Abstract: [Purposes]Addressing the insufficient detection accuracy caused by the “low contrast,strong texture,and cross-scale” characteristics of defects in real-world scenarios,this paper proposes a lightweight road defect detection method combining frequency domain decomposition and contextual multi-scale convolutional attention.[Methods]This method introduces Haar wavelet downsampling into the YOLO11 framework to suppress aliasing and preserve high-frequency details such as cracks in the frequency domain.Simultaneously,it constructs multi-scale convolutional attention to stably respond to slender and small-scale targets.Furthermore,a context-guided module fuses local details and void context to suppress complex background interference and improve target boundary continuity,thereby enhancing detection robustness without significantly increasing computational overhead.[Findings]Training and testing were conducted using the SVRDD street view road damage dataset.The results show that,under lightweight conditions with approximately 3.26×10⁶ parameters and approximately 5.6 ms per inference layer,the model achieves mAP₅₀ and F1 scores of 0.756 and 0.752 7,respectively,which are significant improvements compared to the baseline model.[Conclusions]Furthermore,it achieves a better balance between accuracy and efficiency when compared with typical object detection methods such as Faster R-CNN and YOLOv5,thus verifying the effectiveness of the proposed method.

Key words: road defect identification, convolutional neural networks, wavelet downsampling, multi-scale convolutional attention, context guidance