基于神经网络与边缘修正的隧道衬砌裂缝识别

Crack Identification of Tunnel Lining Based on Neural Network and Edge Correction

针对隧道衬砌图像背景复杂的问题,提出基于卷积神经网络的图像快速筛查与裂缝区域定位,并针对裂缝多存在网状形态以及毛刺形分叉区域的特点,提出基于二叉树与曲线拟合的隧道衬砌裂缝识别方法。首先通过网格化区域划分与卷积神经网络进行区域类别判断,实现海量图片快速筛查与裂缝区域大致定位。对检出的裂缝区域通过Zhang-Suen细化算法获得裂缝原始骨架。根据裂缝点间的关联信息构建裂缝骨架二叉树,采用最大距离化的节点筛查规则修正裂缝骨架,并根据修正结果进行裂缝边缘修正,保留裂缝网状形态的同时,实现裂缝短小分叉和结节等不合理区域的剔除。最后基于最小二乘法曲线拟合修正裂缝轨迹,结合二叉树节点信息对裂缝轨迹断裂部分进行连接从而获取完整轨迹。试验结果表明,提出的方法可准确筛除98.73%不包含裂缝的衬砌图像;采用基于二叉树的边缘修正提高了裂缝的分割精度,并通过轨迹修正确保了裂缝轨迹的完整性。

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.