BDL: transformer-based super-resolution network for degraded underground coal mine images

Abstract Underground coal mine images often suffer from severe blurring and low-resolution degradation due to harsh lighting, dust, and machinery motion, which hinder accurate visual inspection and automated analysis. This study proposes a transformer-based super-resolution (SR) network that integrates local convolution with adaptive interaction mechanisms for effective local–global feature modeling. The network employs a hierarchical architecture consisting of shallow feature extraction, cascaded spatial and channel transformer blocks, and a reconstruction module. Each transformer block incorporates a bidirectional adaptive interaction module (BAIM) to fuse convolutional local features with transformer-based global representations through adaptive reweighting in both spatial and channel dimensions. A dual-group feedforward network (DGFN) decouples channel feature preservation from spatial information enhancement, while cross-group interactions ensure balanced channel modeling and spatial perception without information loss. Additionally, a local convolution block (LCB) with SE-based channel weighting is used to restore fine-grained details. Extensive experiments on both a dedicated coal mine dataset and public benchmarks demonstrate that the proposed method consistently outperforms existing state-of-the-art (SOTA) SR approaches. Specifically, for ×2 super-resolution, it achieves a PSNR/SSIM of 32.07/0.9688 on the coal mine dataset, improving over the previous best by 0.59 dB and 0.0036, respectively. For ×4 super-resolution, it attains 28.10/0.8836, surpassing the previous best by 0.24 dB and 0.0013. Similar improvements are observed on public datasets, confirming the method’s effectiveness in both general and challenging industrial scenarios.