摘要
精准的路面病害检测是开展高效路面养护管理的必要前提.针对现有路面病害检测方法存在精度不足、易受噪声干扰等问题,提出一种基于二维灰度图像与三维深度图像特征融合的卷积神经网络路面病害检测方法.首先依托线结构光路面信息采集系统,同步获取灰度图像与深度图像数据,并完成数据预处理与标注;继而结合图像数据特性,设计2种基于Res2Net架构的网络模型——双通道模型与双编码器模型,并在模型中嵌入注意力机制模块以优化裂缝分割的类别不平衡问题;最后针对不同类型路面病害开展定量分析.实验结果表明,多模态图像(灰度+深度)融合模型可使检测精度显著提升,平均交并比(MIoU)较基准提升了5.48%,达到82.96%,为道路养护的工程应用提供了参考.
Accurate pavement distress detection was regarded as a necessary prerequisite for effective pavement maintenance management.To address the limitations of existed methods,such as insufficient accuracy and susceptibility to noise,it was proposed a convolutional neural network based approach for pavement distress detection by fusing features from two-dimensional grayscale images and three-dimensional depth images.First,a line-structured light pavement information acquisition system was employed to synchronously collect grayscale and depth image data,followed by data preprocessing and annotation.Subsequently,considering the characteristics of the image data,two network models based on the Res2Net architecture—dual-channel and dual-encoder models—were constructed,with an attention mechanism module integrated to mitigate the class imbalance issue in crack segmentation.Finally,quantitative analyses were conducted for different types of pavement distress.Experimental results demonstrated that the multimodal fusion model significantly improved detection accuracy,with the mean Intersection over Union(MIoU)increasing by 5.48%over the baseline to 82.96%,providing a valuable reference for engineering applications in roadway maintenance.
作者
邱欣
张霆锋
陶珏强
梁毅
QIU Xin;ZHANG Tingfeng;TAO Jueqiang;LIANG Yi(College of Engineering,Zhejiang Normal University,Jinhua 321004,China)
出处
《浙江师范大学学报(自然科学版)》
2026年第1期33-44,共12页
Journal of Zhejiang Normal University:Natural Sciences
基金
浙江省自然科学基金资助项目(LJHSQY26E080001)
浙江师范大学青年博士专项(科技)资助项目(ZZ344205020523011159)。
关键词
卷积神经网络
多模态
路面裂缝检测
图像分割
convolutional neural network
multimodal
pavement crack detection
image segmentation
