Foreign object detection for coal mine conveyor belts based on improved YOLOv7

Foreign object detection on mine conveyor belts is of significant importance for ensuring coal mine safety production and enhancing automation levels. However, traditional machine vision methods are limited by low detection accuracy, high missed detection rates, and inadequate performance in identifying multi-scale and small-sized foreign objects, due to challenges such as insufficient underground illumination, strong point light interference, and complex backgrounds. Additionally, the limited availability of foreign object samples constrains model training effectiveness. To address these issues, an improved YOLOv7 foreign object detection algorithm is proposed, which integrates a feature enhancement structure and a focused linear attention mechanism. First, a feature enhancement module with a multi-scale parallel architecture and switchable atrous convolution is introduced in the backbone network to expand the receptive field and improve small target feature extraction. Second, a focused linear attention mechanism is incorporated to suppress complex background interference while enhancing attention to critical foreign object regions. Third, an efficient dynamic serpentine convolution module is designed to adapt to the geometric characteristics of strip-shaped objects and strengthen semantic representation during feature fusion. Furthermore, a direction-aware SIOU loss function is employed to improve bounding box fitting accuracy and training convergence efficiency. Finally, a transfer learning strategy is implemented, where pre-training on generic datasets is leveraged before fine-tuning for coal mine conveyor belt detection tasks to mitigate small-sample training challenges. Experimental results on the CUMT-BELT dataset demonstrate that the proposed algorithm achieves a Precision of 89.3%, which represents a 9.3% improvement over the baseline YOLOv7 model. The method is also shown to significantly outperform other mainstream detection approaches in precision, recall, and mAP@0.5 metrics, while particularly demonstrating superior robustness and adaptability in multi-scale, small-target, and strip-shaped object detection tasks, thereby validating its effectiveness and practicality in complex mining scenarios.