基于融合编码器与双解码器的半监督疲劳裂纹分割  

Semi-supervised fatigue crack segmentation based on fusion encoder and dual decoders

作　　者：香超 邓露[1,2] 王维[1,2] 郭晶晶[1,2] Xiang Chao;Deng Lu;Wang Wei;Guo Jingjing(College of Civil Engineering,Hunan University,Changsha 410082,China;Hunan Provincial Key Laboratory for Damage Diagnosis of Engineering Structures,Hunan University,Changsha 410082,China)

机构地区：[1]湖南大学土木工程学院,长沙410082 [2]湖南大学工程结构损伤诊断湖南省重点实验室,长沙410082

出　　处：《东南大学学报（自然科学版）》2024年第1期89-98,共10页Journal of Southeast University：Natural Science Edition

基　　金：国家自然科学基金资助项目(52278177,52308312);湖南省自然科学基金资助项目(2023JJ30154);长沙市自然科学基金资助项目(kq2208029)。

摘　　要：针对现有基于深度学习的疲劳裂纹分割算法依赖于大量像素级标记的问题,提出了一种半监督疲劳裂纹分割网络SFD-Net.SFD-Net利用对比学习方法进行半监督训练,以减少对大量像素级标记的依赖.同时,它采用融合编码器和双解码器的设计,旨在更好地捕捉裂纹区域的特征并提高分割准确性.通过引入改进注意力模块和边界优化模块,增强了对裂纹特征的表示,提高了裂纹边界的分割质量.在公开的疲劳裂纹数据集上对SFD-Net进行性能验证.结果表明:相较于使用相同标记比例的全监督算法,SFD-Net的分割性能有明显提升;仅使用25%的标记数据时,SFD-Net的交并比(IoU)达到70.6%,超过了使用100%标记数据的其他全监督算法的平均IoU(69.1%);同时,与其他先进的半监督方法相比,SFD-Net在所有标记数据比例下均取得了最高的预测精度.In response to the heavy reliance on extensive pixel-level annotations in existing deep-learning-based fatigue crack segmentation algorithms,a semi-supervised fatigue crack segmentation network named SFD-Net(semi-supervised fusion encoder and dual decoder network)is introduced.SFD-Net uses contrastive learning for semi-supervised training,reducing the reliance on extensive pixel-level annotations.Additionally,it incorporates a design with fusion encoder and dual decoders to better capture features within crack regions,improving segmentation accuracy.By integrating improved attention modules and boundary optimization modules,the representation of crack features is enhanced,leading to a significant improvement in the segmentation quality of the crack boundary.The performance of SFD-Net is validated on a publicly available fatigue crack dataset.The results indicate that the segmentation performance of SFD-Net is significantly improved compared with the fully supervised algorithms with the same annotation proportions.Even with only 25%of the labeled data,SFD-Net achieves an intersection over union(IoU)of 70.6%,surpassing the average IoU(69.1%)of other fully supervised algorithms using 100%labeled data.Moreover,when compared with other advanced semi-supervised methods,SFD-Net consistently achieves the highest predictive accuracy across all levels of labeled data.

关 键 词：疲劳裂纹检测 语义分割 半监督学习 对比学习 复杂背景 

分 类 号：U448.36[建筑科学—桥梁与隧道工程]