本論文針對TFT面板瑕疵分類問題,探討以半監督式域自適應學習的方式,以達到縮短TFT Defect 分類網路在面對相同類別但新產品樣本資料的標記成本與訓練時間。論文的主要研究目標有二:第一,利用非監督式的網路架構,設計一個可拉開不同類別樣本間特徵距離的作法,以提高面板瑕疵分類的效果;第二,針對網路損失函數的設計進行分析與改善,以在維持原有網路架構的分類成效下,更適用於TFT Defect 分類問題。在實驗分析上,我們實際測試於包含5個產品的TFT Defect Dataset,並探討在不同比例的標註資料、不同參數設計、不同網路損失函數等對於分類成效的影響。實驗結果顯示,在只有少量Target Label的情況下,域自適應的深度學習網路架構的確可以成功應用在 TFT Defect 分類問題上,取得接近甚至比重新訓練更好的分類結果,有效地減少整體的模型訓練時間與資源成本。
This thesis focuses on thin-film-transistor (TFT) defect classification via semi-supervised domain adaptation. With the rapidly evolving TFT manufacturing process, automatic defect classification has become indispensable for TFT-LCD industry. Although deep-learning based methods have achieved remarkable performance for many classification tasks, the success of model training heavily relies on large-scaled and accurately-labeled datasets. In this thesis, we study and experiment using semi-supervised domain adaptation to facilitate the model trained on fully-labeled source domain into the target domain with few labeled data. Our goal is two-fold: to enlarge the inter-class feature discriminability alone with the model adaptation, and to study and compare different hyperparameter settings and loss terms for the TFT defect classification. Experimental results on the TFT Defect Dataset verify the effectiveness of our study and demonstrate its applicability to real-world problems.