Integrating retinex theory with atmospheric scattering model for mixed-degradation image recovery

Images captured in low-light, hazy environments suffer from mixed degradation: insufficient illumination compresses the dynamic range, while atmospheric scattering obscures textural details. Existing methods, typically designed for single degradation factors, struggle to address these coupled effects. To overcome this limitation, we propose a unified physics-based model that integrates the atmospheric scattering equation with Retinex theory. Based on this model, we introduce the Mixed-Degradation Image Recovery (MDIR) framework, an end-to-end network specifically engineered for concurrent low-light enhancement and dehazing. To support data-driven training, we construct the Low-Light & Foggy Dataset (LFD), a novel collection of 1900 paired real-world images. Furthermore, MDIR incorporates a superpixel-guided, windowed self-attention mechanism to estimate depth-dependent transmittance, enabling precise, non-uniform haze removal. Extensive experiments on the LFD dataset demonstrate that MDIR significantly outperforms state-of-the-art techniques, achieving average gains of 7.61 dB in PSNR and 0.38 in SSIM. The dataset and code are available at https://github.com/MDIR-LFD/MDIR.