Multi-scale Deep Curve Estimation for Low-light Image Enhancement

Due to the limitation of the device, pictures taken in low-light environment usually consist of unpleasant deterioration, such as low contrast and color distortion. In this paper, we propose a Multi-scale Deep Curve Estimation network (MSDCE) for low-light image enhancement, which formulates the single low-light image enhancement task as a pixel-wise curve estimation by paired learning. To impose more priors of low-light regions, we propose an inverse illuminance map as part of the Curve Estimation network input. The curve estimation network backbone is composed of multi-scale modules which learns information from multi-scale feature streams and ensures the information exchange across different scales. Compared with several state-of-the-art methods, our method is significantly better. From the perspective of visual evaluation, our MSDCE can effectively improve the contrast and illumination of the image, and ensure the color fidelity of the image. CCS CONCEPTS • Computing methodologies • Artificial intelligence • Computer vision • Computer vision problems • Reconstruction