Illumination Compensation And Image Denoising for Low-Light Images Based on Deep Learning

Abstract

Image denoising is one of the basic low-level computer vision problems, but low-light denoising is challenging due to low photon count and low SNR. Therefore, we propose an end-to-end encoded and decoded network of illumination compensation and image denoising in low-light condition based on deep learning, which is used to denoise low-light images and adaptively brighten images without over-amplifying the brighter part of the images with high dynamic range.In the network, the illumination compensation branch network eliminates the disadvantage that the magnification must be selected externally. Different simulation gain and exposure time are used to train the multi-light compensation coefficient, which can eliminate the residual errors caused by inaccurate gain and various exposure time effectively. The results show that the model is suitable for the recovery and reconstruction of natural low-light images with different degrees of degradation due to the advantages of flexibility and data driving.