結合影像處理、電腦視覺與人工智慧之混凝土結構表面裂縫識別研發

Integration of image processing, computer vision, and artificial intelligence to identify concrete surface cracks

謝承穎(Cheng-Ying Hsieh)

指導教授 ： 張家銘

國立臺灣大學/工學院/土木工程學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201902149

摘要

近年為了延長結構物之壽命與避免受到天然災害之二次損傷，結構健康監測於實務上日益重要，透過即時與自動化之方式將結構物之健康資訊傳遞給居民，亦或是工程師以利後續之補強工程。而傳統之結構健康監測方式為利用有線感測器傳遞電子訊號至主機，再由主機從感測器所提供之資訊進行結構健康監測。而本研究目的為提出一自動化方式，有別於一般有線感測器，以無線之影像量測、電腦視覺與深度學習判斷混凝土構件表面之裂縫性質，以非破壞性檢測裂縫之方式，提供裂縫相關資訊使非專業人士也能夠判斷混凝土構件之健康情況。

本研究首先利用基於人工智慧中深度學習以及遷移學習的方式，藉由學習訓練資料中每張影像特徵，訓練出屬於本研究之裂縫辨識模型。訓練完畢後，透過其架構能夠自動卷積每張影像提取特徵，進而自動化地判斷影像中混凝土表面裂縫之有無與將其位置框選出，隨後將深度學習所判斷裂縫之位置，進行影像處理，經過電腦視覺之方法，將混凝土表面上的裂縫萃取出，最後利用影像量測之方式，計算混凝土表面之裂縫長度、寬度。

本研究利用現場拍攝不同混凝土構件表面之裂縫影像，來評估方法的可靠性。利用事先校正好之雙相機模型於現場拍攝裂縫影像，將其輸入至本研究方法中，經由量測裂縫實際長度與寬度作為驗證，本研究方法所計算結果與實際值相當，顯示本研究方法能夠成功地判斷裂縫性質。而該方法能藉由非破壞性的檢測方式，進一步了解混凝土構件的破壞程度，可作為結構健康監測方法之一。

關鍵字

結構健康監測 ； 人工智慧 ； 深度學習 ； 電腦視覺 ； 影像處理 ； 影像量測

並列摘要

Structural health monitoring becomes more and more important in practice because this technology can elongate the structural life cycle as well as protect structures against natural hazards. Moreover, structural health monitoring systems can automatically inform residents and users for the current condition of structures and engineers for the current performance. In past, structural health monitoring relies on the contact sensors to acquire structural responses and then diagnoses structures in accordance with the measurements. In this research, a new method is developed to detect and quantify the concrete cracks based on the noncontact image measurements. This method integrates computer vision and deep learning to identify the crack existence and geometry. The identified cracks can provide indirect information for experts to further investigate the structural conditions.

This study exploits deep learning and transfer learning, e.g., the tools in the category of artificial intelligence, to train and establish a concrete segmentation model that can identify the locations of cracks in images. In this model, the crack features can be obtained from the convolutional neural network and then automatically identify whether the cracks are present and where the cracks are. Then, the image processing and computer vision are implemented to highlight and extract these cracks from images. Finally, the geometry of these cracks (i.e., lengths and widths) can be calculated by image measurement techniques.

To verify the proposed method, this study employs the images of concrete surface cracks obtained from the real-world structures and then evaluate the reliability of this method. In the verification, the pre-calibrated stereo camera model with a two-camera setup is used to verify the actual lengths and widths of cracks. The calculated results are compared with the actual measurements. As a result, the proposed method can successfully determine crack geometry. Moreover, the method also benefits users to obtain crack information and to turn into performance evaluation of concrete structures for structural health monitoring.

並列關鍵字

Structural Health Monitoring ； Artificial Intelligence ； Deep Learning ； Computer Vision ； Image Processing ； Image Measurement

參考文獻

[1] Abdel-Qader, I., Abudayyeh, O., Kelly, M.E. (2003), “Analysis of edge-detection techniques for crack identification in bridges,” Journal of Computing Civil Engineering , 17(4), 255-63.

[2] Kim, H., Ahn, E., Cho, S., Shin, M., Sim S.H. (2017), “Comparative analysis of image binarization methods for crack identification in concrete structures,” Cement Concrete Research, 99, 53-61.

[3] Yamaguchi, T., Hashimoto, S. (2006), “ Automated crack detection for concrete surface image using percolation model and edge information,” IECON 2006 - 32nd Annual Conference on IEEE Industrial Electronics, 3355-3360.

[4] Yamaguchi, T., Nakamura, S., Hashimoto, S. (2008), “An efficient crack detection method using percolation-based image processing,” 2008 3rd IEEE Conference on Industrial Electronics and Applications, 1875-1880.

[5] Yamaguchi, T., Hashimoto, S. (2010), “Fast crack detection method for large-size concrete surface images using percolation-based image processing,” Machine Vision and Application, 21(5), 797-809.