自工業 4.0 開始高科技計畫後，世界各國為了保持其優勢皆開始以智慧工廠為目標， 其中自動光學檢測的技術在半導體領域中扮演著重要的角色，自動光學檢測在業界已有 一定程度的發展，各廠商皆已發展出各自的架構，但現行自動光學檢測會因為外在因素 導致有誤判的行為，後續需要進行人工目檢的二次檢查，而在人工目檢時會倚賴製程人 員的專業知識判斷並受到主觀意識的影響，導致人工目檢沒有統一的檢測標準。 故本研究提出以現有架構為基準，以深度學習模型取代現有人工目檢的方式，達到 統一標準的檢測方式，降低對製程人員的依賴程度，且半導體工廠具備多條自動光學檢 測生產線，以傳統方式訓練的模型普遍具有災難性遺忘及模型更新速度耗時的問題，為 改善傳統模型的問題及符合半導體工廠的生態體系，本研究透過增量學習及深度學習模 型的結合提出 Adaptive Aggregation Networks on Gradient Episodic Memory 的系統架構， 以改善傳統模型的災難性遺忘及取代人工目檢為目的，達到提升產品的良率及減少時間 與人力的成本;Since Industry 4.0 started its high-tech project, countries all over the world have started to aim for smart factories in order to maintain their advantages, among which automatic optical inspection plays an important role in the semiconductor field. However, the existing automatic optical inspection may lead to misjudgment due to external factors, and the secondary inspection by manual visual inspection is required. Therefore, this study proposes to replace the existing manual visual inspection with a deep learning model based on the existing framework to achieve a unified standard inspection method and reduce the dependence on process personnel. In order to improve the problems of traditional models and to meet the ecological system of semiconductor factories, this study proposes a system architecture of Adaptive Aggregation Networks on Gradient Episodic Memory by combining incremental learning and deep learning models to improve the disastrous forgetfulness of traditional models and to replace manual visual inspection, so as to improve the product yield and reduce the time and labor cost.
