4.7 Article

Automatic soil desiccation crack recognition using deep learning

期刊

GEOTECHNIQUE
卷 72, 期 4, 页码 337-349

出版社

ICE PUBLISHING
DOI: 10.1680/jgeot.20.P.091

关键词

clays; in situ testing; laboratory tests; neural networks

资金

  1. National Key Research and Development Program of China [2019YFC1509902]
  2. National Natural Science Foundation of China [41925012, 41902271, 41772280, 41572246]
  3. Natural Science Foundation of Jiangsu Province [BK20171228, BK20170394]
  4. Fundamental Research Funds for the Central Universities

向作者/读者索取更多资源

This study proposes a new automatic soil cracks recognition method based on a U-Net convolutional neural network (CNN) architecture for segmentation on soil desiccation crack images. The method achieves better results than traditional crack image processing techniques and accurately identifies and quantifies soil cracks.
Soil desiccation cracking is a common natural phenomenon. The existence of cracks can negatively impact both the mechanical and hydraulic properties of soil. Accurate acquisition of soil crack networks is not only the basis for obtaining the relevant geometrical parameters of crack networks, but also an important foundation and premise for further study about the formation mechanism of shrinkage and desiccation cracking. This study proposes a new automatic soil cracks recognition method based on a U-Net convolutional neural network (CNN) architecture for segmentation on soil desiccation crack images. The backbone of the U-Net encoder is selected as ResNet and a new loss function combining both binary cross-entropy (BCE) loss and dice loss is used during the training stage to fit an imbalance problem. Subsequently, the U-Net with an encoder based on ResNet and a decoder part is trained from end to end on a subset of 524 labelled crack images with 224 x 224 pixels for semantic segmentation. The U-Net architecture achieves 94 center dot 38%, 74 center dot 43% and 81 center dot 13% for precision, recall and dice scores on test sets, which are better than all results using the Otsu threshold method employed in the traditional crack image processing technique. Experimental results reveal that deep learning can achieve higher accuracy than the traditional method (binarisation by thresholding) in quantifying surface crack ratio, average crack width, total crack length and crack number. Moreover, deep learning can not only accurately identify cracks or spots by means of crack edge features, but also can accurately separate soil cracks and clod areas under a bad photographing condition (such as uneven illumination, a field environment or poor photographing angle). Overall, the proposed deep learning-based method presents a satisfactory performance in soil crack image recognition and quantification. It may also be applied to other materials with cracks.

作者

我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。

评论

主要评分

4.7
评分不足

次要评分

新颖性
-
重要性
-
科学严谨性
-
评价这篇论文

推荐

暂无数据
暂无数据