Integration of Lightweight Network and Attention Mechanism for the Coal and Gangue Recognition Method

Recognition technology for coal and gangue is a crucial aspect of intelligent mine development. It plays a significant role in the advancement of mining technologies. Traditional methods of identification not only require a significant amount of material and financial resources, but also exhibit a low accuracy rate in identification. A novel method, GCEB-YOLO, is proposed for intelligent coal and gangue recognition by integrating a lightweight network with an attention mechanism. This method utilizes YOLO5s as the baseline model. This method aims to address the issues of noise interference, low illumination, motion blur, and the mixing of coal and gangue in the belt transport process at coal preparation plant belt conveyors. These factors have led to challenges in accurately identifying coal and gangue, as well as poor real-time detection, misdetection, and leakage of detection. Firstly, the Ghost Net serves as the backbone network, effectively reducing the number of model parameters, floating-point computations, and overall model size. Secondly, the Coord Attention (CA) mechanism is integrated into the backbone network, effectively enhancing the model’s focus on coal and gangue by emphasizing the positional information of coal and gangue in the image. Subsequently, the Efficient Channel Attention (ECA) mechanism is incorporated into the feature fusion network. This allows the model to prioritize important information from feature maps of different scales during the fusion process, leading to a significant improvement in coal and gangue detection accuracy. Finally, the weighted Bidirectional Feature Pyramid Network (Bi FPN) mechanism is utilized in the feature fusion process to eliminate less efficient feature transmission nodes in the multi-scale feature fusion process. This approach achieves different scale feature fusion in a comprehensive manner, effectively enhancing the efficiency of feature fusion. The GCEB-YOLO model demonstrates excellent detection performance in the presence of noise interference, low illumination, motion blur, and coal and gangue mixing conditions. This model achieves a mean average precision of 91.4%, representing a 2.8% improvement compared to YOLOv5s. It also demonstrates a real-time detection speed of 115 frames/s, showing a significant improvement of 27.6%. Additionally, the model parameters count, floating-point computation, and model size are measured at 3.8 M, 8.5G, and 8.2 MB respectively, all indicating improvements over YOLOv5s. Compared to YOLOv3-Tiny, YOLOv7-Tiny, YOLOv8s, and YOLOv9 models, the GCEB-YOLO model has a high recognition rate and recognition speed. It also effectively addresses issues related to leakage and misdetection, thereby providing a technical solution for intelligent recognition of coal and gangue.