| 摘要: | 本研究旨在利用深度卷積神經網路 (deep convolutional network) 檢測3D球柵陣列 (ball grid array, BGA) 中錫點大小變化不一致的瑕疵。球柵陣列是一種表面黏著技術 (surface mount technology, SMT)，錫點按陣列形式排列在積體電路元件下面，具有縮小積體電路、增加接腳、更好的散熱等優勢。空焊是指錫點上的錫球並未完美貼合在印刷電路板 (printed circuit board, PCB) 及電路元件之間，造成電路不通或不穩定。
由於錫球被夾在兩印刷電路板之間，無法直接觀察其形狀和大小，為了檢測球柵陣列中的錫球是否有瑕疵，因此利用 X 光檢測儀，對錫球的不同高度進行斷層掃描，依據物質吸收能量大小轉換對應之灰階，透過比較影像間灰階差異，可判斷是否為瑕疵影像。本實驗旨在檢測球柵陣列錫點中大小變化不一致的瑕疵，而傳統的影像處理方法無法有效地擷取到這類錫球變化特徵。
在本研究我們蒐集許多X光斷層掃描影像，並將這些影像作為深度卷積神經網路的輸入，讓網路模型擷取出斷層掃描影像間的特徵，找出瑕疵錫球，期望能提供一種端到端 (end-to-end) 且兼具高效率及高精度的自動光學檢測 (automatic optical inspection, AOI) 技術，降低球柵陣列中瑕疵錫球的漏檢率。
在本研究中，我們使用影像處理並搭配修改後CBAM-ResNet-50 架構作為主要的網路骨幹，內容包括：i. 使用強化對比，把斷層掃描影像中錫球與背景區別開，使網路訓練時有效學習到邊界資訊；ii. 增加差分影像，引導網路進一步學習斷層掃描影像中錫球大小變化特徵；iii. 網路架構加入預訓練及注意力模組，加強網路尋找特徵的能力。
本研究的訓練與測試資料是一組三張X光掃描錫球不同高度的斷層掃描影像所組成。資料集分為正常與瑕疵兩種類別，其中正常影像共有4,525組，瑕疵影像共有592組。訓練集正常影像共有3,621組，瑕疵影像共有472組；測試集正常影像共有904組，瑕疵影像共有120組。最後，使用影像擴增方法只將瑕疵的樣本總數提高10倍，總數量為5,920組。
在實驗結果上，我們分別分析資料處理和網路架構的改進對結果的影響。資料處理包括強化對比、差分影像、CLAHE 3.0、資料擴增、及亮暗度調整，使召回率從80.33% 提升到98.59%。網路架構加入使用ImageNet預訓練模型和加入CBAM注意力模組，使召回率從98.59% 提升到99.66%。
;This study aims to use deep convolutional neural network to detect the variation defect of ball grid array (BGA) packaging. Ball grid array packaging is a surface mount technology (SMT) that uses circular solder joints to distribute in an array form under integrated circuit components, with advantages such as reducing integrated circuit size, increasing pins, and better heat dissipation. Voiding refers to the solder ball not perfectly adhering between the printed circuit board (PCB) and the circuit components, causing the circuit to be disconnected or unstable.
Since the solder balls are sandwiched between two printed circuit boards and cannot directly observe their shape and size, in order to detect whether there are defects in the solder balls in the ball grid array, X-ray inspection equipment is used to scan the solder balls. X-ray inspection equipment can use the characteristics of X-rays to perform tomographic scanning at different heights of the solder balls, and convert the absorption energy of different materials into gray levels. By comparing the gray level differences of different tomographic images, we can judge whether there are defects in the solder balls. And traditional image processing methods cannot effectively extract the features of solder balls between different tomographic images.
In this study, we collected many X-ray tomographic images and used these images as inputs to a deep convolutional neural network, allowing the network model to learn the correlation between tomographic images and find defective solder balls. This study expects to provide an end-to-end automatic optical inspection (AOI) technology that is both efficient and accurate, and reduce the missed detection rate of defective solder balls in ball grid arrays.
In this study, we use image processing and modify CBAM-ResNet-50 architecture as the main network backbone, including: i. preprocessing the image, distinguish the solder point from the background using enhanced contrast, so that the network can effectively learn the boundary information when training; ii. increasing differential images, guide network to further learn solder point size change feature; iii. adding pre-training and attention module to network architecture, enhance network′s ability to find features.
The training and testing data are X-ray slice images composed of different heights and intensities. A group of images consists of three slice images. The data set is divided into two categories: normal and defective. Among them, there are 4,525 groups of normal images and 592 groups of defective images. The training set has 3,621 groups of normal images and 472 groups of defective images; The validation set has 904 groups of normal images and 120 groups of defective images. Finally, we use data augmentation method to increase the total number of defective samples by 10 times, with a total number of 5,920 groups.
In terms of experimental results, we analyze the impact of data processing and network architecture improvement on the results respectively. Data processing includes enhanced contrast, differential image, CLAHE 3.0, data augmentation and brightness adjustment, which improves recall rate from 80.33% to 98.59%. The network architecture adds ImageNet pre-trained model and CBAM attention module, which improves recall rate from 98.59% to 99.66%. |
