運用智慧型彩色影像辨識於鋼橋生鏽檢測

Smart Color Image Recognition for Steel Bridge Rust Inspection

楊雅晴(Ya-Ching Yang)

指導教授 ： 張陸滿

共同指導教授 ： 陳柏翰(Po-Han Chen)

國立臺灣大學/工學院/土木工程學研究所/碩士(2009年)

https://doi.org/10.6342/NTU.2009.01484

摘要

影像處理已廣泛用於學術研究及產業，於公共工程維護檢測之應用，包括鋼橋塗漆檢測及下水道管壁檢測等。雖然相關文獻顯示，K-Means聚類法是最有效的鋼橋生鏽偵測法，但此法仍無法穩定地辨識光度不均之影像，以及輕微生鏽部分；且在過去應用影像處理的鋼橋檢測研究中，尚無有效的模型可解決照片中光度不均之問題，亦未發展出自動化的彩色辨識系統。因此，本研究以鋼橋塗漆生鏽檢測為例，處理此二問題，並以發展自動化模型為目標。
為選取抵抗光度不均能力較佳之彩色座標，本研究首先從現今十四個常見的彩色座標系統中，選取相對最佳的生鏽辨識之彩色座標。經由實驗決定a*b*座標為最具抵抗光度不均能力之座標，本研究並以此座標發展以下兩個模型：Adaptive ellipse approach (AEA) 及 Box and ellipse-based neural fuzzy approach (BENFA)。
第一個模型Adaptive ellipse approach (AEA)中，一張生鏽影像被分為三個區域，生鏽、背景(即塗漆顏色)及輕微生鏽到背景顏色之漸變色區域。此模型可適當處理漸變色區域，排除光度不均之影響，以達到輕微生鏽辨識的目的。透過自動偵測背景，可決定基本的背景色；由收集的生鏽照片，作者以基本橢圓形定義生鏽顏色。本模型藉由擴大基本橢圓形加強偵測輕微生鏽顏色之成效，其擴大百分比取決於生鏽顏色與塗漆顏色之關係。與K-Means聚類法之處理結果比較後，顯示此模型可更適當地辨識輕微生鏽區域。
然而，當生鏽影像顏色分佈近似平行於基本橢圓形長軸時，AEA無法適當地辨識輕微生鏽顏色。有鑑於此，作者發展第二個模型Box and ellipse-based neural fuzzy approach (BENFA)以強化漸變色區域處理。本模型應用調適性網路模糊推論系統(Adaptive-network-based fuzzy inference system)描述漸變色。為達到自動化辨識之目的，此模型引用自動偵測背景、光度調整及基本橢圓形，以決定輕微生鏽和嚴重生鏽的門檻值。研究發現，相較於Fuzzy C-Means聚類法，此模型可更穩定地辨識鋼橋表面的生鏽程度。
最後，為修正光度不均之生鏽照片，作者發展第三個模型BEMD-morphology approach (BMA)。此模型應用二維經驗模態分解法(bidimensional empirical mode decomposition)降低陰影之影響，並且應用影像形態學(morphology)重建反光點之顏色。結果顯示，以K-Means聚類法處理經由此模型修正後之影像的結果，比起處理未修正影像時更接近實況。

關鍵字

塗層缺陷辨識 ； 影像處理 ； K-Means聚類法 ； 調適性網路模糊推論系統(Adaptive-network-based fuzzy inference system) ； Fuzzy C-Means聚類法 ； 二維經驗模態分解法(Bidimensional empirical mode decomposition) ； 形態學(morphology)

並列摘要

Image processing has been widely utilized in scientific research and prevalently adopted in industries. Application in infrastructure condition assessment includes defect recognition on steel bridge painting and underground sewer systems. Nevertheless, there is still no robust method to overcome the non-uniform illumination problem. Although, the K-Means is recognized as one of the best rust defect recognition methods, it cannot recognize the non-uniform illuminated images and the mild rust color well. Also, there is lack of an automated color image recognition system in this field.
This research starts with an investigation of 14 color spaces in order to find out a comparatively proper color configuration for non-uniformly illuminated rust image segmentation. Among the 14 color spaces, the color configuration of a*b*, which has moderate ability to filter light, is utilized to develop the proposed two models, adaptive ellipse approach (AEA) and box and ellipse-based neural fuzzy approach (BENFA).
In the adaptive ellipse approach (AEA), a rust image is partitioned into three parts, background, rust, and the gradual change color from mild-rust to background. The main idea is to deal with the gradual color change properly for mild rust color extraction. The background colors can be automatically detected from a rust image. A fundamental ellipse is previously defined by the collection of rust colors. The AEA enlarges the fundamental ellipse to include part of the gradual change in color, and the enlarged size depends on the relationship between the rust color and the color of coating. The AEA is expected to deal with the boundary between background color and rust color properly. In addition, illumination adjustment is adopted in this model in order to overcome the non-uniform illumination problem. Finally, the processing results of the AEA are compared with the K-Means clusters method to show that it can recognize the mild-rust-colors.
When the color distribution is almost parallel to the major axis of the fundamental ellipse, the proposed AEA may not recognize the mild-rust-colors well. Therefore, the box and ellipse-based neural fuzzy approach (BENFA) is proposed to deal with the gradual color change from mild-rust to background. The BENFA applies the adaptive-network-based fuzzy inference system (ANFIS) to describe the gradual change colors. In order to achieve automated detection, the BENFA applies the automated detection of background, illumination adjustment, and the fundamental ellipse to determine the thresholds of serious rust and mild rust. Compared to the Fuzzy C-Means (FCM), the BENFA can stably recognize the rust intensity.
The third model which is called BEMD-morphology approach (BMA) aims to adjust the color of a non-uniformly illuminated rust image. The BMA applies the bidimesional empirical mode decomposition (BEMD) to mitigate the shade/shadow effect, and morphology to substitute the highlight points by the neighboring colors. Processing a rust image with the BMA is more reliable than processing without the BMA.
Finally, conclusions will be drawn and recommendations for future work will be made.

並列關鍵字

Coating defect recognition ； image processing ； K-Means ； adaptive-network-based fuzzy inference system (ANFIS) ； Fuzzy C-Means ； bidimensional empirical mode decomposition (BEMD) ； morphology

參考文獻

Chang, J. L. and L. C. Chang, Eds. (2007). Neural Network. Taipei, 東華.

Adachi, Y., A. Imai, et al. (2000). "Extraction of face region by using characteristics of color space and detection of face direction through an eigenspace." The Fourth International Conference on Knowledge-Based Intelligent Engineering Systems and Allied Technologies, 2000.

Bezdek, J. C., R. Ehrlich, et al. (1984). "FCM: The fuzzy C-Means clustering algorithm." Computers & Geosciences 10(2-3): 191-203.

Bhagavatula, R. and M. Savvides (2007). "Analyzing Facial Images using Empirical Mode Decomposition for Illumination Artifact Removal and Improved Face Recognition. IEEE International Conference on Acoustics, Speech and Signal Processing, 2007.

Carr, J. C., W. R. Fright, et al. (1997). "Surface interpolation with radial basis functions for medical imaging." Medical Imaging, IEEE Transactions on 16(1): 96-107.