基于半监督学习的路面裂缝检测  

Pavement crack detection based on semi-supervised learning

作　　者：郭文浩[1,2] 张德津 GUO Wenhao;ZHANG Dejin(Faculty of Geosciences and Engineering,Southwest Jiaotong University,Chengdu 611756,China;Key Laboratory of Urban Spatial Information Engineering,Shenzhen 518060,China;School of Architecture and Urban Planning,Shenzhen University,Shenzhen 518060,China)

机构地区：[1]西南交通大学地球科学与工程学院,成都611756 [2]广东省城市空间信息工程重点实验室,广东深圳518060 [3]深圳大学建筑与城市规划学院,广东深圳518060

出　　处：《交通科技与经济》2024年第5期52-58,共7页Technology & Economy in Areas of Communications

基　　金：国家重点研发计划项目(2019YFB2102703)。

摘　　要：针对裂缝自动检测任务中难以获取大量精确标注样本数据的问题,提出LGS-Net(Local Global Similarity-Network)模型。LGS-Net的核心在于利用裂缝图像区域的语义相似性,有效结合少量已标注数据和大量未标注图像数据,通过半监督学习实现裂缝自动检测。为全面评估LGS-Net的性能,实验在GAPs384和Crack500数据集上进行验证。结果表明,在标注资源有限的情况下,LGS-Net能够实现高精度的裂缝检测。通过对检测结果的可视化分析,证明LGS-Net具有在复杂环境下有效识别裂缝的能力。LGS-Net利用路面裂缝图像的语义相似性特征进行检测,能为路面裂缝检测的工程应用提供技术支持。To address the challenge of obtaining a large volume of precisely annotated sample data for automatic crack detection tasks,LGS-Net(Local Global Similarity Network)is proposed in the paper.The core of LGS-Net is to utilize the semantic similarity of the crack image region,effectively combine a small amount of labeled data and a large amount of unlabeled image data,and realize the automatic crack detection by semi-supervised learning through to realize automatic crack detection.To comprehensively evaluate the performance of LGS-Net,the experiments are validated on GAPs384 and Crack500 datasets.The results show that LGS-Net is able to achieve high-precision crack detection with limited labeling resources.The ability of LGS-Net to effectively identify cracks in complex environments is demonstrated through the visual analysis of the detection results.LGS-Net utilizes the semantic similarity features of pavement crack images for detection,which provides technical support for engineering applications of pavement crack detection.

关 键 词：道路工程 裂缝检测 语义相似性 半监督学习 对比学习 

分 类 号：U416.2[交通运输工程—道路与铁道工程]