Abstract:
The existing deep learning based foreign object detection models for conveyor belts are relatively large and difficult to deploy on edge devices. There are errors and omissions in detecting foreign objects of different sizes and small objects. In order to solve the above problems, a foreign object detection method for coal mine conveyor belts based on improved YOLOv8 is proposed. The depthwise separable convolution, squeeze-and-excitation (SE) networks are used to reconstruct the Bottleneck of the C2f module in the YOLOv8 backbone network as a DSBlock, which improves the detection performance while keeping the model lightweight. To enhance the capability to obtain information from objects of different sizes, an efficient channel attention (ECA) mechanism is introduced. The input layer of ECA is subjected to adaptive average pooling and adaptive maximum pooling operations to obtain a cross channel interactive MECA module, which enhances the global visual information of the module and further improves the precision of foreign object recognition. The method modifies the 3 detection heads of YOLOv8 to 4 lightweight small object detection heads to enhance sensitivity to small objects and effectively reduce the missed and false detection rates of small object foreign objects. The experimental results show that the improved YOLOv8 achieves a precision of 91.69%, mAP@50 reached 92.27%, an increase of 3.09% and 4.07% respectively compared to YOLOv8. The detection speed of improved YOLOv8 reaches 73.92 frames/s, which can fully meet the demand for real-time detection of foreign objects on conveyor belts in coal mines. The improved YOLOv8 outperforms mainstream object detection algorithms such as SSD, Faster-RCNN, YOLOv5, and YOLOv7-tiny in terms of precision, mAP@50, number of parameters, weight size, and number of floating point operations.