Multi-Scale Low-Light Image Enhancement Network Based on U-Net

Abstract

Low light is a common phenomenon when shooting at night.Insufficient illumination causes serious loss of image details and reduces visual quality.The existing low-light image enhancement methods have insufficient perception and expression of features at different scales.To address the problem of existing low-light image enhancement methods being inadequate in their ability to perceive and express features at different scales, a multi-scale low-light image enhancement network based on U-Net(MSU-LIIEN) is proposed.Firstly, the feature pyramid is used as the basic processing framework of this article to achieve feature extraction for low-light images.Then, the U-Net is used as the backbone in all three branch structures of the feature pyramid construction to encode and decode the extracted shallow image features, while structural detail residual fusion blocks are introduced to enhance the network's ability to extract and characterize low-light image feature information.Finally, the extracted feature information is fused layer by layer to recover the final image.The experimental results show that, compared with the second-performing KinD algorithm in LOL-datasets, the average Peak Signal-to-Noise Ratio(PSNR) value increased by 16.21%, and compared with the second-performing model on the Brighting Train dataset, the average PSNR value increased by 49.67%.The proposed algorithm outperforms other classical low-light image enhancement algorithms in terms of both subjective visual field perception and objective evaluation metrics.Not only does it effectively enhance the overall brightness of low-light images, but it also maintains detailed information in the image and clear object outlines, making the overall picture of the enhanced image realistic and natural.