Abstract: Intelligent coal gangue sorting is one of the key technologies for advancing intelligent and green coal mining. Accurate identification of coal gangue is a prerequisite for intelligent sorting, and overcoming the challenges posed by complex underground environments has become a critical issue affecting machine vision recognition. For underground coal transportation systems, we constructed an 11-class image dataset with single, dual-mixed, and triple-mixed backgrounds, including interference factors such as water stains, coal dust, broken coal, broken gangue, and components of transportation equipment. We propose a high-precision coal gangue recognition method that integrates multiple attention mechanisms and residual connections. In the high-resolution stage, residual convolution blocks are used to quickly generate high-quality tokens, and in the subsequent stages, deep feature representation learning is performed using a cascaded Channel Spatial Swin Transformer Block (CSSTB). To improve the model's robustness against background noise, the network incorporates global, channel, and spatial attention mechanisms, enhancing feature expression. The CSSTB leverages a LeakyReLU-based linear attention mechanism to model global information, strengthening sparse activation through its negative slope characteristics, while the Convolutional Block Attention Module (CBAM) is utilized to optimize attention distribution and improve model generalization. Additionally, considering the scale differences between coal, gangue, and equipment components, residual connections are applied across stages to enhance communication and information flow between multi-scale features. The results show that the proposed model achieves average accuracies of 95.06%, 97.77%, and 95.65% on single, dual-mixed, and triple-mixed backgrounds, respectively, representing improvements of 7.01%, 4.83%, and 1.03% compared to the baseline Swin Transformer-Tiny network. Visualization experiments demonstrate that, unlike the baseline model, which struggles to accurately distinguish between coal and gangue under complex background interferences such as water stains, low light, and reflections, the proposed model can precisely focus on key feature regions of coal and gangue, exhibiting strong anti-interference capabilities. The findings provide a theoretical reference for efficient coal-gangue sorting in underground coal transportation.