摘要
不同区域道路服役环境的差异影响着裂缝的形成和发展,多年冻土区与普通地区的路面裂缝在空间分布、几何特征等方面显著不同,对裂缝识别方法的适应性与性能提出了更高要求。为了分析上述问题,本文构建了一种融合多特征尺度与注意力机制的快速区域卷积神经网络(Faster-R-CNN)裂缝识别模型。该模型在原始结构基础上,引入分离注意力网络提升特征提取能力,引入特征金字塔融合模块增强多特征尺度裂缝检测能力,同时采用Soft-NMS算法优化候选框筛选策略。结合多年冻土区路面图像与普通地区公开图像数据集进行训练与测试,评估模型性能,并提取多种裂缝几何特征参数开展区域差异性分析。结果表明,改进后的模型在普通地区和多年冻土区路面检测中,各项指标均有所提升,尤其是多年冻土区中的精确率(precision)、召回率(recall)与平均精确率(mAP)(IoU=0.5~0.95)分别提升12.59%、14.43%、13.69%,均优于普通地区的提升幅度,显示出较强的环境适应性与稳定性。进一步分析发现,多年冻土区裂缝在长度、宽度、密度与连通性等指标上分布离散,参数间呈显著正相关,裂缝网络结构更复杂,反映出冻融循环主导下的系统性变化特征。本研究成果可为寒区道路病害评价及裂缝发展预测提供技术支撑与理论依据。
Pavement cracks are one of the earliest manifestations of road structure degradation,particularly prevalent in regions subject to extreme climatic variability.In permafrost regions,frequent freezing-thawing cycles and drastic ground temperature fluctuations induce mechanical stresses in roads that differ from those in ordinary regions,leading to significantly distinct crack formation mechanisms.Understanding these differences is crucial for developing robust crack detection systems adapted to diverse environmental conditions and formulating reasonable road maintenance strategies.This study aims to establish an intelligent and high-precision framework for crack detection capable of accurately identifying crack locations and extracting their geometric structural features that reflect regional differences.To achieve this objective,this study established an improved deep learning detection model based on the Faster Region-Based Convolutional Neural Network(Faster-R-CNN),integrating multi-feature channels and multi-scale information.The model integrated three major innovations:(1)the introduction of a split-attention network to improve inter-channel feature discrimination and suppress background interference;(2)the integration of convolutional block attention module(CBAM)with ResNeXt50 in the feature pyramid network(FPN)to strengthen the representation of cracks at different scales;and(3)the use of the soft non-maximum suppression(Soft-NMS)algorithm to replace traditional NMS to retain more valid bounding boxes for slender or densely distributed cracks.This model was trained and validated on two independent datasets:UAV-based orthophoto data from the permafrost sections of the Qinghai-Xizang Highway that represented the permafrost regions,and urban pavement crack images from the widely used VOC-2007 dataset that represented the ordinary regions.All images were annotated manually,and training samples were expanded through data augmentation.The end-to-end optimization was performed using a joint classification and regression loss function.During evaluation,model performance was comprehensively measured using indicators including precision,recall,mean average precision(mAP)at both IoU=0.5 and IoU=0.5~0.95 thresholds.On the test set of the ordinary regions,the model achieved a precision of 92.20%,a recall of 89.25%,and a mAP(IoU=0.5~0.95)of 81.14%.On the permafrost dataset,although the overall accuracy was slightly lower due to complex backgrounds,the performance improvement over the baseline was more significant,demonstrating the model’s adaptability and stability in complex environments.Ablation experiments indicated that Soft-NMS increased mAP by approximately 5%and the integrated multi-module mechanism provided an 11%~14% improvement.Based on crack detection,this study developed a crack geometric parameter extraction module,capable of extracting indicators such as crack length,width,area ratio(Rc),length density(Lc),connectivity index(K),and curvature to quantitatively characterize the spatial structure of cracks.The results showed that cracks in permafrost regions exhibited greater scale variability and geometric expansion,with typical transverse cracks exceeding 2.5 m in length,widths often over 10 mm,and connectivity indices generally above 0.6.The statistical distributions of Lc and Rc in permafrost regions were right-skewed,indicating high crack density and coverage.Further Pearson correlation analysis showed stronger coupling among geometric parameters in permafrost regions.For example,the correlation coefficient between Lc and Rc exceeded 0.65,reflecting a synergistic expansion of crack networks likely driven by thermal and hydraulic forces.In contrast,cracks in ordinary regions were more isolated,with weaker coupling between parameters.The contributions of this study are mainly reflected in two aspects.First,a highly accurate,generalizable crack detection model is proposed,suitable for automated inspection tasks across different road surface environments.Second,a standardized workflow of crack geometric analysis is developed,which can be applied to evaluate crack morphology and reveal deterioration mechanisms under specific climatic and geological conditions.These findings can inform maintenance agencies in prioritizing repair tasks based on indicators such as crack length density and connectivity and provide crack control parameters for road structure design in permafrost regions,thereby promoting the construction of more resilient road infrastructure.Moreover,the geometric parameters can serve as input variables for predictive models,facilitating intelligent evaluation of road performance and long-term early warning systems.In conclusion,this study integrates computer vision technology with environmental pavement science,representing significant progress in both automated detection technology and the mechanistic understanding of pavement deterioration processes.
作者
姚常杰
柴明堂
郭子龙
张楷昕
刘微
YAO Changjie;CHAI Mingtang;GUO Zilong;ZHANG Kaixin;LIU Wei(School of Civil and Hydraulic Engineering,Ningxia University,Yinchuan 750021,China;Key Laboratory of Digital Water Management for Yellow River Water Network in Ningxia,Yinchuan 750021,China)
出处
《冰川冻土》
2025年第5期1490-1500,共11页
Journal of Glaciology and Geocryology
基金
国家冰川冻土沙漠科学数据中心2023年度开放基金项目(E01Z790201)
宁夏回族自治区重点研发计划项目(引才专项)(2023BSB03021)
宁夏回族自治区高等学校一流学科建设项目(NXYLXK2021A03)资助。
关键词
裂缝识别
几何特征
多年冻土
多特征尺度
注意力机制
crack detection
geometric features
permafrost
multi-feature scale
attention mechanism
