應用共變異數矩陣與小波轉換於BGA基板線路瑕疵檢測

Automatic Visual Approaches for Ball Grid Array (BGA) Substrate Conduct Paths Inspection Using

葉繼豪(Chi-Hao Yeh)

指導教授 ： 蔡篤銘

元智大學/工程學院/工業工程與管理學系/博士(2001年)

摘要

本研究利用機器視覺技術對BGA (Ball Grid Array)基板線路進行短路(Short)、斷路(Open)、線路突出(Spur)、缺口(Mousebite)等重要瑕疵之偵測與分類。由於重要瑕疵均屬多重轉折之幾何結構，故線路之曲率會劇烈變化，所以本研究依此特性共提出兩種瑕疵檢測方法，方法一利用轉折偵測(Corner detection)技術偵測BGA基板線路曲率(Curvature)陡變位置，而線路上每一邊緣點之曲率值是由其相鄰邊緣點座標的共變異矩陣的特徵值(Eigenvalue)求得。當瑕疵位置偵測之後﹔利用線路邊緣點之特徵值變化模式(Pattern)與幾何形狀資訊可將瑕疵進一步分類為一正常線路轉折、短路、斷路、線路突出或缺口。
本研究之第二方法為利用一維小波轉換(1-D Wavelet transform)進行BGA基板線路瑕疵位置偵測。小波轉換可選擇不同之小波基底(Wavelet basis)將輸入訊號多階分解為低頻(Low-pass)平滑與高頻(High-pass)細節兩大部份。小波轉換已證實可將輸入訊號中局部變異(local deviation)凸顯於細節部份，因此，方法二先以BGA基板線路上每一邊緣點之斜率角度做為輸入訊號，再進行一維小波轉換以求得每一邊緣點相對應之一維小波係數值，藉由局部性之小波係數值變化以偵測瑕疵之位置。
由於方法一使用曲率變化偵測瑕疵，而曲率為二階微分(Second-order difference)求得，故需要較高之影像解析度以降低雜訊之影響。方法二之輸入訊號為每一邊緣點之斜率角度，而斜率角度為一階微分(First-order difference)，需要之影像解析度較低，故方法二可以有較大之檢測範圍。再者，若同樣以偵測瑕疵位置所需之計算時間比較，由於方法二需進行一維小波轉換，故計算時間較方法一為長。
本研究提出之兩種方法均不需模板(Template)訓練、無需參考樣本的特徵比對且不受待測物旋轉影響，適用於多種產品之小批量生產，能夠改良傳統視覺檢測印刷電路板線路瑕疵之缺失。在適當之影像解析度下，方法一與方法二之實驗範例皆可達100%瑕疵偵測率。

關鍵字

BGA基板線路 ； 瑕疵偵測 ； 瑕疵分類 ； 共變異矩陣 ； 小波轉換

並列摘要

The aim of this study is to exploit automatic visual approaches for detecting the boundary defects such as open, short, mousebite, and spur on Ball Grid Array (BGA) substrate conduct paths. Two proposed approaches are respectively based on the detection of local deviations of path boundaries using covariance matrix and 1-D wavelet transform. In the first approach, boundary defects are detected by a boundary-based corner detection method using eigenvalues of the covariance matrix obtained from the coordinates of neighboring boundary points. Detected defects are then classified by discrimination rules derived from variation patterns of eigenvalues and the geometrical shape of each defect type. Experimental results achieve 100% correct identification for BGA substrate boundary defects under a sufficient image resolution.
In the second approach, the 2-D boundaries of BGA substrate conduct paths are initially represented by a 1-D tangent curve. The tangent angles are evaluated from the eigenvector of a covariance matrix constructed by the boundary coordinates over a small boundary segment. Since defective regions of boundaries result in irregular tangent variations, the 1-D wavelet transform can decompose the 1-D tangent curve and capture irregular angle variations. Boundary defects can be located easily by evaluating the wavelet coefficients of the 1-D tangent curve in its high-pass decomposition. Experimental results also show 100% correct identification for numerous defective BGA substrate samples by selecting appropriate wavelet basis, decomposition level and image resolution.
Both defect detection approaches proposed in this study are invariant with respect to the orientation of BGA substrates, and do not require pre-stored templates for matching. These methods are suitable for inspecting various types of BGA substrates in small batch production because precise alignment of BGA substrates and the prestored templates are not required.

並列關鍵字

BGA substrate conduct paths ； Defect detection ； Defect classification ； Covariance matrix ； Wavelet transform

參考文獻

1.N. Chandler and S. G. Tyler, “Ultra-fine feature printed circuits and multi-chip

2.U. D. Perera, “Evaluation of reliability of mBGA solder joints through twisting

and bending”, Microelectronics Reliability, 39 (3), pp. 391-399, 1999.

4.D. Shen and H. H. S. IP, “Discriminative wavelet shape descriptors for recognition

5.S. Pittner and S. V. Kamarthi, “Feature Extraction From Wavelet Coefficients for