Research on high-precision recognition model for multi-scene asphalt pavement distresses based on deep learning

Zhang, Sheng; Bei, Zhenghao; Ling, Tonghua; Chen, Qianqian; Zhang, Liang · 2024 · Crossref

DOI: 10.21203/rs.3.rs-4412199/v1

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the challenge of accurately detecting asphalt pavement distresses in complex, multi-scene environments, a task critical for road maintenance and traffic safety. Traditional manual inspection is inefficient and subjective, while conventional image processing and machine learning methods struggle with complex backgrounds and lack precise localization capabilities. Although deep learning models like YOLO offer speed and accuracy, existing versions often suffer from high parameter counts that hinder deployment on mobile devices and may fail to capture fine-grained features in blurry or low-resolution images. To resolve these issues, the authors propose SMG-YOLOv8, an improved object detection model based on the YOLOv8s framework, designed to enhance detection precision while reducing computational complexity. The SMG-YOLOv8 model integrates three specific architectural modifications. First, it incorporates a Space-to-Depth Convolution (SPD Conv) module in the backbone to preserve fine-grained information during downsampling, addressing issues with blurry images and small targets. Second, it employs a Multi-Scale Convolutional Attention (MSCA) mechanism to adaptively focus on distress targets of varying scales and distinguish them from complex backgrounds. Third, it replaces the standard C2f structure with a lightweight G-GhostC2f structure, which reduces parameters and computational burden while maintaining feature extraction efficiency. The model was trained on the RDD 2022 dataset, comprising 6,084 augmented images of four distress types: longitudinal cracks, transverse cracks, mesh cracks, and potholes. Training utilized an RTX 4090 GPU with PyTorch, employing stochastic gradient descent, mosaic data augmentation, and cosine annealing learning rate decay over 200 epochs. Experimental results demonstrate that SMG-YOLOv8 significantly outperforms the baseline YOLOv8s model and other variants, including YOLOv5n, YOLOv5s, YOLOv6s, and YOLOv8n. The proposed model achieved a macro-averaged precision (Pmacro) of 81.1% and a mean average precision at 0.5 IoU (mAP@50) of 79.4%, representing increases of 8.2% and 12.5%, respectively, over the YOLOv8s baseline. The model exhibited strong generalization capabilities across diverse scenarios and effectively identified all four distress types. Furthermore, the structural optimizations reduced the number of parameters, making the model more suitable for deployment on resource-constrained devices such as drones and inspection vehicles. The significance of this work lies in providing a high-precision, lightweight solution for intelligent pavement distress detection. By improving detection accuracy in complex, real-world conditions while lowering computational requirements, SMG-YOLOv8 offers practical technical support for automated road maintenance systems. This advancement facilitates more efficient and reliable monitoring of asphalt pavement conditions, contributing to enhanced traffic safety and reduced economic losses associated with road deterioration.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• distraction detection algorithms