| 摘要: | 隨著深度學習的快速發展，將工業生產與深度學習結合，發展智慧製造工廠已經成為產業發展趨勢。對製造產業而言，提升生產線上的良率更是核心議題，除了提升生產技術，製程檢測準確率對於確保生產線的品質與效率更是關鍵議題。儘管現行自動化AOI光學檢測已經泛用於生產線中，然而其嚴苛的門檻值設定，雖然避免了瑕疵產品流入市場，卻導致大量良品被誤判為瑕疵品，導致生產線效能降低。針對AOI檢測的問題，需要額外的人力複檢成本解決此問題。
人力複檢成本高、效率低、且因人眼疲勞問題可能導致準確度下降。為了解決這個問題，我們提出利用深度學習來代替人工複檢程序。然而深度學習在工業檢測上又面臨樣本比例不均、瑕疵樣本種類未知性等問題，造成開發演算法上的困難。我們提出基於半監督式深度學習—GANomaly之優化方法；GANomaly是運用生成對抗式網路解決異常檢測問題的一個方法，然而其準確率與漏檢率尚未能達到應用於產線上的標準，因此我們提出如轉換色彩空間、調整損失函數和修改異常評分公式等研究方向以提升準確率，最後使此優化瑕疵檢測方法在資料集上達到高準確率與低漏檢率的表現。
;With the rapid development of deep learning, how to combine industrial manufacture with deep learning in order for developing smart factories has become one of the most recent trends. For the manufacturing industry, how to improve the yield rate of the production line is the core issue. In addition to improving the production techniques, the accuracy of the product defect inspection is a crucial issue to ensure the quality and efficiency of the production line. Automated optical inspection (AOI) is a technique in computer vision that combines image processing and automatic control techniques. In contrast to the traditional way of using optical instruments for product inspection by humans, the AOI techniques can lower the labor cost and shorten the inspection time. Although the current AOI detection techniques have been widely used in production lines, in practical cases, there are still many mis-classified samples which need to be double checked by humans. The reason is that the traditional AOI technique typically applies a hard threshold on extracted features as decision criterions which is not flexible enough to deal with practical manufacturing situations and therefore results in misclassification, which, in turn, increases the cost of the quality inspection.
Defective product confirmation by human labor is expensive, with low efficiency, and mis-classification may happen again in this stage due to eye fatigue. To solve this problem, we propose to use deep learning for the AOI problem. However, deep learning is faced with problems such as unbalanced sample size and unknown types of defective samples in industrial inspection, which causes difficulties in developing algorithms. In this thesis, we propose an optimization method based on the semi-supervised deep learning method: GANomaly. GANomaly is a method of using a generative adversarial network to solve anomaly detection problems, but its accuracy and missed detection rates have not yet reached the standards suitable for manufactural production lines. Therefore, we propose research directions such as color space transformation, the rethinking of the loss functions, and modifying the anomaly score to improve accuracy. Finally, our optimized method can achieve high accuracy and low false-positive rate and outperforms baseline methods like GANomaly and AnoGAN on our dataset. |
