SUnet plus plus :Joint Demosaicing and Denoising of Extreme Low-Light Raw Image

Despite the rapid development of photography equipment, shooting high-definition RAW images in extreme low-light environments has always been a difficult problem to solve. Existing methods use neural networks to automatically learn the mapping from extreme low-light noise RAW images to long-exposure RGB images for jointly denoising and demosaicing of extreme low-light images, but the performance on other datasets is unpleasant. In order to address this problem, we present a separable Unet++ (SUnet++) network structure to improve the generalization ability of the joint denoising and demosaicing method for extreme low-light images. We introduce Unet++ to adapt the model to other datasets, and then replace the conventional convolutions of Unet++ with M sets of depthwise separable convolutions, which greatly reduced the number of parameters without losing performance. Experimental results on SID and ELD dataset demonstrate our proposed SUnet++ outperform the state-of-the-arts methods in term of subjective and objective results, which further validates the robust generalization of our proposed method.