CUs-YOLO : Enhanced feature fusion model for coal and gangue recognition in complex environment of coal mine

Abstract This study proposes a novel model, CUs-YOLO, to address the challenges of detecting coal and gangue under complex conditions, including light spots from illumination, image blurring due to noise, and colour distortion. It also tackles limitations in local feature extraction and the tendency of existing models to lose target information. In the backbone network, the convolutional layers in CSPNet are enhanced using CondConv, which employs weighted convolutional kernels to increase model capacity while reducing computational cost. To mitigate information loss during upsampling in the pyramid structure of the original model's neck, this study improves the CARAFE operator by adding convolutional layers and replacing the original upsampling structure, thereby enhancing detail retention and reconstruction quality. Additionally, a dedicated coal and gangue data acquisition and detection device was developed, and a dataset was constructed to support experimentation. Experimental results demonstrate that the CUs-YOLO model achieved an average detection accuracy of 98%, a GFLOPs of only 2,570,599, and a real-time recognition speed of 60.2 FPS, confirming the effectiveness of the proposed enhancements. Comparative experiments further validate the superior performance of CUs-YOLO, which combines lightweight design with high recognition accuracy. This offers a promising solution for the intelligent identification of coal and gangue in complex environments, with significant practical application value.