当前位置: X-MOL 学术Neural Process Lett. › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Gaussian Pyramid of Conditional Generative Adversarial Network for Real-World Noisy Image Denoising
Neural Processing Letters ( IF 2.6 ) Pub Date : 2020-03-06 , DOI: 10.1007/s11063-020-10215-w
Ruijun Ma , Bob Zhang , Haifeng Hu

Image denoising is an essential and important pre-processing step in digital imaging systems. However, most of existing methods are not adaptive in real-world applications due to the complexity of real noise. To address this problem, a novel pyramidal generative structural network (PGSN) is proposed for robust and efficient real-world noisy image denoising. Specifically, we consider the denoising problem as a process of image generation. The procedure is to first build a Gaussian pyramid where a cascade of encoder-decoder networks are used to adaptively capture multi-scale image features and progressively reconstruct the corresponding noise-free image from coarse to fine granularity. Then, we train a conditional form of GAN at each pyramid level. By integrating the conditional GAN approach into the Gaussian pyramid, the proposed network can well combine the image features from different pyramid levels, and an incremental distinction between the real noise and image details is dynamically built up, hence greatly boosting the denoising performance. Extensive experimental results demonstrate that our PGSN gives satisfactory denoising results, and achieves superior performance against the state-of-the-arts.

中文翻译:

条件生成对抗网络的高斯金字塔用于真实世界的噪声图像去噪

图像去噪是数字成像系统中必不可少的重要预处理步骤。但是,由于真实噪声的复杂性,大多数现有方法在现实世界的应用中均不适用。为了解决这个问题,提出了一种新颖的金字塔生成结构网络(PGSN),用于健壮和有效的现实世界中的噪声图像降噪。具体来说,我们将去噪问题视为图像生成过程。该过程首先要建立一个高斯金字塔,其中使用级联的编码器/解码器网络来自适应地捕获多尺度图像特征,并从粗粒度到细粒度逐步重建相应的无噪声图像。然后,我们在每个金字塔级别训练GAN的条件形式。通过将条件GAN方法整合到高斯金字塔中,所提出的网络可以很好地组合来自不同金字塔等级的图像特征,并且动态建立真实噪声和图像细节之间的增量区别,从而大大提高了降噪性能。大量的实验结果表明,我们的PGSN具有令人满意的降噪效果,并且相对于最新技术具有出色的性能。
更新日期:2020-03-06
down
wechat
bug