Reconstructing the Noise Manifold for Image Denoising

Deep Convolutional Neural Networks (DCNNs) have been successfully used in many low-level vision problems such as image denoising, super-resolution etc. Although the conditional generative adversarial networks (cGAN)[44,16] have led to large improvements in this task, there has been little effort in providing a cGAN with an explicit way of understanding the image noise for object-independent denoising reliable for real-world applications. The task of leveraging structures in the target space (clean image) of the cGAN is very unstable due to the complexity of patterns in natural scenes, so the presence of unnatural artifacts or over-smoothed image areas cannot be avoided. To fill the gap, in this work we introduce the idea of a cGAN which explicitly leverages structure in the image noise space. By using the residual learning in our model, the generator promotes the removal from the noisy image only that information which spans the manifold of the image noise. Based on our experiments, our model significantly outperforms existing state-of-the-art architectures.