應用快速之紋路一致性指標於表面瑕疵檢測

Fast regularity measures for surface defect detection

陳明君(Ming-Chun Chen)

指導教授 ： 蔡篤銘

元智大學/工程學院/工業工程與管理學系/碩士(2009年)

https://doi.org/10.6838/YZU.2009.00085

摘要

本研究利用機器視覺技術針對規律性之隨機性紋路及無紋路表面上細微變化之瑕疵進行檢測，由於檢測影像上的瑕疵區域並沒有特殊的結構或形狀特性，因此從視覺上看起來和正常部分並沒有太明顯的差異且不易觀察，使得檢測的工作極為困難。

本研究藉由觀察在固定視窗中影像的灰階變異情形及其在空間中的相關程度開發兩項技術來設計紋路一致性之特徵指標。第一項技術為透過主成分分析(Principal component analysis)法所設計之特徵指標，此指標是在視窗內以灰階值當成密度計算x軸和y軸兩個方向的共變異數，利用主成分分析法計算共變異矩陣所得到的固有值(Eigenvalue)之較小者λ2，而紋路不一致之瑕疵區域的λ2值會較正常區域為小。第二項技術為將每個視窗區分成數個大小相等但不重疊之小區塊，若檢測之表面具均勻性，則每一個區塊之灰階加總值應相類似，本研究針對此條件提出卡方χ2與熵(Entropy)兩個特徵指標，利用特徵指標在瑕疵區域和其他正常區域之間的差異，透過一個適當之指標閥值區分瑕疵和正常部份。為了加快指標值之運算速度，本研究利用積分影像(Integral image)的三步加減運算取代視窗或區塊內之逐點加總運算，使用積分影像前後λ2的處理時間改善了約400倍，而χ2與熵則改善了10倍。經測試筆記型電腦塑膠外殼、皮革表面、LCD背光模組中的擴散膜(Diffuser)之紋路影像及太陽能電池反面部份之無紋路表面影像後驗證本研究所提之特徵指標都能有很好的檢測效果。

關鍵字

表面瑕疵檢測 ； 紋路分析 ； 主成分分析 ； 紋路指標設計 ； 積分影像

並列摘要

This research proposes machine vision schemes for detecting subtle defects in non-textured and homogeneously textured surfaces. The defects to be inspected are ill-defined and hardly visible in the surfaces, which make the automatic surface inspection task extremely difficult.

In this study, regularity features of a small window sliding through the whole image are extracted based on the consistence of spatial distribution of gray levels in each window. Two methods are proposed. The first method is based on principal component analysis (PCA) that calculates the eigenvalues of the covariance matrix formed by the covariance of x- and y-coordinates with the gray level as the weight. The smaller eigenvalue λ2 is used as the regularity feature, where a defective region will generate a feature value smaller than that of a homogeneous defect-free region. The second method divides the sliding window into a set of small non-overlapped blocks. The sum of the gray levels in each block should be similar to each other if the window of the sensed image contains no defects. The Chi-square (χ2) that measures the difference between the gray-level sum of a block and the mean gray-level sum, and the entropy that measures the complexity of the gray-level sums in all the blocks are then used as the regularity measures. By using the integral image technique, the sum operations for all three proposed regularity measures can be efficiently calculated for on-line, real-time implementation.

The experiments on a variety of textured and non-textured surfaces including plastic case images of laptop computers, leather, TFT-LCD backlight panels and backsides of solar wafers have shown the effectiveness of the proposed methods. The computation times for an image of size 400×400 are only 0.032 seconds for λ2 and 0.28 seconds for χ2 and the entropy measures on a typical personal computer.

並列關鍵字

Machine vision ； surface defect detection ； texture analysis ； regularity measure ； principal component analysis ； integral image

參考文獻

Bodnarova, A., Bennamoun, M. and Kubik, K. K. (2000), “Suitability Analysis of Techniques for Flaw Detection in Textiles using Texture Analysis,” Pattern Analysis & Applications, Vol. 3, pp. 254–266

Bradley, D. and Roth, G. (2007), "Adaptive Thresholding Using Integral Image," Graphics Tools, Vol. 12, pp. 13-21.

Corneloup, G., J. Moysan and I. E. Magnin (1996), “Bscan image segmentation by thresholding using co-occurrence matrix analysis,” Pattern Recognition, Vol. 29, pp. 281-296.

Conners, R. W., C. W. Mcmillin, K. Lin and R. E. Vasquez-Espinosa (1983), “Identifying and locating surface defects in wood : part of an Automated Lumber Processing System,” IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. PAMI-5, pp. 573-583.

Chang, Y. J., Tyan, H. R. and Liao, H. Y. (1994), “Texture classification using C-matrix and fuzzy Min-Max neural network,” Visual Communications and Image Processing, Vol. 2308, pp. 809-818.