期刊文献+
任意字段
题名或关键词
题名
关键词
摘要
作者
第一作者
机构
刊名
分类号
参考文献
作者简介
基金资助
栏目信息

基于Transformer与CNN架构的路面裂缝语义分割 认领

Semantic Segmentation of Pavement Cracks Based on Transformer and CNN Architecture
暂未订购 下载PDF
引用 收藏 分享
摘要 为解决传统路面裂缝分割精度差、完整度不高的问题,提出一种基于Transformer与CNN架构的路面裂缝语义分割模型TCU-Net.该模型以Transformer架构的FastViT作为主干提取全局语义表征,引入EVC ...展开更多 To address the issues of low accuracy and poor integrity in traditional pavement crack segmentation,this study proposes a semantic segmentation model for pavement cracks—TCU-Net—based on a hybrid Transformer and CNN architecture.This model uses FastViT,based on the Tran...MORE To address the issues of low accuracy and poor integrity in traditional pavement crack segmentation,this study proposes a semantic segmentation model for pavement cracks—TCU-Net—based on a hybrid Transformer and CNN architecture.This model uses FastViT,based on the Transformer architecture,as the backbone to extract global semantic representations,introduces the EVC Block to enhance local texture and boundary responses,and designs a CAFusion module with integrated coordinate attention to achieve efficient collaboration between global and local features.During the training phase,Focal Dice Loss is used as the loss function to mitigate the impact of sample imbalance.Comparative experiments with models such as UNet,SegNet,DcsNet,and DTrc-Net,as well as ablation studies,validate the effectiveness of the proposed method.Results demonstrate that TCU-Net achieves state-of-the-art performance on two pavement crack datasets:Pavementscapes and Crack500.Specifically,its mIoU reaches 77.62% and 73.28%,mPA reaches 86.2% and 82.73%,and F1-score reaches 85.70% and 82.19%,respectively—outperforming existing mainstream semantic segmentation models.This reflects its strong expressive capability and generalization performance in crack detection tasks.As the backbone network,FastViT significantly enhances the model's performance in complex image segmentation,with substantial improvements in all metrics.The EVC Block strengthens the model's perception of local details,further boosting segmentation accuracy.Compared with traditional fusion methods,CAFusion adaptively assigns weights to each channel,fully integrates low-dimensional local information and high-dimensional global information,and notably improves the model's ability to represent edge details and structural continuity.TCU-Net,based on the Transformer-CNN hybrid architecture,enables complete and accurate pavement crack segmentation,providing a precise decision-making basis for road maintenance.FEWER
作者 赵全满 马志浩 葛鲁民 刘朝晖 朱琳琳 刘继法 郭桂宏 ZHAO Quanman;MA Zhihao;GE Lumin;LIU Zhaohui;ZHU Linlin;LIU Jifa;GUO Guihong(School of Traffic Engineering,Shandong Jianzhu University,Jinan 250101,China;Shandong High-speed Engineering Inspection Co.,Ltd.,Jinan 250002,China;Shandong Taishan Traffic Planning and Design Consulting Co.,Ltd.,Tai’an 271000,Shandong,China)
机构地区 山东建筑大学交通工程学院 山东高速工程检测有限公司 山东泰山交通规划设计咨询有限公司
出处 《昆明理工大学学报（自然科学版）》 CAS 北大核心 2026年第2期144-156,共13页 Journal of Kunming University of Science and Technology(Natural Science)
基金 山东省交通运输科技计划项目(2024B39,2023B32).
关键词 路面工程 深度学习 语义分割 Transformer CNN pavement engineering deep learning semantic segmentation Transformer CNN
分类号 U418.6 [交通运输工程—道路与铁道工程] TP183 [自动化与计算机技术—控制理论与控制工程] TP391.41 [自动化与计算机技术—控制科学与工程]
  • 相关文献

参考文献15

二级参考文献364

共引文献700

同被引文献1

昆明理工大学学报（自然科学版）
2026年 第2期
科创助手
论文智评
维普砚承
维普斟知

相关作者

内容加载中请稍等...

相关机构

内容加载中请稍等...

相关主题

内容加载中请稍等...

浏览历史

内容加载中请稍等...
;
在线咨询 使用帮助 返回顶部 意见反馈