基於深度學習方法之軟性印刷電路板瑕疵檢測研究

Abstract

自動視覺檢測(Automated Visual Inspection，AVI)是結合了機器視覺與高精度光學影像檢測系統，運用於檢測產品瑕疵的一項技術，廣泛運用在產品製造的品質管理上。在印刷電路板(PCB)產業發展面對品質檢驗須達到更快速與更精準的目標，使用自動視覺檢測(AVI)取代人工目視檢驗(Visual Inspection)這是目前業界廣泛運用的技術。目前AVI檢測技術容易因檢測規格加嚴而造成AVI過篩現象及人力複檢成本提高。本研究針對AVI檢測不良瑕疵，透過卷積神經網路(convolutional neural network，CNN)設計實驗，將瑕疵影像分類經CNN機器學習，建立AVI瑕疵檢測辨識模型，以降低AVI過篩及人力複檢成本。本實驗分別以兩種產品及三種主要瑕疵為資料集，以驗證深度學習以驗證深度學習方法檢測模型之效力，兩種產品最多樣本總計為8000樣本與20000樣本。研究主要發現：(1)良品與不良品辨識模型：分類製品為良品或不良品，隨著訓練圖資批次增加數量，瑕疵檢測模型最終識別平均漏檢率可達小於5%與偽瑕疵篩檢率大於90%。(2)瑕疵分類辨識模型：依不同瑕疵分類建立訓練模型，瑕疵檢測模型最終識別平均漏檢率為小於10%、偽瑕疵篩檢率大於92%。綜言之，本研究發現藉由細部瑕疵分類辨識模型，可有效找出過篩的瑕疵類型及辨識較差的瑕疵分類來進行訓練模型調整，以輔助AVI檢測提高訓練模型的辨識正確率，減少AVI過篩情況，達到降低人員複檢成本的目標。Automated Visual Inspection (AVI) is a technology that combines machine vision and high-precision optical image inspection systems to detect product defects, and is widely used in the quality management of product manufacturing. In the face of the development of the printed circuit board (PCB) industry, quality inspection must achieve faster and more accurate goals. The use of automatic visual inspection (AVI) to replace manual visual inspection (Visual Inspection) is a technology widely used in the industry. At present, the AVI detection technology is prone to increase the AVI screening phenomenon and the cost of manpower re-inspection due to stricter detection specifications. In this study that aims at the detection of AVI defects, an experiment was designed through a convolutional neural network (CNN), and the defective images were classified by CNN machine learning to establish an AVI defect detection and identification model, so as to reduce the cost of AVI screening and manual re-inspection.In this experiment, two products and three major flaws were used as data sets to verify the effectiveness of deep learning to verify the effectiveness of the detection model of the deep learning method. The maximum samples of the two products are 8,000 samples and 20,000 samples in total. The main findings of the study: (1) good and bad product identification model: classified products as good or bad products, with an increase in the number of training image batches, the final identification of the defect detection model can reach an average missed detection rate of less than 5% and false defect screening. The detection rate is greater than 90%. (2) The defect classification and identification model: a training model was established according to different defect classifications. The average missed detection rate of the final identification of this model was less than 10%, and the false defect screening rate was greater than 92%. In conclusion, this study found that the detailed defect classification and identification model can effectively identify the types of defects that have been screened and classify poorly identified defects to adjust the training model to assist AVI detection, improve the identification accuracy of the training model, and reduce AVI. This model reduces the cost of personnel re-inspection.