State Key Laboratory of Mining Response and Disaster Prevention and Control in Deep Coal Mines, Anhui University of Science and Technology
School of Mechanical Engineering, Anhui University of Science and Technology

由于煤矿井下高噪声、低照度、运动模糊的复杂工况和煤矸易聚集现象，导致煤矸目标检测模型特征提取困难及煤矸分类、定位不准确问题。针对该问题，提出一种基于CED−YOLOv5s模型的煤矸识别方法。首先，在YOLOv5s主干网络中引入坐标注意力 (CA) 机制，通过将坐标信息嵌入信道关系和长程依赖关系中对特征图进行编码，充分利用通道注意力信息和空间注意力信息，使模型更加关注重要特征，抑制无用信息。其次，在YOLOv5s的检测头部引入EIoU回归损失函数，将目标框与锚框的宽高差异最小化，以增强目标的位置和边界信息，提高模型在密集目标下的定位精度和收敛速度；最后，在YOLOv5s的检测头部引入轻量化解耦头，解耦出单独的特征通道，分别用于分类任务和回归任务，解决了原模型中耦合头部分类任务与回归任务的相互干扰问题，进一步提升了模型的并行运算效率与检测精度。实验结果表明： CED−YOLOv5s模型与其他YOLO系列目标检测模型相比，综合性能最佳，平均检测精度达94.8%，相较于YOLOv5s模型提升了3.1%，检测速度达84.8 帧/s，可充分满足煤矿井下煤矸实时检测需求。

Due to the complex working conditions of high noise, low illumination, and blurred movement in coal mines underground, as well as the phenomenon of coal gangue easily gathering, it is difficult to extract features from coal gangue object detection models. The classification and positioning of coal gangue are inaccurate. In order to solve the above problems, a coal gangue recognition method based on the CED-YOLOv5s model is proposed. Firstly, the coordinate attention (CA) mechanism is introduced into the YOLOv5s backbone network, which encodes feature maps by embedding coordinate information into channel relationships and long-range dependencies. The method fully utilizes channel attention information and spatial attention information to make the model focus more on important features and suppress irrelevant information. Secondly, the EIoU regression loss function is introduced in the detection head of YOLOv5s to minimize the width and height difference between the object box and anchor box. It enhances the position and boundary information of the object, improves the positioning precision and convergence speed of the model in dense objects. Finally, a lightweight decoupling head is introduced in the detection head of YOLOv5s, decoupling separate feature channels for classification and regression tasks. It solves the interference problem between the coupling head part of the class task and the regression task in the original model, further improving the parallel operation efficiency and detection precision of the model. The experimental results show that the CED-YOLOv5s model has the best overall performance compared to other YOLO series object detection models. It has an average detection precision of 94.8%, an improvement of 3.1% compared to the YOLOv5s model, and a detection speed of 84.8 frames/s. The results can fully meet the real-time detection requirements of coal gangue in coal mines.

安徽高校协同创新资助项目（GXXT-2021-076）。

何凯，程刚，王希，等. 基于CED−YOLOv5s模型的煤矸识别方法研究[J]. 工矿自动化，2024，50（2）：49-56, 82.

HE Kai, CHENG Gang, WANG Xi, et al. Research on coal gangue recognition method based on CED-YOLOv5s model[J]. Journal of Mine Automation，2024，50（2）：49-56, 82.