中国科技论文

针对煤矿井下传送带作业场景中光照不足与检测目标形态多变等因素造成检测模型存在错检、漏检现象，且检测模型难以在边缘设备部署的问题，提出了一种基于改进YOLOv8的煤矿井下传送带异物检测算法(SENet-CCFM-Dynamic Head-YOLO,SCD-YOLO)。首先，采用SENetV2改进C2f(CSP bottleneck with 2 convolutions)模块，增强特征表示，提高模型性能；其次，使用CCFM(cross-scale feature-fusion module)对颈部进行改进，在保证检测效果的同时，大幅减少模型参数；此外，将传统的检测头替换为动态检测头(Dynamic Head)，强化模型对小目标的检测能力；最后，采用ShapeIoU(shape intersection over union)损失函数替换CIoU损失函数，更精细地衡量预测目标与真实目标的形状匹配程度。采用煤矿井下传送带异物检测数据集对所提出的模型进行测试，实验结果表明：SCD-YOLO的mAP@50、mAP@50:95、召回率(Recall)分别达到96.3%、82.8%、92.2%，相较于基线模型YOLOv8n分别提升了3.6%、2.3%和3.8%;SCD-YOLO的参数量仅有2.65×10⁶，计算量为6.7×10⁹，相较于基线模型YOLOv8n分别下降了11.7%和18.3%。

The poor illumination and high variability in object shapes bring significant challenges to foreign object detection in underground coal mine conveyor belt scenarios, which leads to not only false detections and missed detections, but also difficulties in deployment on resource-constrained edge devices. To address these issues, a foreign object detection algorithm in underground coal mine conveyor belt(SCD-YOLO) based on an improved YOLOv8 architecture was proposed. Firstly, the C2f(CSP bottleneck with 2 convolutions) module was enhanced with SENetV2 to improve feature representation and boost detection performance. Secondly, crossscale feature fusion module(CCFM) was introduced in the neck to reduce model parameters while preserving accuracy. Thirdly, the conventional detection head was replaced with a Dynamic Head to enhance the model's capability in detecting small objects. Finally, the ShapeIoU loss function was employed to replace the traditional CIoU loss, enabling more precise evaluation of the geometric alignment between predicted and ground truth objects. The proposed model was evaluated on a customized data set of foreign object detection in coal mine conveyor belts. Experimental results demonstrate that SCD-YOLO(SENet-CCFM-Dynamic Head-YOLO) achieves an mAP@50 of 96. 3%, mAP@50:95 of 82. 8%, and a Recall of 92. 2%, outperforming the baseline YOLOv8n by 3. 6%, 2. 3%, and 3. 8%, respectively. Furthermore, SCD-YOLO contains only 2. 65×10⁶ parameters and requires 6. 7×10⁹ FLOPs, representing reductions of 11. 7% and 18. 3% compared to the baseline, thereby proving its suitability for edge deployment in industrial environments.