Zhu Lei, Gu Wenzhe, Liu Chengyong, Zhang Beiyan, Liu Wentao, Yuan Chaofeng
China Coal Energy Research Institute Co., L td., Xi'an, Shaanxi, China.
China Coal Xi'an Design Engineering Co., L td., Xi'an, Shaanxi, China.
PLoS One. 2025 May 9;20(5):e0321249. doi: 10.1371/journal.pone.0321249. eCollection 2025.
Coal and gangue recognition technology is one of the key technologies in the intelligent construction of coal mines. With the deepening of the research, only the coal and gangue recognition, the pixel segmentation of the coal gangue image is needed. Aiming at the gangue segmentation algorithm with low accuracy, easy to miss detection, wrong detection and large amount of detection data, slow detection speed and other problems. A coal gangue segmentation model based on improved YOLOv8 is proposed to achieve fast and accurate recognition of coal gangue images, and the overall computational volume of the model is not large, which has achieved better application results. Using the YOLOv8 model as the base model, the standard convolutional modules in the first, second & third C2f modules were replaced with depth separable convolution (DSC) modules in the YOLOv8 model backbone network, reducing the overall computational effort of the model. Adding the CBAM module before the second convolution of the up-sampling module and down-sampling stage in the model neck network improves the differentiation of the model for gangue and enhances the recognition accuracy. The original dataset was expanded from 1980 to 11,265 sheets using data expansion techniques and some hyperparameters were adjusted. Results show that the improved YOLOv8 model has an accuracy (Precision) of 95.67%, a recall (Recall) of 95.74%, a transmitted frames per second (FPS) of 32.11 frames/s, and a mean average precision (mAP) of 96.88%, which is an improvement of 5.6% in accuracy, 7.12% in recall, and the mean average precision (mAP) is improved by 4.65% and FPS is improved by 8.83 frames/s. By comparing with YOLOv3, YOLOv5, YOLOv7, and YOLOv8 models, the improved model is optimal in terms of accuracy and speed. Finally, the model is successfully applied to underground coal gangue image segmentation through transfer learning, and the effect of coal gangue image segmentation is good, which verifies the re-liability of the algorithm.
煤矸识别技术是煤矿智能化建设的关键技术之一。随着研究的深入，仅对煤和矸石进行识别，需要对煤矸石图像进行像素分割。针对矸石分割算法存在的准确率低、易漏检、误检、检测数据量大、检测速度慢等问题，提出了一种基于改进YOLOv8的煤矸石分割模型，以实现对煤矸石图像的快速准确识别，且模型的整体计算量不大，取得了较好的应用效果。以YOLOv8模型为基础模型，将YOLOv8模型主干网络中第一、第二和第三C2f模块中的标准卷积模块替换为深度可分离卷积（DSC）模块，降低了模型的整体计算量。在模型颈部网络的上采样模块和下采样阶段的第二次卷积之前添加CBAM模块，提高了模型对矸石的区分能力，增强了识别准确率。利用数据扩充技术将原始数据集从1980张扩充到11265张，并调整了一些超参数。结果表明，改进后的YOLOv8模型准确率（Precision）为95.67%，召回率（Recall）为95.74%，每秒传输帧数（FPS）为32.11帧/秒，平均精度均值（mAP）为96.88%，准确率提高了5.6%，召回率提高了7.12%，平均精度均值（mAP）提高了4.65%，FPS提高了8.83帧/秒。通过与YOLOv3、YOLOv5、YOLOv7和YOLOv8模型进行比较，改进后的模型在准确率和速度方面最优。最后，通过迁移学习将该模型成功应用于井下煤矸石图像分割，煤矸石图像分割效果良好，验证了算法的可靠性。
