School of Mechanical Engineering, Anhui University of Science and Technology
Collaborative Innovation Center of Mine Intelligent Equipment and Technology, Anhui University of Science and Technology
Anhui Zhongke Photoelectric Color Sorter Machinery Co., Ltd., Postdoctoral Research Station

现有基于深度学习的煤矸识别方法在煤矿井下低照度、高噪声及运动模糊等复杂工况下存在煤矸识别精度低、小目标煤矸容易漏检、模型参数量和运算量大，难以部署到计算资源有限的设备中等问题，提出了一种基于HGTC−YOLOv8n模型的煤矸识别算法。采用HGNetv2网络替换YOLOv8n的主干网络，通过多尺度特征的有效提取，提高煤矸识别效果并减少模型的存储需求和计算资源消耗；在主干网络中嵌入三重注意力机制模块Triplet Attention，捕获不同维度间的交互信息，增强煤矸图像目标特征的提取，减少无关信息的干扰；选用内容感知特征重组模块（CARAFE）来改进YOLOv8n颈部特征融合网络上采样算子，利用上下文信息提高感受视野，提高小目标煤矸识别准确率。实验结果表明：① HGTC−YOLOv8n模型的平均精度均值为93.5%，模型的参数量为2.645×106，浮点运算量为8.0×109 ，帧速率为79.36帧/s。② 平均精度均值较YOLOv8n模型提升了2.5%，参数量和浮点运算量较YOLOv8n模型分别下降了16.22%和10.11%。③ 与YOLO系列模型相比，HGTC−YOLOv8n模型的平均精度均值最高，且参数量和浮点运算量最少，检测速度较快，综合检测性能最佳。④ 基于HGTC−YOLOv8n模型的煤矸识别算法在煤矿井下复杂工况下，改善了煤矸识别精度低、小目标煤矸容易漏检等问题，满足煤矸图像实时检测要求。

The existing deep learning based coal gangue recognition methods have problems in complex working conditions such as low lighting, high noise, and motion blur in coal mines, such as low precision of coal gangue recognition, easy omission of small target coal gangue, large model parameter and computational complexity, and difficulty in deploying to devices with limited computing resources. A coal gangue recognition algorithm based on the HGTC-YOLOv8n model is proposed. The method replaces the backbone network of YOLOv8n with HGNetv2 network, effectively extracts multi-scale features to improve coal gangue recognition performance and reduces model storage requirements and computational resource consumption. The method embeds a Triplet Attention mechanism module in the backbone network to capture interaction information between different dimensions. The method enhances the extraction of target features in coal gangue images, and reduces the interference of irrelevant information. The method selects the content aware reassembly of features(CARAFE) to improve the upsampling operator of YOLOv8n neck feature fusion network, utilizing contextual information to enhance perceptual field of view and improve the accuracy of small target coal gangue recognition. The experimental results show the following points.① The average precision of the HGTC-YOLOv8n model is 93.5%, the parameters number of the model is 2.645×106, the number of floating-point operation is 8.0×109, and the frame rate is 79.36 frames/s. ② The average precision of the YOLOv8n model has increased by 2.5% compared to the YOLOv8n model, and the number of parameters and floating-point operations have decreased by 16.22% and 10.11%, respectively. ③ The comparison results with the YOLO series models show that the HGTC-YOLOv8n model has the highest average precision, the least number of parameters and floating-point operations, fast detection speed, and the best overall detection performance. ④ The coal gangue recognition algorithm based on the HGTC-YOLOv8n model has improved the low precision of coal gangue recognition and the easy omission of small target coal gangue under complex working conditions in coal mines. The method meets the requirements of real-time detection of coal gangue images.

机械工业联合会矿山采选装备智能化重点实验室开放基金项目（2022KLMIO4）；安徽理工大学引进人才基金项目（13230411）。

滕文想，王成，费树辉. 基于HGTC−YOLOv8n模型的煤矸识别算法研究[J]. 工矿自动化，2024，50（5）：52-59.

TENG Wenxiang, WANG Cheng, FEI Shuhui. Research on coal gangue recognition algorithm based on HGTC-YOLOv8n model[J]. Journal of Mine Automation，2024，50（5）：52-59.