Retinex-Based Network for Low-Light Image Enhancement with Noise Reduction and Artifact Alleviation

Low-light image enhancement is a crucial low-level task in computer vision, with Retinex theory widely employed to model low-light images by decoupling color information from illumination. In conjunction with Retinex theory, deep learning approaches have been introduced to enhance low-light images for their powerful feature learning capabilities. However, CNN-based methods often face challenges in effectively capturing global information, and Transformer-based models require significant computational resources. This paper proposes a lightweight Retinex-based approach consisting of brightening and denoising stages. We design the Illumination-Guided Simple Channel Attention (IGSCA) module for feature processing and adaptive enhancement, specifically tailored to guide attention according to illumination characteristics. The Channel Expansion Self-Activating Gate (CESAG) module is applied to activate and transmit features, effectively preserving textures and reducing noise. The proposed method achieves efficient noise reduction and artifacts alleviation with fewer parameters and lower computational cost. Quantitative and qualitative results confirm the effectiveness of this approach.