請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/91971
標題: | 結合深度學習模型與區塊影像技術應用於混凝土表面裂縫辨識 Automatic Recognition Method for Concrete Surface Cracks: Integrating Image Patch Technique and Deep Learning Models |
作者: | 林禹齊 Yu-Qi Lin |
指導教授: | 張家銘 Chia-Ming Chang |
關鍵字: | 結構健康檢測,非破壞式檢測,電腦輔助,自動化裂縫判讀,深度學習,直方圖均衡化,資料增強,圖像分類, structural health monitoring,non-destructive testing,computer-aided,automatic crack detection,deep learning,histogram equalization,data augmentation,image classification, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 對建築結構而言,結構健康檢測是至關重要,其可於災後觀察結構物的受損情形以評估補強方式,或是對老舊建物進行壽命評估,恰當的結構物健康檢測及補強可以延續建築物生命週期甚至避免生命及財產的損失。目前結構健康檢測分為破壞式及非破壞式檢測,為了進行大量的建物健康檢測,例如災後建物受損程度的評估,或是定期性的對老屋進行評估,通常以非破壞式檢測作為主要健康檢測的方式。非破壞性檢測主要觀察建物的外觀受損情形,以評估建築物的受損程度,其中以觀察裂縫之型態作為重要判斷依據,包括裂縫之長度、寬度及走向。過去之檢測方式需要技師至現場實地勘驗,以目視觀察並於劣化評估表上標註破壞,而人力之觀察及紀錄流程通常是耗費時間及精力且不精確的,因此近年來開始以電腦輔助裂縫之判別,期盼以自動化之判別流程節省時間及精力,並利用自動化判讀之優點使判斷標準統一。本研究提出流程化之自動化裂縫判讀流程,使用深度學習之方式作為主要裂縫識別之方式,改善以往以深度學習裂縫辨識之缺點,同時可以達到裂縫訓練資料標註過程簡化,並提高訓練後之模型於辨識時之精度,提出新的辨識判別流程,以計算裂縫可能出現之機率,再進一步利用計算出之裂縫出現機率建立裂縫之走向。
本研究提出之裂縫辨識方法主要分為三個部分,第一部分為深度學習模型之訓練,本研究主要以VGG19及Resnet50以圖像分類為任務的深度學習模型,作為辨識區塊圖像之裂縫之深度學習模型。為了比較不同訓練模型之精度性,訓練及預測時將原始資料集進行灰階圖片之轉換,及配合限制對比度自適應直方圖均衡化(CLAHE)以校正圖片之色階分布,並於訓練過程中使用資料增強(Data Augmentation)的方式,提高訓練照片之多元性。第二部分則為提出之裂縫辨識流程,此方法將原始裂縫圖片切割成區塊圖片,同時調整區塊圖片之重疊率,再將區塊圖像放入深度學習模型中預測,以圖像分類之方式達到像素等級的裂縫辨識結果,並進一步利用裂縫辨識結果之機率分布,計算裂縫之粗略位置與走向。第三部分則為裂縫辨識流程之驗證,使用於某停車場所拍攝之裂縫圖像,以驗證模型之可行性。因為自行拍攝之圖像與原始訓練之裂縫照片資料集於拍攝條件、裂縫型態具有一定的差異性,因此可同時驗證本研究提出之裂縫辨識方法之精度。經驗證過程中發現,以Resnet50搭配原始裂縫訓練資料集擁有最好的辨識精度,而以相鄰區塊圖片之步長為14計算出的裂縫機率分布圖可以有效的顯現出裂縫的位置,最後以Zhang-Suen骨架化演算法計算出的裂縫位置與原始圖片之裂縫標註相比,發現其結果可以有效的預測裂縫位置與趨勢,僅有少部分裂縫區域未被辨識出。 For architectural structures, structural health monitoring is crucial. It can be used to observe the damage to structures after a disaster in order to assess the reinforcement methods. It is also used for evaluating the lifespan of old buildings. Proper structural health monitoring and reinforcement can extend the lifecycle of a building and even prevent loss of life and property. Currently, structural health monitoring can be categorized as destructive and non-destructive testing. Non-destructive testing is usually the preferred method for conducting extensive building health assessments, such as evaluating the extent of damage after a disaster or conducting regular assessments of old houses. Non-destructive testing primarily examines the appearance of the building to assess the level of damage, with a focus on observing crack patterns. This includes measuring the length, width, and direction of cracks. In the past, inspection methods required technicians to physically visit the site, visually observe and mark the damage on a deterioration assessment form. However, the manual observation and recording process was time-consuming, labor-intensive, and not precise. Therefore, in recent years, there has been a shift towards computer-assisted crack identification to save time and effort. The advantages of automated interpretation are utilized to establish unified criteria for crack assessment. This study proposes a process-oriented automated crack identification approach that utilizes deep learning as the primary method for crack recognition. It aims to improve upon the limitations of previous deep learning crack recognition methods. The proposed approach simplifies the crack labeling process during training and enhances the generalizability of the trained model to various types of cracks. It introduces a new identification process that calculates the probability of crack occurrence and utilizes this probability to determine the crack direction. The crack recognition method proposed in this study consists of three main parts. The first part involves training deep learning models. VGG19 and ResNet50, commonly used deep learning models for image classification, are employed as the crack recognition models for block images. To compare the generalizability of different training models, the original dataset is transformed into grayscale images during training and prediction. Additionally, contrast-limited adaptive histogram equalization (CLAHE) is applied to correct the grayscale distribution of images. Data augmentation techniques are also employed during the training process to increase the diversity of training photos. The second part introduces the crack recognition process. This method involves segmenting the original crack images into block images and adjusting the overlap rate of these blocks. The block images are then fed into the deep learning models for prediction, enabling pixel-level crack recognition results through image classification. Furthermore, the probability distribution of the crack recognition results is utilized to calculate the approximate position and direction of the cracks. The third part is the validation of the crack recognition process, using crack images captured in a parking lot to assess the feasibility of the proposed model. As there may be differences in shooting conditions and crack patterns between the self-captured images and the original training crack photo dataset, the generalizability of the crack recognition method proposed in this study can be evaluated simultaneously. During the validation process, it was found that using ResNet50 with the original training crack dataset achieved the best recognition accuracy. The crack probability distribution calculated using a stride of 14 for adjacent block images effectively displayed the crack positions. Finally, the crack positions and trends predicted by the Zhang-Suen skeletonization algorithm were compared to the crack annotations in the original images. The results showed that the proposed method effectively predicted the crack positions and trends, with only a small portion of the crack areas remaining undetected. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/91971 |
DOI: | 10.6342/NTU202304163 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 土木工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-112-1.pdf | 6.56 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。