Abstract: Visual monitoring images at underground coal-mine mining and excavation faces and in adjacent working areas are affected by dust, water mist, low illumination, and local strong light, resulting in reduced contrast, obscured textures, color distortion, and local glare. Existing learning-based dehazing methods mainly rely on synthetic training data. When applied to real coal-mine scenes, they are susceptible to non-uniform degradation and domain shift, which can cause residual haze, distorted details, and unstable visual quality. To address these problems, a dust-fog image dehazing method based on depth prior and adaptive extinction modeling is proposed. A monocular relative inverse-depth prior is first used to construct an initial haze representation, providing the network with relative geometric information about the scene. Pixel-wise relative extinction intensity is then predicted using the Beer-Lambert attenuation formulation to generate a refined haze map. A haze attention mechanism enhances feature representations in heavily hazed regions and distant structures. Smoothness and consistency regularization of the predicted relative extinction intensity improve the stability of adaptive extinction prediction. To address the lack of strictly paired clear ground truth for real coal-mine dust-fog images, an enhanced pseudo-supervised real-domain fine-tuning strategy combines multi-candidate pseudo-ground-truth construction, selective updating, and joint synthetic-real supervision to improve transfer stability. Experiments show that the proposed method achieves favorable peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) results on public synthetic datasets. On the coal-mine dust-fog test set, the fine-tuned model obtains fog aware density evaluator (FADE), naturalness image quality evaluator (NIQE), and blind/referenceless image spatial quality evaluator (BRISQUE) values of 0.529, 11.707, and 28.117, respectively. Compared with hazy inputs, the pre-trained model, and the comparison methods, the proposed method achieves a favorable balance among residual haze suppression, structural detail preservation, and overall visual naturalness, providing a more stable image basis for underground visual monitoring and subsequent detection and recognition tasks.