Abstract: Under complex underground coal-mine conditions, image acquisition is affected by non-uniform illumination, coal-dust and water mist interference, and coupled electromagnetic noise. These adverse factors cause resolution degradation, blurred edges, and loss of fine texture details, thereby reducing the visual quality of mine images. Although conventional super-resolution methods can restore some information, their complex architectures and large memory footprints hinder their real-time deployment on resource-constrained edge devices. Therefore, mine image reconstruction requires a model that improves image quality while maintaining low parameter complexity and efficient feature processing. This requirement is particularly important in underground environments, where edge equipment must process degraded images with limited memory resources while maintaining reliable reconstruction quality as well as demonstrate practical real-time operating efficiency. To address these challenges, this paper proposes a lightweight coal-mine image super-resolution reconstruction method based on multiscale features and frequency attention. The network first introduces a multireceptive-field attention module combining pixel attention with multiscale feature perception. By capturing spatial contextual information through receptive fields of different sizes, this module aggregates complementary local and contextual features. Therefore, it strengthens the representation of texture boundaries, structural edges, and other regions necessary for recovering fine image details. The multiscale design also mitigates the limitations caused by extracting spatial information under only a single receptive field. Next, a Transformer-based adaptive weighted multiscale feature fusion mechanism is designed to integrate features extracted via multiscale convolutions. Learnable weights dynamically regulate the contribution of each feature scale, thus allowing the network to select information based on image content. This mechanism enhances interactions among features from different levels and improves the consistency of the reconstructed global structure. A gated depth-wise convolutional feed-forward network is incorporated to refine the fused representation while controlling the additional computational burden. Collectively, adaptive weighting and gated depth-wise convolution improve cross-level feature correlation as well as preserve complementary structural and texture information. Finally, a frequency attention mechanism is introduced during upsampling to guide the reconstruction of high-frequency details using frequency-domain information. The mechanism emphasizes frequency components related to edges and textures, thus allowing visually meaningful high-frequency content to be prioritized during reconstruction. Consequently, the model can recover sharper textures and clearer boundaries without substantially increasing the network depth or parameter count. The spatial and frequency modules function synergistically to improve detail reconstruction while satisfying the lightweight-design requirement of edge deployment. Experimental results demonstrate that the proposed method outperforms existing mainstream algorithms in terms of objective evaluation metrics and subjective visual comparisons. Additionally, it achieves a favorable balance between reconstruction performance and model compactness. At a scale factor of 4, the proposed model contains only 617k parameters, representing a 22% reduction compared with the lightweight IMFRN model. On the complex-texture Urban100 dataset, the model improves the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) by 0.05 dB and 0.0041, respectively. On the coal-mine test set, the corresponding improvements are 0.11 dB and 0.0031, respectively. These results indicate that the method effectively extracts multiscale texture features and establishes correlations across different feature levels. Under lightweight constraints, it reconstructs high-resolution images with clear high-frequency details and well-preserved structural information. The proposed method provides an efficient super-resolution solution for deployment on edge devices in underground coal mines.