A lightweight algorithm based on YOLOv5 for relative position detection of hydraulic support at coal mining faces

To solve the existing problems that the relative position detection method of hydraulic support in comprehensive coal mining faces, such as complex detection equipment, has poor flexibility and maintainability, we propose a method of hydraulic support relative position detection in coal mining faces based on object detection Lightweight-Ghost-YOLO (LG-YOLO). To better deploy the deep learning model on end-side devices, the GhostNet convolution module and the Ghost residual module are integrated into the YOLOv5s network to reduce the number of parameters and the occupancy of computing resources. Additionally, the Parametric Rectified Linear Unit (PReLU) activation function is integrated to achieve faster inference speed. In the postprocessing stage, Distance Intersection over Union Nonmaximum Suppression (DIoU-NMS) is used to improve the detection accuracy of closely aligned targets. The network model is further compressed by channel pruning and knowledge distillation. Experiments show that compared with YOLOv5s, the proposed algorithm can effectively detect the status of hydraulic supports. Finally, the model size is reduced by 73% and the computational amount is reduced by 69%, which can meet the requirements of end-side device deployment and real-time detection.