Embedded lightweight detection system for road surface defect using model pruning and knowledge distillation

Accurate and real-time road surface defect inspection is critical for ensuring traffic maintenance. However, existing defect detection models for road surface often suffer from large parameter sizes and low computational efficiency, posing challenges for deployment on embedded or resource-constrained platforms. This paper developed an embedded lightweight detection system for road surface inspection using an improved YOLOv8 model. To reduce the parameter number, a parameter-sharing detection head and a slimming pruning model are employed to decrease model complexity while preserving detection performance. In addition, a matching guided distillation method is applied to enhance the representation learning capability of the pruned model without additional parameters. Finally, an embedded road surface defect detection system is developed, with a user interface implemented using the PyQt5 framework. Field tests on municipal roads in Guangzhou, China, demonstrate that the proposed system achieves high precision in defect detection and classification.