A Hybrid Residual U-Net Architecture for Enhanced Low-Light Image Restoration

Low-light image enhancement (LLIE) is very essential for night-time surveillance, autonomous driving, and photography, yet traditional methods struggle with color distortion and texture loss. This paper proposes a Hybrid Residual U-Net, integrating residual learning with U-Net's multi-scale feature extraction to enhance brightness while preserving fine details. A hybrid loss function (SSIM + L1 loss) ensures perceptual quality and pixel accuracy. Trained on the LOL dataset, the model achieves 27.09 dB PSNR and 0.9031 SSIM, surpassing CNN and Transformer-based methods while maintaining real-time inference at 30.5 FPS. Future improvements will focus on self-supervised learning, adaptive enhancement for extreme low-light, and efficiency optimization. The proposed model offers a high-quality, real-time LLIE solution.