鋼橋鏽蝕評估─以傅立葉轉換結合支持向量機增進照度不均勻影像辨識效果之研究

Abstract

鋼橋維護管理之良窳，關係日後使用年限與安全。評估鋼橋表面塗漆品質，一般是以表面塗漆缺陷面積所占的比例表示；儘管在塗漆缺陷分級上有明確的規範，但依靠人工目視判斷塗漆缺陷面積，無法擺脫不客觀、難維持一致性以及曠日費時的缺點。因此，運用數位影像處理技術執行自動化評估塗漆缺陷，有其存在的價值和必要。 過去的相關研究大多著重於灰階影像處理，其辨識效果會大幅受非均勻照度問題的影響。雖然 Adaptive Ellipse Approach (AEA) 和 box-ellipse-based ANFIS (BE-ANFIS) 的辨識機制能處理受非均勻照度影響的彩色影像，卻無法用於以紅色或棕色為背景顏色的鏽蝕影像。 本研究以支持向量機 (Support Vector Machine) 做為學習的機器，將彩色影像轉換至18種不同的色彩空間後，分別輸入54種色彩元素的組合加以測試，尋找即使受非均勻照度影響仍能表現出鋼橋鏽蝕影像特徵的模式。 測試結果顯示，xyY色彩空間中的x元素與y元素的組合，在非均勻照度和隨機混入雜訊的兩種條件下，能使得支持向量機在鏽蝕辨識上有顯著的效果；此一組合對於正常影像亦有優秀表現。 最後，本研究將彩色影像以xyY色彩空間中的x與y元素的組合搭配支持向量機，作為鋼橋進行自動化評估鏽蝕面積的工具，並利用傅立葉轉換與同態濾波器去調整非均勻照度影像的亮度和對比度，進一步提高鏽蝕辨識的精確率。

In many industries and fields of applications, image pattern recognition has been widely adopted over the last two decades. However, there are few robust methods for infrastructure maintenance to overcome non-uniform illumination problems in steel bridge rust assessment. Although the Adaptive Ellipse Approach (AEA) and the box-and-ellipse-based ANFIS (BE-ANFIS) methods have been proposed to deal with non-uniform illumination problems, they cannot work well on the images in red or brown background colors. The purpose of this research is to resolve non-uniform illumination problems for rust images in red or brown background colors. In order to find out the best color configurations for uniformly illuminated, non-uniformly illuminated, and random pepper-like noise rust image segmentation, an investigation of 18 color spaces is done. Among the 18 color spaces, the x and the y color configurations of the xyY color space have substantial performance over the simulation test. Therefore, the x and the y color configurations are used with Fourier transform, homomorphic filter and support vector machine (SVM) to improve the recognition accuracy. In an advanced SVM, Fourier transform with a homomorphic filter is used to adjust the lightness and contrast of non-illumination images for building up the training set, which is used to train SVM. The results show that the homomorphic filter can quickly adjust non-uniformly illuminated images of the training set for making steel bridge rust recognition more reliable than processing without illumination adjustment.