Wavelet-Based Diffusion Model for Low-Light Image Enhancement under Nonuniform Illumination in Tunnel Environments

Cameral surveillance has become a crucial data collection method in the operation and maintenance of tunnel environments. However, because it relies entirely on artificially arranged light sources for illumination, the image data collected are often affected by insufficient lighting or localized overexposure. These issues significantly hinder downstream recognition tasks, such as detecting personnel activities, monitoring system status, and assessing environmental conditions within tunnels. To address these challenges, this study proposes a low-light enhancement deep learning model (DTLL). The model integrates diffusion-based enhancement techniques with a customized detail restoration module and an innovative combination of adaptive wavelet decomposition to improve low-light image quality in tunnel scenarios. On the publicly available LoLv1 data set and a real-world tunnel data set, the DTLL model achieved a peak signal-to-noise ratio (PSNR) of 24.690, indicating reduced noise and higher reconstruction fidelity; a structural similarity index measure (SSIM) of 0.879, suggesting a high degree of structural preservation; a Brenner score of 0.0304, reflecting improved image sharpness; entropy of 5.1862, representing richer image information; and edge intensity of 0.0271, highlighting clearer edge features. These metrics collectively confirm the model’s ability to enhance image clarity, detail, and overall visual quality. The proposed method has strong potential for real-time deployment in tunnel monitoring systems, enabling more accurate detection and decision-making in transportation, construction, and emergency response scenarios.