度学习方法被提出用于图像去噪。然而，由于大多数此类方法主要关注合成噪声，因此可以通过更复杂的网络和训练方案进一步改善它们对空间变化的真实世界噪声的去噪效果。在本文中，提出了一种采用新颖网络架构和精心设计的训练方案的多尺度生成对抗网络（MSGAN）。具体来说，首先提出级联多尺度模块作为 MSGAN 的基本构建块，以利用多尺度上下文并增加网络学习能力，然后将空间注意力机制应用于 MSGAN 以细化去噪结果. 最后，一个复杂的训练计划，它将像素级损失与对抗性损失相结合，旨在抑制现实世界的噪声，同时同时恢复高频和低频图像细节。对几个典型的数据集进行了大量实验，以验证 MSGAN 的有效性。结果表明，MSGAN 在定量指标（PSNR、SSIM）和视觉质量方面都有望用于现实世界的图像去噪。