現存結構中裂縫分割與量化之通用流程整合

Abstract

建築物上的裂縫是結構健康監測的依據之一，裂縫的寬度在日本的鋼筋混凝土建築震後損傷評估指南多次出現做為損害評估的依據。現存的文獻針對裂縫辨識或裂縫分割已有許多的研究，然而多數的研究僅考量單一情況，且未提供完整的辨識前置作業方法，因此本研究整合了各種裂縫量化和拍攝技術，建立一個完整的量化流程。同時，我們考量了現有流程圖對於量化精確度及快速量化能力的不足，為此本研究發展出了一種附有單個編碼器及兩個解碼器的多重任務深度學習模型，該模型能同時夠有效地提取中心線並生成準確的分割結果，以達到快速量化的效果。我們在模型中引入了Lovasz Hinge Loss作為訓練中的損失函數，它在處理裂縫邊界時相對於傳統的損失函數有一些優勢。 在裂縫寬度的計算成果中表明了在所提供的測試數據集下，本研究基於U-Net架構所提出的多重任務模型相較傳統方法在誤差評估指標中能夠有1~2個百分點的領先。證實了本研究提出的方法除了在時間上的節省外也在精確度中獲得了提升。

此外，本研究探討了一些與裂縫量化相關的問題，如拍攝環境或尺度因子對於裂縫寬度計算造成的影響。我們提出了一個基於3D重建和平面擬合的方法，克服了影像變形的問題，並證實了所提出的投影矩陣對精度修正的有效性。

Cracks on RC buildings are considered to be one of the key parameters used for structural health monitoring. The width of cracks has been repeatedly used as a basis for damage assessment in the post-earthquake damage assessment guidelines for reinforced concrete buildings in Japan. Existing literature has focused on crack recognition or segmentation, but most studies have only considered specific scenarios and have not provided a complete pre-processing method for crack identification. Therefore, in this study, integration of various crack quantification and imaging techniques was done to establish a comprehensive workflow for quantification of cracks.

Furthermore, the limitations of existing flowchart were considered in terms of quantification accuracy and efficiency. To address this, a deep learning model was developed based on a single encoder and two decoders. The developed model can effectively extract the centerline and generate accurate segmentation results simultaneously, enabling fast quantification. The Lovasz Hinge Loss was used as the training loss function, which has advantages in handling crack boundaries compared to traditional loss functions. The results of crack width calculation demonstrate that our proposed multi-task model based on the U-Net architecture outperforms traditional methods by 1-2 percentage points in error evaluation metrics, indicating improved accuracy and time savings with our proposed approach.

In addition, issues related to crack quantification were explored, such as the impact of imaging environment or scale factors on crack width calculation. A method was proposed based on 3D reconstruction and plane fitting to overcome the challenges of image deformation, and the effectiveness of the proposed projection matrix was validated in improving accuracy.