AMSF ‐ DETR : A Lightweight Foreign Object Detection Algorithm for Coal Mine Conveyor Belts Based on Adaptive Multi‐Scale Feature Fusion

ABSTRACT

Aiming at the major risks of conveyor belt foreign objects on conveyor belts to coal mine production safety and efficiency, as well as the issues associated with inadequate feature extraction and limited adaptability to multi‐scale foreign objects in complex coal mine environment by existing deep learning methods, this study proposes a lightweight coal mine conveyor belt foreign object detection algorithm, AMSF‐DETR, based on dual‐domain enhancement and dynamic multi‐scale feature fusion. The algorithm innovatively designs three core modules: First, a new backbone network, Adaptive Perception Network (APN), is designed to significantly enhance the model's ability to capture the textural and morphological features of foreign objects and improve its adaptability to objects of varying shapes. Secondly, the Bi‐domain Enhanced Transformer (BDET) is proposed. By combining Bipolar Linear Attention (BLA) and Frequency Domain Modulation Network (FDMN), the collaborative optimization of spatial and frequency domain features is realized, and the perception ability of the model to foreign objects is enhanced. In addition, the Dynamic Multi‐Scale Fusion Module (DMFM) is proposed to significantly enhance the representation ability of foreign object features through multi‐path feature extraction and fusion mechanism. Experimental results show that AMSF‐DETR achieves 92.1% accuracy and 87.5% mAP0.5, with only 13.6 million parameters and an inference speed of 90.4 FPS. Compared with the baseline model, it achieves a relative improvement of 3.8% in accuracy and 3.9% in mAP0.5 respectively, achieves a relative reduction of 33.3% in parameters and a relative increase of 35.9% in inference speed, providing an efficient solution for foreign object detection on coal mine conveyor belts.