Foreign object detection algorithm for mine conveyor belts based on YOLOv10n-CND

The detection of foreign objects on coal mine conveyor belts is a crucial step in ensuring the safety of coal transportation. A dataset containing 6,279 sample images with three types of detection targets was annotated using labelimg for the mine conveyor belt foreign object detection task. To address issues such as complex background in foreign object images, weak feature extraction capability, low detection accuracy for small adhered targets, and distortion in detection box positioning and scale, a mine conveyor belt object detection model based on YOLOv10n-CND is proposed. The C2f-DySnakeConv module is used to replace the C2f module, improving sensitivity to target shapes and boundaries, thereby solving the problem of weak feature extraction in underground foreign object recognition. A normalized Gaussian Wasserstein distance regression loss function is employed to improve the precision and efficiency of small target detection and localization. A scale-unified target detection head is added, combining three attention mechanisms—scale, spatial, and task— to obtain detection boxes for objects of different sizes, thus enhancing the model’s detection performance. A dataset and validation set were created using underground coal mine conveyor belt images. The final detection results are compared with eight existing detection algorithms, and experimental results show that the YOLOv10n-CND model achieves precision, recall, and average detection accuracy of 91.4%, 82.6%, and 87.9%, respectively. These improvements represent increases of 5.3%, 1.5%, and 2.6% compared to the original model, effectively ensuring the healthy development of coal mine transportation.