應用深度學習技術於手機玻璃表面之瑕疵檢測

Abstract

科技不斷進步，智慧型產品與人類生活息息相關，如手機、平板電腦與智慧手錶。由於防水、耐用、高透光與不易被腐蝕等特性，使得玻璃成為智慧型產品螢幕表面之保護材料。為了提高產品的品質與良率，檢測成為生產線上不可或缺的重要環節；然而，人工檢測的主觀性判斷使得檢測標準參差不齊，且其帶來的高人事成本，促使自動化檢測導入漸漸成為一種趨勢。 自動化檢測中，以影像處理相關之演算法為現行最被廣泛使用的技術。仰賴長期所累積的知識與經驗，影像處理技術具備解決大部分瑕疵檢測的能力；但是，在一般生產環境下，影像處理技術易受光源與環境影像，使參數設定泛用性降低，且愈來愈嚴格的檢測標準將提高研發成本並拉長研發所需時間。為了克服現行傳統影像處理演算法技術瓶頸，本研究提出一應用深度學習技術之手機玻璃表面瑕疵檢測方法。研究架設包括取像系統、光源與電腦。本研究執行步驟，首先進行影像前處理，包括取像影像分割、資料強化、瑕疵分類等步驟建立影像資料庫；緊接著，進行神經網路訓練以不同的卷積神經網路與優化器訓練獲得最佳手機玻璃表面瑕疵檢測方法。本研究展現如何建立一以深度學習技術檢測手機玻璃表面瑕疵的方法與流程，並透過神經網路訓練結果證實其應用於手機玻璃表面瑕疵的突破。

Inspection systems have been widely applied in industries to examine defects in products. Investigations into the method of defect inspection are important for manufacturing process. In this paper, we present an inspection approach based on deep learning, with a view to distinguishing flawed smartphone cover glass and to classify different kinds of glass defects. In contrast with traditional techniques, image processing techniques (IPTs), which often creates challenges to algorithm calculations due to inconsistent real-world variables like lighting and shadow changes, deep learning-based defect inspection technique is less time-consuming. This study proposes a method using deep learning architecture of convolutional neuron networks (CNNs) for detecting cover glass defects. We cropped the images into pieces sized 300'×' 300 pixel resolutions and classified defects features like scratches, dust, dirt and fur, which were subsequently learned automatically by different models of CNN, including LeNet-5, AlexNet, VGG and GoogLeNet. As the CNN finished auto-learning, it was then capable of detecting and classifying different kinds of defects. The best results among the used models showed that the accuracy of a trained network was 98% when classifying cover glass defects in the right categories, proven to perform better compared with IPT. This paper concludes that the proposed method can indeed find the defects in realistic situations with a more ease manner.