LRD-Net: A Lightweight Deep Network for Low-light Image Enhancement

Abstract

The computer vision algorithms demonstrate superior performance in the real world when provided with high-quality image or video inputs. However, low-light or non-uniformly illuminated conditions result in low visibility in the captured images, thus demanding low-light image enhancement. Therefore, the paper presents a lightweight deep learning-based method, referred to as LRD-Net, for enhancing low-light images. The proposed approach initially estimates the illumination map using maximum of RGB color channels and then utilizes the retinex theory to derive the corresponding reflectance. Further, we propose a lightweight deep network for refining the estimated reflectance, which is capable of restoring the lost structural and textural details. Our proposed approach employs a lightweight structure yet demonstrates competitive performance to various state-of-the-art methods. The fast computation speed escalates the practical usage of our method without compromising on quality, as demonstrated by extensive experiments.