针对异物图像背景复杂、特征提取能力弱、粘连小目标的检测精度低、检测框定位及尺度失真等问题，提出了一种基于YOLOv5n−CND的矿用输送带目标检测算法。首先，采用C2f对特征金字塔进行优化，使用更少参数解决在井下异物图像采集背景复杂且存在复杂目标干扰对小目标检测不敏感的问题；然后，采用归一化高斯瓦萨斯坦距离(NWD)回归损失函数替代CIoU，改善多尺度异物检测效果不佳的问题，实现粘连小目标的精准检测；最后，添加目标检测头(Dy Head)，将尺度、空间和任务3种注意力机制结合，提高对异物轮廓的特征提取能力，增强对多尺度目标的适应能力。实验结果表明：YOLOv5n−CND的mAP@0.5、mAP@0.5∶0.95、参数量及检测速度分别为87.9%，55.9%，4.49×10⁶个，85.5帧/s，满足煤矿井下异物检测需求；YOLOv5n−CND的mAP@0.5和mAP@0.5∶0.95较YOLOv5n分别提高了2.6%和3.4%，较YOLOv5s−CBAM分别提高了1.7%和3.8%；模型参数量在YOLOv5n的基础上略有提升，但较其他模型参数量均有所降低。选取异物与背景相近的细长检测物、光照比较低的锚杆检测物、大量煤矸石混杂的检测物、含有多个异物4种场景进行测试，结果表明：基于YOLOv5n−CND的矿用输送带异物检测算法未出现误检及重复检测的情况，漏检较少，检测框定位准确，对粘连小目标的处理效果更好，能够实现输送带异物的准确检测。

To address the issues of complex background in foreign object images, weak feature extraction, low detection accuracy for adhering small objects, and distortion in detection box positioning and scale, a foreign object detection algorithm for mining conveyor belts based on YOLOv5n-CND is proposed. First, the C2f module was used to optimize the feature pyramid, using fewer parameters to address the issue of poor sensitivity to small object detection caused by the complex background in foreign object images and interference from complex objects in underground environments. Second, the normalized Gaussian Wasserstein distance (NWD) regression loss function was used to replace CIoU, improving the performance of multi-scale foreign object detection and accurately predicting the detection of adhering small objects. Finally, a detection head (DyHead) was added, combining three attention mechanisms: scale, spatial, and task, to enhance feature extraction for foreign object contours and improve the adaptability to multi-scale targets. Experimental results demonstrated that YOLOv5n-CND achieved an mAP@0.5 of 87.9%, an mAP@0.5:0.95 of 55.9%, a parameter count of 4.49×10⁶, and a detection speed of 85.5 frames per second, meeting the requirements for underground foreign object detection in coal mines. The mAP@0.5 and mAP@0.5:0.95 of YOLOv5n-CND were 2.6% and 3.4% higher than YOLOv5n, and 1.7% and 3.8% higher than YOLOv5s-CBAM, respectively. Although the model’s parameter count slightly increased compared to the YOLOv5n model, it was still lower than that of other models. Tests were conducted in four scenarios: foreign objects with elongated shapes resembling the background; anchor bolts with relatively low lighting; objects heavily mixed with coal gangue; and multiple foreign objects. The results indicated that the foreign object detection algorithm for mining conveyor belts based on YOLOv5n-CND did not result in false detections or duplicate detections, with very few missed detections. The detection box positioning was accurate, and the handling of adhering small objects was more effective, enabling precise detection of foreign objects on conveyor belts.