Abstract: In the complex working conditions of coal mines characterized by non-uniform illumination distribution, interference from coal dust and water mist, and coupled electromagnetic noise, mine images commonly suffer from challenges such as resolution degradation and blurred texture details. Traditional super-resolution algorithms, due to their high network complexity and large memory footprint, struggle to meet the real-time deployment requirements on edge devices. To address these issues, a lightweight coal mine image super-resolution reconstruction method based on multi-scale features and frequency attention is proposed. Firstly, the network introduces a multi-receptive-field attention module based on pixel attention and multi-scale feature perception. This module enhances the representation of image textures and edges by capturing spatial contextual information at different scales. Secondly, an adaptive weighting multi-scale feature fusion mechanism based on Transformer is designed, which integrates features extracted by multi-scale convolutions through learnable weights. Combined with a gated depth-wise convolutional feed-forward network, it further improves global structural consistency of the reconstructed image. Finally, a frequency attention mechanism is introduced during the upsampling stage to guide the reconstruction of high-frequency details using frequency domain information. Experimental results demonstrate that the proposed method not only outperforms existing mainstream algorithms in both quantitative evaluation metrics and subjective visual analysis but also achieves a better balance between model performance and lightweight design. At a scaling factor of 4，compared to the lightweight model IMFRN, the proposed model has a parameter count of only 617K, representing a 22% reduction, while improving the objective metrics PSNR and SSIM by 0.05 dB and 0.0041, respectively, on the complex-texture Urban100 dataset, and by 0.11 dB and 0.0031 on the coal mine test set. The results prove that this method can effectively extract multi-scale texture features of images, achieve cross-level feature correlation，and reconstruct high-resolution images with clearer high-frequency details under lightweight constraints, providing an efficient solution for deployment on mine edge devices.