基於半監督生成對抗網路的仿生膠薄膜缺陷檢測框架

Semi-supervised Generative Adversarial Network-Based Defect Detection Framework for Dry Adhesive Tape

本研究針對仿生壁虎膠帶（Bio-inspired gecko tape）在製程中面臨的多樣化瑕疵檢測挑戰，提出一種融合深度學習技術的自動光學檢測（Automated Optical Inspection, AOI）框架，以提升生產線上瑕疵識別的準確率與泛化能力。此框架核心結合半監督學習（Semi-Supervised Learning, SSL）與生成對抗網路（Generative Adversarial Networks, GAN），在有限標註資料下有效利用大量未標註數據，並透過條件生成對抗網路（Conditional GAN, cGAN）合成稀有且多樣的缺陷影像，解決資料不平衡問題，而系統採用卷積神經網路（Convolutional Neural Network, CNN）進行多類別瑕疵分類和膜厚度均勻性回歸預測，並藉由偽標籤生成與一致性正則化等半監督策略，促使模型在未標註資料上學習更穩健的特徵，提升對未知製程與光照條件變化的適應能力，研究配合多種測試場景，包括同一批次、跨批次與跨光照源影像，評估模型於不同條件下的魯棒性與泛化性。實驗結果顯示，整合模型在缺陷分類準確率達93.5%、F1-score達0.929，厚度品質回歸誤差顯著降低，並於跨域測試中保持較低的性能衰減，展現出強大的適應力與實務可用性，此框架不僅減少了對昂貴人工標註的依賴，同時提升了智能製造中薄膜膠帶品質監控的效率和精度，為高效、精準的工業瑕疵檢測提供了一種具備前瞻性及廣泛應用潛力的解決方案。

This study developed a deep-learning-based automated optical inspection framework to address challenges related to defect detection in the manufacturing of bio-inspired gecko tape. This framework integrates semi-supervised learning and generative adversarial networks to effectively learn defect features from large volumes of data with limited annotations. It then employs a convolutional neural network to perform multiclass defect classification and estimate film thickness uniformity. The semi-supervised learning module generates pseudolabels and applies consistency regularization, considerably enhancing the framework’s ability to learn robust representations from unlabeled data and improving its adaptability to process and lighting variations. Experiments were conducted across single-batch, cross-batch, and cross-illumination scenarios to validate the proposed framework’s robustness and generalization ability. This framework achieved a defect classification accuracy of 93.5% and an F1 score of 0.929 while substantially reducing thickness estimation error and maintaining stable cross-domain performance. The proposed framework reduces dependence on costly manual annotations and enhances the efficiency and precision of quality monitoring, thereby serving as a practical and robust solution for the intelligent detection of manufacturing defects in adhesive films.