To address the challenges of complex underground coal mine environments characterized by uneven lighting, strong background interference, and the simultaneous presence of various foreign objects of different sizes and materials on belt conveyors, which makes it difficult for existing detection methods to achieve both high accuracy and real-time performance, this paper proposes a lightweight real-time detection model, MSC-DETR. First, to enhance the model’s ability to model long-range dependencies of multi-scale foreign object features and significantly reduce parameter redundancy, a Multi-branch Adaptive Feature Enhancement Aggregation Module (MAFEA) is designed. Second, to improve the model’s robustness under complex noise and lighting variations, a Spectrum Enhancement Feed-Forward Network (SEFFN-AIFI) is proposed, combining multi-scale convolution and adaptive frequency domain filtering mechanisms. Finally, to improve the model’s cross-scale semantic consistency modeling and context awareness capabilities in complex multi-object scenarios, a Context-Guided Feature Enhancement Aggregation Pyramid Network (CEAFPN) is constructed. Through multi-scale contextual collaborative modeling and dynamic feature fusion, dual adaptive calibration of spatial and semantic features is achieved. Experiments on the CUMT-BelT dataset show that, compared to the RT-DETR baseline model, MSC-DETR reduces the number of parameters and GFLOPs by 55.2% and 70.2%, respectively, achieves an mAP50 of 93.3%, and increases FPS to 124.6. The model outperforms mainstream detection algorithms in multiple metrics, particularly demonstrating superior recognition stability and localization accuracy in multi-target, small-target, and complex background scenarios. It achieves an effective balance between accuracy, lightweight design, and inference speed, providing an efficient and reliable solution for real-time detection of various foreign objects on coal mine belt conveyors.
Zhao et al. (Thu,) studied this question.