MSFE-GAN: multi-scale feature extraction GAN for perceptually enhanced low-light images

Abstract

Low-light image enhancement plays a crucial role in downstream computer vision applications such as autonomous driving, semantic segmentation, and security surveillance. Conventional enhancement methods often overlook non-uniform illumination handling, which leads to overexposure, detail, and texture loss within the brightness distribution. As to address these limitations, MSFE-GAN (Multi-Scale Feature Extraction GAN) is proposed with a novel generative adverserial network (GAN) that incoporates spatial and frequency-domain processing as to achieve the overlooked perceptual quality and structural integrity in the enhanced low-light images. Unlike the traditional methods that apply uniform brightness adjustments, MSFE-GAN introduces a U-Net-based generator for multi-scale feature extraction and a Fourier-based refinement module for high-frequency detail and texture preservation. A dual-Markovian discriminator is employed within the GAN network to ensure global consistency and local texture fidelity, which produces a high-quality, visually coherent image enhancement result.