Lightweight pruning model for road distress detection using unmanned aerial vehicles

The size and complexity of the multiobjective detection model restrict its applicability to real-time road distress detection with unmanned aerial vehicles (UAVs). To address this issue, this paper proposes a lightweight approach that integrates a performance-aware approximation global channel pruning (PAGCP) algorithm and a channel-wise knowledge distillation method. YOLOv7-RDD was selected as the baseline model, and ablation tests were conducted to analyze the modules. The SIoU loss function demonstrated superior performance to CIoU, Wise IoU, and EIoU, while SimAM exhibited enhanced results compared to SE, CBAM, LSKA, and ELA attention mechanism modules. The integration of the PAGCP pruning model and the channel-wise knowledge distillation method resulted in a 17 % reduction in model size and a 79 % reduction in computational complexity while maintaining accuracy. The model exhibited satisfactory performance in the detection of four types of pavement distress based on UAV-collected image data, with an mAP of 0.712.