首页|复杂背景下基于改进Mask R-CNN的路面裂缝检测算法

复杂背景下基于改进Mask R-CNN的路面裂缝检测算法

扫码查看
原文链接
  • 万方数据
  • 维普
裂缝检测对路面养护具有重要意义,深度学习在该领域取得一定成效.然而,在实际应用中,图像中的噪声纹理背景、复杂的裂缝拓扑结构和图像采集设备给裂缝检测带来了一定的挑战.为了提升在复杂场景下的路面裂缝检测精度,提出了一种改进掩码区域卷积神经网络(Mask R-CNN)模型的实例分割算法.使用ConvNeXt-T替代Mask R-CNN的ResNet50框架作为特征生成网络,在自下而上捕获长期依赖的同时保持裂缝特征多样性;设计高维特征提取模块(HFEM)获取高级语义信息,消除背景噪声;引入感受野模块(RFB),扩大感受野,增强多尺度特征信息交互能力.在多结构裂缝图像(MSCI)数据集上进行对比实验,结果表明,提出的改进方法能显著提升Mask R-CNN模型的分割精度,优于经典的Cascade Mask R-CNN,最佳模型F1得分84.15％,相较原算法提高了6.29％.在DeepCrack数据集上进行泛化性实验,表现优异.
Road crack detection algorithm based on improved Mask R-CNN in complex backgrounds
Crack detection is of great significance to road maintenance,and deep learning has made some achievements in this field. However,in practical applications,the noise texture background,complex crack topology,and image acquisition equipment bring some challenges to crack detection. In order to enhance the accuracy of road crack detection in complex backgrounds,this paper proposed an instance segmentation algorithm based on the improved mask region-based convolutional neural network (Mask R-CNN). ResNet50 of Mask R-CNN was replaced by ConvNeXt-T as the feature generation network,which captures long-term dependence from bottom to top while maintaining crack feature diversity. A high-dimension feature extraction module (HFEM) was designed to extract high-level semantic information,effectively eliminating background noises. Additionally,the paper introduced the receptive field block (RFB) to expand the receptive field and enhance multi-scale feature interaction capabilities and conducted comparative experiments on the multi-structure crack image (MSCI) dataset. The results demonstrate that the improved algorithm proposed in this paper can significantly improve the segmentation accuracy of the Mask R-CNN model and outperforms classical Cascade Mask R-CNN. The F1 score of the best model is 84.15％,which is 6.29％ higher than that of the original algorithm. Moreover,it shows excellent performance in the generalization experiments on the DeepCrack dataset.

road crack detectioncomplex scenariosmask region-based convolutional neural network (Mask R-CNN)instance segmentation

张晓华、李小龙、艾金泉、舒兆翰

展开 >

东华理工大学测绘与空间信息工程学院,江西南昌,330013

东华理工大学自然资源部环鄱阳湖区域矿山环境监测与治理重点实验室,江西南昌,330013

核工业湖州勘测规划设计研究院股份有限公司,浙江湖州,313000

路面裂缝检测 复杂场景 掩码区域卷积神经网络(Mask R-CNN) 实例分割

国家自然科学基金湖南省自然资源厅科技项目江西省重点研发计划江西省地质局科技研究项目

422610782022-2620223BBE510302022JXDZKJKY08

2024

10.19580/j.cnki.1007-3000.2024.03.027

北京测绘
北京市测绘设计研究院,北京测绘学会

北京测绘

影响因子：0.55
ISSN：1007-3000
年,卷(期)：2024.38(3)