Research on coal gangue detection in coal preparation plant based on YOLOv5s-FSW model

Abstract

A coal gangue detection method in coal preparation plant based on YOLOv5s-FSW model is proposed to address the problems of insufficient feature extraction, large parameter quantity, low detection precision, and poor real-time performance in existing coal gangue detection models. This model is improved on the basis of YOLOv5s. Firstly, the C3 module in the Backbone section is replaced with a FasterNet Block structure, which improves detection speed by reducing the number of model parameters and computation. Secondly, in the Neck section, a parameter free SimAM attention mechanism is introduced to enhance the model's attention to important targets in complex environments, further improving the model's feature extraction capability. Finally, in the Prediction layer, the CIoU bounding box loss function is replaced with Wise-IoU, and the model focuses on ordinary quality anchor boxes to improve convergence speed and bounding box detection precision. The results of the ablation experiment indicate that compared with the YOLOv5s model, The mean average precision (mAP) of the YOLOv5s-FSW model has been improved by 1.9%, the model weight has been reduced by 0.6 MiB, the number of parameters has been reduced by 4.7%, and the detection speed has been improved by 19.3%. The comparative experimental results show that the YOLOv5s-FSW model has a mAP of 95.8%, which is 1.1%, 1.5%, and 1.2% higher compared to the YOLOv5s-CBC, YOLOv5s-ASA, and YOLOv5s-SDE models, respectively, and compared to YOLOv5m, YOLOv6s improved by 0.3%, 0.6% respectively. The detection speed of the YOLOv5s-FSW reaches 36.4 frames per second, which is 28.2% and 20.5% higher than the YOLOv5s-CBC and YOLOv5s-ASA models, respectively. Compared to YOLOv5m, YOLOv6s and YOLOv7, the detection speed of the YOLOv5s-FSW has increased by 16.3%, 15.2%, and 45.0%, respectively. The visualization experiment results of the thermal map show that the YOLOv5s-FSW model is more sensitive to the target feature areas of coal gangue and has higher attention. The detection experiment results show that in complex scenes with dim environments, blurred images, and mutual occlusion of targets, the YOLOv5s-FSW model has a higher confidence score for coal gangue target detection than the YOLOv5s model, and effectively avoids the occurrence of false positives and missed detection.

Keywords

• coal gangue detection
• yolov5s
• fasternet block
• simam attention mechanism
• wise iou bounding box loss function