A Real-Time Pavement Crack Detection Method Integrating Lightweight YOLO Model and Multi-Scale Attention Mechanism

To improve the detection accuracy and computational efficiency of pavement crack identification, this study proposes an improved YOLOv8 framework. The model combines the lightweight Ghost module and the multi-scale attention mechanism, aiming to significantly reduce the model parameters and inference cost while maintaining high accuracy. The Ghost convolution structure is embedded in the YOLOv8 backbone network, effectively reducing computational complexity and improving the running speed. To further enhance the feature extraction capability, this paper introduces the cross-scale feature fusion module (CCFM) and the cascade group attention mechanism (CGA), which significantly improves the model’s ability to identify subtle cracks. The study constructs an image dataset covering multiple types of pavement surface diseases for model training and verification. Experimental results show that the improved model outperforms the existing YOLOv5, YOLOv7-Tiny and original YOLOv8 models in key indicators such as mAP, precision and recall, achieving a detection accuracy of 90.3% and an inference speed of 43 frames per second (FPS). The model shows good generalization ability and engineering application prospects and can provide effective technical support for intelligent pavement detection and rapid distress assessment.