Abstract To address the problem of complex imaging background of tunnel lining, this paper proposes a fast image screening and crack region localization based on convolutional neural network, as well as a tunnel lining crack identification method based on binary tree and curve fitting, in view of the feature that cracks mostly exist in reticular form and burr-shaped bifurcation areas. First, the rapid screening of massive amounts of pictures and locating of approximate position of crack areas are conducted by gridding the area segmentation and using convolutional neural networks to judge the category of the areas. Then the original crack skeleton is obtained by the Zhang-Suen refinement algorithm for the detected crack areas. The skeleton binary tree of the cracks is constructed based on the correlation information between crack points and the crack skeleton is corrected by using the node screening rule of maximum distance. Based on the correction results, the crack edges are further corrected to retain the crack mesh pattern while realizing the elimination of unreasonable areas such as areas with short bifurcations and joints Finally, the crack trajectory is corrected based on least-square curve fitting and the broken parts of the crack trajectory are connected based on the binary tree node information to obtain the complete trajectory. The test results show that the proposed method can accurately screen out 98.73% of the lining images that do not contain cracks; the use of binary tree-based edge correction has improved the segmentation accuracy of the cracks and ensured the integrity of the crack trajectory by trajectory correction.
Abstract：
To address the problem of complex imaging background of tunnel lining, this paper proposes a fast image screening and crack region localization based on convolutional neural network, as well as a tunnel lining crack identification method based on binary tree and curve fitting, in view of the feature that cracks mostly exist in reticular form and burr-shaped bifurcation areas. First, the rapid screening of massive amounts of pictures and locating of approximate position of crack areas are conducted by gridding the area segmentation and using convolutional neural networks to judge the category of the areas. Then the original crack skeleton is obtained by the Zhang-Suen refinement algorithm for the detected crack areas. The skeleton binary tree of the cracks is constructed based on the correlation information between crack points and the crack skeleton is corrected by using the node screening rule of maximum distance. Based on the correction results, the crack edges are further corrected to retain the crack mesh pattern while realizing the elimination of unreasonable areas such as areas with short bifurcations and joints Finally, the crack trajectory is corrected based on least-square curve fitting and the broken parts of the crack trajectory are connected based on the binary tree node information to obtain the complete trajectory. The test results show that the proposed method can accurately screen out 98.73% of the lining images that do not contain cracks; the use of binary tree-based edge correction has improved the segmentation accuracy of the cracks and ensured the integrity of the crack trajectory by trajectory correction.
CHEN Yingying1,
2 LIU Xingen1,
2 HUANG Yongliang3 etc
.Crack Identification of Tunnel Lining Based on Neural Network and Edge Correction[J] MODERN TUNNELLING TECHNOLOGY, 2022,V59(6): 24-34
2022, Vol. 59 