Research on Road Crack Detection Based on Lightweight Networks

To address the issue of limited computational resources when deploying road crack detection on edge devices, this paper proposes a lightweight detection framework that replaces the original backbone network in YOLOv5 with GhostNet and uses the differential evolution algorithm (DE) to recluster anchor boxes to match the target size distribution. Experiments were conducted on the China_Drone subset of the RD2022 dataset, with 600 training epochs and the Adam optimizer set as unified hyperparameters. Results show that GhostNet-DE reduces GFLOPs from 15.8 to 7.7 while maintaining nearly the same performance as the original model, halving the model size and improving inference efficiency by approximately 1.9 times. Compared to the GhostNet version without DE, [email protected] enhanced by 2.8 %. This method provides an efficient, low-cost solution for real-time road crack detection, suitable for deployment in edge scenarios such as roadside sensing units and drones.