A novel lightweight coal gangue image super-resolution network based on the global adaptive feature interaction transformer

Images captured in coal washing scenarios, the edges and textures of the coal gangue are often blurred, leading to misidentification in coal gangue detection. A novel lightweight Global Adaptive Feature Interaction Transformer model was proposed to achieve super-resolution of coal gangue images. First, a new multi-dimensional interactive attention mechanism was designed to extract global and local feature information. Then, a spatial gate structure was adopted and integrated with the Squeeze-and-Excitation module to form an efficient feedforward network, enhancing the model’s feature representation ability. Furthermore, a frequency domain loss was introduced to maintain consistency between global and local features, assisting the model in better capturing and restoring high-frequency information in images. Finally, the proposed method was tested on public datasets. At scaling factors of 2 and 4, it achieved PSNR values of 40.2415 dB and 33.0397 dB, SSIM values of 0.9637 and 0.8662, and LPIPS values of 0.0826 and 0.2618, respectively, outperforming existing models in all cases. Furthermore, the reconstructed images were tested on YOLOv8n, RT-DETR, Faster R-CNN and SSD, with an average precision improvement of 3.875%, particularly reaching an 8% improvement on YOLOv8n. The proposed method will effectively promote the identification of coal gangue.