MSFA-Net: A Low-Light Image Enhancement Method Based on Multi-Scale Feature Fusion and Attention Mechanism

Abstract

Low-light images are degraded by uneven illumination, sensor noise, color distortion, and weak texture boundaries. To address insufficient detail recovery, noise amplification, and inadequate coordination between global illumination and local texture, this paper proposes MSFA-Net, a compact residual enhancement network based on multi-scale feature fusion and hybrid channel–spatial attention. Each MSFA block extracts complementary information through standard 3×3, dilated 3×3, and 5×5 convolutional branches. A 1×1 fusion layer integrates local edges and large-receptive-field context, after which channel attention recalibrates global feature responses and spatial attention emphasizes informative underexposed regions. A learnable residual gate stabilizes optimization and limits over-enhancement. The network is trained with a composite objective combining pixel, structural, gradient, color-consistency, and total-variation losses. With four MSFA blocks and 32 feature channels, the baseline configuration contains approximately 0.192 million trainable parameters. A reproducible evaluation framework is defined on LOL and LOL-v2 using PSNR, SSIM, LPIPS, NIQE, edge preservation, noise statistics, and computational complexity. The proposed formulation focuses on a single algorithmic problem and provides an explicit mechanism for the coordinated improvement of illumination, texture detail, and visual fidelity.