透過您的圖書館登入
IP:18.191.37.17
  • 學位論文

自動化影像辨識應用於鋼橋樑鏽蝕缺陷評估

Automatic Color Image Recognition for Steel Bridge Rust Defects Assessment

指導教授 : 張陸滿
共同指導教授 : 陳柏翰

摘要


影像處理技術已應用於北美地區的鋼橋樑表面塗漆檢測,然而實際使用時仍有光度不均、缺陷部位相較背景不明顯及雜物干擾等環境因素影響辨識結果,另外過去應用影像處理的鋼橋樑檢測研究中,尚無有效的模型可以辨別鏽蝕的嚴重程度,而且也需要較長的處理時間,無法做到即時檢測。 本研究首先應用傅立葉轉換(Fourier Transform)發展Fourier-transform-based steel bridge coating defect detection approach (FT-DEDA),利用鏽蝕像素間的顏色差異大於背景像素間的顏色差異這項特徵,來區分輸入影像是否包含缺陷部分,再結合彩色影像處理技術發展出第一個模型Rust defect recognition method (RUDERM),將輸入影像區分為無缺陷影像(non-defective)、有鏽蝕缺陷影像(rust-defective)及其他缺陷影像(other-defective)三種,如輸入影像被分類為無缺陷影像或其他缺陷影像,檢測系統就會停止,得以減少整體處理時間及避免誤判。其次本研究利用Support Vector Machine (SVM)發展第二個模型support-vector-machine-based rust assessment approach (SVMRA),區分有鏽蝕缺陷影像中的鏽蝕像素及背景像素,並計算鏽蝕部分占全部影像的比率。緊接著,本研究以root-mean-square standard deviation (RMSSTD)及Artificial Neural Networks (ANNs)發展第三個模型Artificial-Neural-Networks-based method for Rust Intensity recognition(ANNRI),將鏽蝕依顏色深淺區分適當的群組,藉以描述鋼橋樑鏽蝕的腐蝕程度。 為有效進行鋼橋樑表面塗漆品質檢測,上述三個模型將結合為Rust Defects Assessment Hybrid Model (RDAHM),並與過去類似研究進行比較,如SKMA、RUDA及BE-ANFIS。實驗結果顯示,RDAHM辨識鏽蝕的準確度高於其他三種方式,而且不受不均勻光源影響,亦可以辨識出紅色塗漆中的鏽蝕區域;然而RDAHM在處理時間方面卻需要2.05秒,雖然遠低於BE-ANFIS所需的18.56秒,卻高於RUDA的需0.26秒及SKMA的0.27秒,如果能夠再縮短處理時間,RDAHM將可以發展為即時的鋼橋樑表面塗漆品質自動檢測系統。

並列摘要


In North America, image processing has been applied on steel bridge coating inspection since the late 1990s. This newly proposed application, however, still has problems that haven’t been solved. Particular environment conditions like non-uniform illumination, low contrast, and noises often affect the assessment results. Also, the current methods cannot recognize rust intensity, and required long processing time that need to be shortened. Here we applied Fourier Transform to determine the existence of rust defects before assessing rust ratio in order to shorten the processing time and avoid recognition error. The Fourier-transform-based steel bridge coating defect detection approach (FT-DEDA) makes use of the fact that differences between background pixels are not as big as differences between defect pixels, to detect the existence of defects. To differentiate rust from other defects, a steel bridge coating rust defect recognition method (RUDERM), combining FT-DEDA and color features, was developed. RUDERM defines a rust color range and classifies a normal or non-normal steel bridge coating image into “non-defective,” “rust-defective,” and “other-defective.” After detecting rust defects in the images by RUDERA, a brand-new support-vector-machine-based rust assessment approach (SVMRA), which integrates color image processing, Fourier transform, and support vector machine (SVM), is applied to assess the rust ratio. The rust intensity was assessed by ANNRI, a combination of RMSSTD and ANNs, which takes advantage of the different similarity of RGB color within the groups. ANNRI is able to classify the rust images into suitable number of clusters, which is close to the classification by human naked eyes, for describing gradually changed colors. The Rust Defects Assessment Hybrid Model (RDAHM) proposed in this research includes three processors: RUDERM as the pre-processor, SVMRA as the main processor, and ANNRI as the post-processor. RUDERM detects the existence of rust defects in the steel bridge images. If the image is judged as rust-defective, the images would be preceded to SVMRA; otherwise, the process would be stopped. The pixels of rust defects are identify by SVMRA to assess the rust ratio. After assessing the rust ratio, ANNRI classifies the images into suitable number of groups to describe gradually changed colors. The rust recognition performance of RDAHM method was validated through extensive comparisons with previous rust defect assessment methods, including SKMA, RUDA, and BE-ANFIS methods, under various situations. According to the experimental results, it is found that RDAHM is superior to SKMA, RUDA, and BE-ANFIS in rust recognition. It is apparent that RDAHM has the best recognition result, followed by RUDA, BE-ANFIS, and SKMA. Additionally, RDAHM is more reliable than other methods based on smallest standard deviation. We showed that RDAHM overcomes the problem of non-uniform illumination that SMKA cannot handle, and is more stable than RUDA and BE-ANFIS in dealing with rust images with red-color-tone background. RDAHM avoids recognition errors better than SMKA and BE-ANFIS if the acquired image is homogeneous. In terms of processing time, RUDA has the best performance, taking only 0.26 s; SKMA has the second best performance, taking 0.27 s; RDAHM takes 2.05 s to process an image; and BE-ANFIS takes 18.56 s to assess an image. RDAHM requires slightly longer time than SKMA and RUDA, but is faster than BE-ANFIS. Moreover, RDAHM and RUDA could avoid recognition errors and save more time when assessing images containing no rust. In summary, RDAHM is an effective and reliable steel bridge rust assessment method which has been advanced in comparison with previous methods. With further modifications, RDAHM will be promising for real-time steel bridge rust assessment and worthy for further development to put automated inspection into practice.

參考文獻


Chang, J. L. and L. C. Chang (2007). Neural Network. Taipei, Tunghua.
AbdelRazig, Y. A. and L.-M. Chang (2000). "Intelligent Model for Constructed Facilities Surface Assessment." Journal of Construction Engineering and Management 126(6): 422-432.
Alex, J. S., S. Bernhard, et al. (2004). "A tutorial on support vector regression." Statistics and Computing 14(3): 199-222.
Besag, J. (1986). "On the statistical analysis of dirty pictures." Journal of the Royal Statistical Society Series B 48: 259-302.
Caelli, T. and D. Reye (1993). "On the classification of image regions by colour, texture and shape." Pattern Recognition 26(4): 461-470.

延伸閱讀