RNMT-Net: A Retinex-Guided NAFNet–Mamba–Transformer Network for Low-Light Image Enhancement

Abstract

Low-light image enhancement (LLIE) remains challenging because a single network must simultaneously correct global illumination, suppress sensor noise, and preserve fine texture and color fidelity. Existing approaches typically specialize in only one of these sub-problems: Retinex-based methods model illumination well but under-model noise; convolutional restoration networks such as NAFNet are efficient but lack long-range receptive fields; and Transformer- or Mamba-based restorers capture global context but are rarely combined with an explicit noise model or a principled multi-task loss balance. We propose RNMT-Net, a unified encoder–decoder architecture that fuses (i) Nonlinear-Activation-Free (NAFNet) convolutional blocks for efficient local feature extraction, (ii) a cross-scan Selective State-Space (Mamba) module for linearcomplexity long-range modeling, (iii) a lightweight channel-wise Transformer block, (iv) a Signal-to-Noise-Ratio (SNR)- guided local/global fusion gate, and (v) an amplitude–phase Fourier enhancement module, all supervised through a Retinexinspired reflectance–illumination–noise decomposition with automatic uncertainty-based loss weighting and a progressive patch-size curriculum. RNMT-Net is trained in two stages — synthetic pre-training on LOL-v2-Synthetic followed by finetuning on LOL-v1 — and evaluated with exponential moving average (EMA) weights and flip-based test-time augmentation. On the LOL-v1 benchmark, RNMT-Net achieves 26.26 dB PSNR and 0.8468 SSIM, outperforming representative Retinexbased, CNN-based, and Transformer-based baselines reported in the literature. Beyond PSNR/SSIM, we additionally report LPIPS and NIQE to assess perceptual quality and naturalness.