缺陷影像擴增對半導體封裝檢驗卷積神經網路模型之績效評估

台灣半導體產值居世界第二，封裝測試占世界第一，隨著半導體產業製造技術不斷精進，對於品質的監控也帶來更大的考驗。透過人工智慧技術進行深度學習來判別半導體晶圓封裝的良品與缺陷品，取代過往使用人員目檢的方式，以提升檢驗效率。本研究的缺陷影像樣本以脫層缺陷為主，依脫層的發生位置可分為三種。本研究考慮影響績效的因子有三個，第一個為依照缺陷好發區擴增和隨機擺放缺陷擴增之兩種擴增方式，以此探討缺陷擴增的位置是否對於模型之績效有顯著影響。第二個因子為影像擴增倍率，以40倍之擴增倍率為上限，以間隔10倍的方式向下調整至20倍，以此探討擴增倍率對於模型之績效表現是否會有影響。第三個因子為卷積核尺寸大小，考量缺陷影像形狀大小不同，藉由測試不同卷積核大小，來探討此因子是否對績效有顯著影響。實務上進行檢驗績效討論時，主要指標是預測模型對缺陷品的偵測能力，此論文以準確率和漏檢率作為評斷標準。本研究將盲測樣本好壞品比例設為1:1，並採用判斷準確率高於90%之實驗結果進行分析。經實驗結果發現依缺陷好發區擴增方式，所得到的準確率較高，缺陷品漏檢率也較低，原因為其擴增影像較接近真實的缺陷品影像，預測模型對於缺陷品影像判別能力可以提升。在本研究中卷積核尺寸越大，準確率與降低漏檢率的表現較好，原因為預測模型對缺陷品影像上較大的缺陷有較佳的偵測能力。

關鍵字

半導體封裝；影像擴增；缺陷位置；卷積神經網路

並列摘要

Taiwan's semiconductor industry value ranks as second in the world, and packaging and testing industry ranks as first in the world. With the continuous improvement of manufacturing technology in the semiconductor industry, quality inspection also brings greater challenges. Through the deep learning of artificial intelligence technology, we can distinguish the good products and the defective products of IC packaging, replacing manual visual inspection, so as to improve the inspection efficiency. In this study, the defect samples are mainly delamination defects, and the defect images can be grouped into three types based on the location of delamination. There are three factors considered as independent variables to the performance. The first factor is to explore whether the location of defect augmentation has a significant impact on the performance of the model, therefore, one augmentation method is based on defect prone area, and the other is to randomly assign defect location. The second factor is the image augmentation ratio, which is from 20 to 40 times with 10 times as the difference, to explore whether the augmentation ratio will affect the performance of the model. The third factor is the convolution kernel size. Because the size of defect image is different, it is assumed that models with different convolution kernel sizes may have different performance. In real industrial applications, it focuses on the ability of the prediction model to identify defective products, and accuracy rate and under-kill rate are used as indexes. The ratio of good and defective products in the blind test samples is 1:1, and prediction results with accuracy rate higher than 90% are used for analysis. The experimental analysis shows that the accuracy rate of the prediction models with defect prone area augmented images are higher, and missing rates of defect detection are lower. The reason is that augmented images are closer to the real defect images, therefore, the model has better ability to distinguish the defect images. In this study, the larger the size of the convolution kernel, the better the accuracy rate and the lower the under-kill rate. The reason is the prediction model has better ability to detect the larger defects on the defective products images.