摘要
针对混凝土表面图像裂纹强度不均匀,噪声干扰严重,人工标注成本高等问题,提出了一个两阶段的无监督裂纹分割方法,该方法充分利用丰富的无裂纹图像,构建一个由粗到细的裂纹分割框架.首先,该方法仅使用无裂纹的背景图像块训练一个裂纹图像块检测器,过滤大部分的背景区域,以便粗略地定位裂纹;其次,提出一种多尺度融合的混合阈值图像块分割方法,对裂纹图像块进行像素级的精细分割.多尺度融合策略有利于裂纹保持连续性以及抑制噪声.在经典的数据集CFD上的实验结果表明:对比无监督裂纹分割方法,本文的两阶段分割方法取得了更好的性能,比最好的分割方法 FFA,F1-score值提高了2.38%.
A two-stage unsupervised crack detection method is proposed for the problems of uneven crack strength,severe noise interference and high cost of manual labelling in concrete surface images.This method makes full use of the rich crack-free images to construct a coarse-to-fine crack detection framework.Firstly,a crack image block detector is trained using only crack-free samples,aiming to filter most of the background regions so as to coarsely locate the crack.Secondly,a multi-scale fusion method based on hybrid threshold image patch segmentation is proposed to generate the finer segmentation of the crack image block at the pixel level.The multiscale fusion strategy is beneficial for crack continuity maintenance as well as noise suppression.Experimental results on the classical Crack Forest Dataset(CFD) crack dataset show that the proposed two-stage method achieves better performance than the classical unsupervised crack segmentation method,with a 2.38% improvement in F1-score value over the state-of-the-arts of crack segmentation.
作者
梁凤娇
李清勇
柴雪松
王雯
LIANG Fengjiao;LI Qingyong;CHAI Xuesong;WANG Wen(Beijing Key Laboratory of Traffic Data Analysis and Mining,Beijing Jiaotong University,Beijing 100044,China)
出处
《北京交通大学学报》
CAS
CSCD
北大核心
2023年第2期122-128,共7页
JOURNAL OF BEIJING JIAOTONG UNIVERSITY
基金
中国国家铁路集团有限公司科技研究开发计划(P2020T001)
国家自然科学基金(U2034211)。
关键词
裂纹分割
无监督学习
多尺度融合
crack segmentation
unsupervised learning
multi-scale fusion
