Blind super-resolution reconstruction method for underground coal mine images based on degradation kernel diffusion

Existing image super-resolution reconstruction methods have difficulty coping with multi-source coupled degradations such as coal dust scattering and non-uniform blur in underground coal mine environments, and they are limited by local receptive fields, making it difficult to capture global structures, while excessive model complexity prevents lightweight deployment on underground edge devices. To address these issues, a blind super-resolution reconstruction method for underground coal mine images based on degradation kernel diffusion was proposed. In the degradation modeling stage, degradation kernel diffusion modeling was introduced, and the reverse sampling process of a diffusion probabilistic model was used to explicitly simulate the degradation kernel distribution in complex underground coal mine scenes, thereby correcting reconstruction artifacts caused by degradation estimation bias at the early stage. In the image reconstruction stage, a hybrid Transformer-CNN encoder and a dynamic invertible decoder were designed, in which a parallel dual-branch structure was used to complementarily extract local textures and global dependencies, and a dynamic proportional fusion mechanism was employed to achieve adaptive interaction between degradation features and image content, reducing the number of model parameters while ensuring lossless transmission of deep features. By combining L1 loss, Structural Similarity (SSIM) loss, and perceptual loss, a multi-metric joint loss function was constructed to enhance perceptual image quality while ensuring pixel-level accuracy. Experiments were conducted on the CMUID underground coal mine image dataset and public benchmark datasets. The results showed that, in terms of objective evaluation metrics, the proposed method achieved overall superior performance in peak signal-to-noise ratio and SSIM compared with competing methods while maintaining a lower parameter count. In terms of subjective visual quality, the proposed method effectively suppressed low-light noise, sharpened edge structures, and clearly restored texture details of conveyor belts and coal blocks in underground coal mine scenes.