應用電腦視覺與類神經網路於BGA檢測系統

Abstract

本文應用電腦視覺技術與改良式SOM網路(Self-Organizing Map進行BGA量測與缺陷的檢測，系統之基本架構主要是以擷取到的BGA原始影像，利用彩色像素群聚與類神經網路做最佳化的影像分割，最後用影像處理的技術，將BGA構裝型式的銲球直徑、銲球真圓度、銲球密度、球位偏差、球距、雙球、銲球破損或缺球與否的性質一一檢測出來。在BGA自動檢測系統的研究方法與處理程序上，首先需將銲球與BGA原始影像中銲墊、通孔與綠色導線分割出來，同時將所有之雜訊一併濾除，因此，如果要將BGA銲球與其它視為雜訊之影像分離出來，也是本文所究研的範圍。主要目的是希望從帶有雜訊的BGA影像中，利用類神經網路與形態學處理程序把非銲球之影像與雜訊去除，然後將所得BGA完整影像，利用區塊點標記、區塊分析，和區塊像素的邊緣檢測，將區塊內的銲球座標點找出，然後再利用最小誤差圓方程式與簽名(signature) 等，計算出銲球的大小、半徑、球距、面積、真圓度是否和標準值相近，最後配合相關係數的選定檢測銲球密度是否有缺陷。綜合以上的檢測結果，將缺陷的數量、種類與缺陷位置清楚的標示在原始BGA影像上，作為良品分類之參考。

This paper applies computer vision and the improved Self-Organizing Map (SOM) network techniques to perform BGA measurement and defect inspection. The basic structure of the investigated system is, mainly, to automatically measure or inspect, one by one, the high solder ball, the solder ball diameter of the fine-pitched BGA structure and mounting pattern, solder ball roundness, solder ball density, ball offset, ball pitch, double ball, the characteristics of solder ball damage or ball missing, using the retrieved BGA original image, the optimal image separation with color pixel grouping and neural network methods, and finally digital image processing techniques. In the aspects of investigation method and treating process of a BGA automated inspection system, we first need to separate solder balls from the solder pads, holes and green conductors in the original BGA image. Meanwhile, all noises must be filtered out. Consequentially, it is within the scope of this study to separate BGA solder balls from other images viewed as noises. The main purpose of this study is to filter out non-solder ball images and noises in the noise-corrupted BGA images, using neural network techniques and image filtering processes, to locate the solder ball coordinates within the regions in the retrieved full BGA image applying region point labeling, region analysis and edge detection of region pixels, and, then, to compute the size, diameter, ball pitch, area, and roundness-- whether they approach standard values-- of the solder balls using least error circle equation / best fit circle equation, signature, etc. Lastly, by matching the selected correlation coefficient, we inspect whether the solder ball density is defected. Combining the above-mentioned inspection results and applying various categories of defects as categorization standards, we clearly depict the quantities and categories of the defects and the locations of the defects in the original BGA image for the categorization reference for GOOD quality product.