ColorBoost-LLIE: A multi-loss guided low-light image enhancement algorithm with decoupled color and luminance restoration

Abstract

Low-light image enhancement (LLIE) aims to improve image brightness and achieve natural-looking images under normal lighting conditions. Existing low-light image enhancement algorithms primarily focus on increasing image brightness, significantly improving the quality of enhanced low-light images. However, these images often suffer from varying degrees of color distortion. To address this issue, we have designed a neural network that separately processes image brightness and color information through distinct sub-networks. Specifically, the input low-light image is converted from the RGB color space to the HSV and Lab color spaces. We then extract features from the brightness and color channels both globally and locally. Finally, the extracted features are fused and decoded to produce the enhanced image. Additionally, we introduce edge loss and color loss functions combined with histogram matching loss and perceptual loss to optimize the training process of the model. By enhancing the restoration of edge and color information, the resulting enhanced images exhibit natural colors with clear details and textures. Extensive experiments demonstrate the effectiveness of our proposed algorithm, particularly in color recovery during low-light image enhancement. It achieves a PSNR of 38.40 and an LPIPS of 0.16 on the LSRW-Huawei dataset and a PSNR, SSIM, and LPIPS of 37.85, 0.85, and 0.14 respectively on the LOL-v2 dataset, Additionally, the algorithm demonstrates impressive enhancement performance on the RichIQA metric (Exploring Rich Subjective Quality Information for Image Quality Assessment in the Wild) and the NLIEE metric (A No-Reference Evaluation Metric for Low-Light Image Enhancement).