| 陈栋,康飞,李新宇,张亚平.基于注意力引导重构网络的大坝无监督裂缝分割方法[J].水利学报,2025,56(10):1360-1371 |
| 基于注意力引导重构网络的大坝无监督裂缝分割方法 |
| An unsupervised crack segmentation method based on attention-guided reconstruction network |
| 投稿时间：2024-11-30 修订日期：2025-09-24 |
| DOI：10.13243/j.cnki.slxb.20240773 |
| 中文关键词: 大坝表面裂缝 无监督裂缝分割 重构网络 注意力引导机制 缺陷检测 |
| 英文关键词: dam surface crack unsupervised crack segmentation reconstruction network attention-guided mechanism defect detection |
| 基金项目:国家重点研发计划项目（2022YFB4703404） |
|
| 摘要点击次数: 28 |
| 全文下载次数: 26 |
| 中文摘要: |
| 针对坝体表面裂缝缺陷检测中样本稀缺、标注困难以及背景噪声干扰多等问题，提出了一种基于注意力引导重构网络的无监督裂缝分割方法AGR-CrackNet。该网络在训练阶段，通过随机生成符合裂缝形貌特征的掩膜模拟完好混凝土图像的损坏过程，以构造训练样本，并将其输入到包含注意力引导机制的重构网络进行训练。在测试阶段，利用预训练的重构模型对裂缝图像进行重构，通过计算重构图像与裂缝图像的绝对差值，并结合Otsu算法，实现了裂缝区域的精确分割。消融实验结果表明，新提出的掩膜损坏过程及训练机制显著提升了裂缝的分割精度，同时有效抑制了噪声区域的错误分割。在实际坝面采集的裂缝图像以及水下裂缝图像数据集中，AGRCrackNet的分割综合评价指标F1较其他无监督方法分别提升了10.0%~66.8%和26.4%~31.2%，并且优于大部分有监督算法。此外，讨论分析表明，AGR-CrackNet生成的高质量标签数据可用于有监督网络训练，拓展了该方法的应用场景。 |
| 英文摘要: |
| To address issues such as sample scarcity，annotation difficulty，and background noise interference in dam surface crack detection，an unsupervised crack segmentation method based on an attention-guided reconstruction network，named AGR-CrackNet，is proposed. During the training phase，masks conforming to crack morphology characteristics are randomly generated to create damaged process of intact concrete images，thereby creating training samples for the AGR-CrackNet. During testing，the pre-trained model reconstructs crack images，and precise segmentation is achieved by calculating the absolute differences between the reconstructed and original images，combined with the Otsu algorithm. Ablation experiments demonstrate that the proposed mask generation process and training mechanism significantly improve segmentation accuracy while effectively suppressing noise-induced errors. AGRCrackNet improves F1 scores by ranges of 10.0%-66.8% on dam surface and 26.4%-31.2% on underwater datasets， compared to other unsupervised methods and surpassing most supervised algorithms. Furthermore，analysis indicates that the high-quality labeled data generated by AGR-CrackNet can be utilized to train supervised networks，further expanding its application scenarios. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |