相對次數分配為基之獨立成份分析模式及其瑕疵檢測與動態影像偵測之應用

A relative frequency-based independent component analysis model for defect detection and motion detection

賴夏枝(Shia-Chih Lai)

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(2006年)

https://doi.org/10.6838/YZU.2006.00171

摘要

本研究提出以相對次數(Relative frequency)分配為基之獨立成份分析(Independent component analysis, ICA) 模式，利用相對次數分配估計出聯合機率密度函數及邊際機率密度函數做為量化獨立性的量測準則，並結合粒子群最佳化(Particle Swarm Optimization, PSO)演算法的方法來搜尋最佳解所建立的獨立成分分析模式，可以還原具相關性的原始訊號。傳統上獨立成份分析法是以訊號資料的非高斯特性(Non-gauissianity) 來量測其獨立性，利用非高斯特性建立的獨立成分分析模式來還原混合訊號，原始訊號間的相關性也會被去除，因此使用傳統的獨立成份分析法無法有效還原具高相關性之訊號。由於本研究方法之聯合機率密度函數在訊號個數大於或等於3時的估算不穩定，目前僅適合訊號個數2筆之應用。

本研究所發展之獨立成份分析模式，主要應用於瑕疵檢測 (Defect detection)與動態影像偵測 (Motion detection)。在瑕疵檢測方面乃以TFT-LCD液晶面板為討論對象，利用LCD面板區域元件皆呈現週期排列之特徵，與本研究之獨立成份分析模式可還原具有高相關性之原始訊號的特性，發展出一個不需對檢測影像做定位之自我影像(Self-reference)檢測的機器視覺演算法。而本研究於動態影像偵測的應用上，主要是針對室內固定位置之影像，利用動態物體與靜態背景互相獨立的特徵，經由本研究之獨立成份分析模式，發展出一個對於光源變化有較高的抵抗性，且可大量減少運算時間而達到及時偵測出前景目標物的演算法。

關鍵字

機器視覺 ； 獨立成份分析 ； 相對次數分配 ； 瑕疵檢測 ； 動態影像偵測

並列摘要

In this study, an independent component analysis (ICA) model that directly measures the difference of the joint probability density function (p.d.f.) and the product of marginal p.d.fs is proposed. The p.d.fs are estimated from the relative frequency distributions, and the particle Swarm Optimization (PSO) algorithm is used to search for the best solution in the ICA model. The proposed ICA model can separate highly correlated data, which is not achievable by the well-known FastICA algorithm that uses non-gaussianity as the independency measure. Since the p.d.f. estimated in the ICA model is simply based on the count of relative frequency, it can only well separate mixture of two source signals. When it comes to more than two signals, the p.d.f. estimation performs unreliably. Therefore, the applications of the proposed ICA model in this research are restricted to the separation of two sources.

The proposed ICA model is applied to defect detection and motion detection, where the underlying signals show high correlation. For defect detection, panel surfaces of TFT-LCD and color filter are the main targets of study. In a panel image, each scan line shows a periodical pattern. By dividing a scan line into two segments of equal length, the two segments are only different by their translations. The proposed ICA model is applied to filter translation changes. A cross correlation-based similarity measure can then be used to identify anomalies in the inspection surface. For motion detection, the proposed ICA model is applied to separate foreground objects from the stationary background. The implementation of the proposed method is computationally fast, and is insensitive to illumination changes. Experimental results have shown that the proposed methods are very efficient and effective for defect detection and motion detection applications.

並列關鍵字

Machine vision ； Independent component analysis ； Relative frequency ； Defect detection ； Motion detection

參考文獻

曾彥馨，2003年，「應用機器視覺於TFT面板之表面瑕疵檢測與分類」，碩士論文，私立元智大學工業與管理工程研究所。

林品杰，2005年，「應用獨立成份分析(ICA)濾波器於背光板與TFT-LCD面板之瑕疵檢測」，碩士論文，私立元智大學工業與管理工程研究所。

Bartlett, M. S., J. R. Movellan and T. J. Sejnowski, 2002, “Face recognition by independent component analysis,’’ IEEE Transactions on Neural Networks, Vol. 13, pp. 1450-1464.

Boscolo R., H. Pan, P. Roychowdhury, 2004, “Independent Component Analysis Based Nonparametric Density Estimation,” IEEE Transaction on Neural Networks, Vol. 15, pp. 55-65.

Comon, P., 1994, “Independent component analysis: a new concept?,’’ Signal Processing, Vol. 36, pp. 287-314.