基于U-net神经网络算法和改进的细化算法的水坝混凝土裂缝测量

Visual Measurement of dam concrete cracks based on U-net and improved thinning algorithm

以水坝裂缝的视觉识别与计算为目标,融合U-net神经网络算法和改进图像细化算法,提出大坝裂缝识别和宽度计算的方法。首先,采用U-net处理1500张水坝裂缝图像样本,得到裂缝预分割模型;接着,采用改进的细化算法提取裂缝骨架分割模型;最后,结合预分割结果和骨架分割模型得到水坝裂缝的准确信息。结果表明,使用U-net神经网络算法作为裂缝检测的预处理算法,可显著提高算法的鲁棒性。改进的细化算法在应对细小裂缝时,骨架提取的命中率在80.17%以上。基于U-net神经网络算法和改进的细化算法分割裂缝的平均准确率、召回率和F1分数分别为92.75%、73.45%和81.86,视觉测量裂缝宽度的误差平均值为8.7%。与主流深度学习方法相比,本研究不依赖于大量训练样本且避免了设置过多的人工阈值,具有显著的实用性和稳定性。本文所述方法可以迁移应用到更多的类似场景中,为实现基础设施智能自动监测与预警提供数据和技术支撑。

Aiming to realize the visual automatic identification and measurement of the crack width and real-time health monitoring,the U-net neural network algorithm and the improved image refinement algorithm are integrated to automatically detect dam concrete cracks.The U-net algorithm is adopted for 1,500 images of dam concrete crack samples training,and the crack pre-segmentation model is constructed.Then the proposed improved-refinement algorithm is utilized for the crack skeleton extraction.The dam crack information is finally obtained by combining the pre-segmentation result with the segmentation.The results illustrate that U-net as a pre-improved thinning algorithm for crack detection has notably enhanced the robustness of the pre-segmentation algorithm.This pre-improved thinning algorithm has a hit-rate of more than 80.17%when dealing with small cracks.The proposed algorithm has an accuracy rate,recall rate,and F1 score of 92.75%,73.45%,and 81.86,respectively.The error deviation of visual measurement of crack width is 8.70%.Compared with traditional deep learning algorithms with a heavy reliance on large numbers of training samples,the proposed method has firm practicability and stability,and can be transferred to more similar scenarios.This research can provide the data and technical support for the realization of intelligent automatic monitoring and early warning of infrastructures.