AM-CFDN: semi-supervised anomaly measure-based coal flow foreign object detection network

In underground coal mine transportation systems, conveyor belts are often affected by foreign objects such as gangue, steel wires, and wooden planks. Foreign objects not only potentially damage the equipment but also reduce the quality of the coal. Addressing the challenges of uneven underground lighting and low model detection accuracy in coal flow foreign object detection, this paper introduces the Anomaly Measure-based Coal Flow Detection Network (AM-CFDN). By introducing the Structural Similarity Index Measure-based Anomaly Evaluation Module (SAE), the AM-CFDN quantitatively assesses the anomaly degree of foreign objects in the coal flow by comparing the feature maps of input samples with selected template samples. Additionally, AM-CFDN employs a multi-task supervised training strategy, integrating both supervised and self-supervised loss functions, where the supervised losses include Cross-Entropy (CE) and mean squared error (L2) loss. The self-supervised loss utilizes Structural Similarity Index Measure (SSIM) loss. The detection performance of the model at both image and pixel levels is effectively enhanced by this approach. Experimental results on the Multi-Class Low-Resolution Foreign Objects in Coal Flow (LFOC) dataset, collected in a real underground coal mine environment, show that AM-CFDN achieves a detection speed of 30.30 frames per second (FPS), with an image-level AUROC of 0.952 and a pixel-level AUROC of 0.693. Compared to existing techniques such as PaDiM and DRAEM, AM-CFDN improves the image-level AUROC by 5.426% and 4.272%, respectively. These results demonstrate that AM-CFDN can effectively detect foreign objects in coal flow while maintaining suitable detection speed, making it suitable for practical coal flow foreign object detection scenarios and effectively ensuring the safety and efficiency of coal mine production.