Crack Detection on Road Surfaces Based on Improved YOLOv8

Abstract

Road defect detection is vital for road maintenance but remains challenging due to the complexity of backgrounds, low resolution, and crack similarity. This paper introduces YOLOv8-VOS(VOS means ‘vanillaNet+ODConv+SEAttention’), an enhanced road crack detection algorithm that incorporates an improved Vanilla Net backbone with Squeeze-and-Excitation (SE) attention and ODConv modules. The loss function is replaced with WIoU to better balance bounding box regression. Experiments on the RDD2022 dataset demonstrate a 2% improvement in average accuracy over the original YOLOv8, achieving 53.7%. The proposed model effectively identifies road cracks in complex traffic backgrounds, contributing to safer and more efficient road maintenance.