基於GAI之瑕疵影像生成研究

面對勞動力老齡化與經驗斷層，製造業正加速導入人工智慧與自動化光學檢測技術至品質檢測中，以提升檢測效能。然而高良率產線常面臨瑕疵樣本稀缺問題，也成為訓練高穩健性AI檢測模型的瓶頸。為此，本研究開發模仿式瑕疵影像生成模組，將資料收集策略由被動收集轉為主動生成。模組使用局部填補技術，將瑕疵大小與位置參數化，納入生成模型的約束條件，使生成瑕疵特徵趨近於真實影像。實測顯示，在僅有100張真實瑕疵影像的情況下，透過GAI擴增瑕疵影像，將其加入檢測模型訓練，使召回率從72%提升至99%，有效降低漏檢率，並快速建置檢測系統。

To address workforce aging and the technical discontinuity, industries are accelerating the adoption of artificial intelligence (AI) and automated optical inspection (AOI) in quality inspection processes to improve the efficiency. However, high-yield production lines often face a scarcity of defect samples, creating a major bottleneck for training robust AI inspection models. This study develops an imitation-based defect image generation module that shifts data acquisition from passive collection to active generation. The module applies local inpainting to incorporate user-defined parameters—such as defect size and location—as controllable constraints in a generative model, so that the synthesized defects closely match those observed in real images. In an evaluation setting with only 100 real defect images, additional defect images were augmented using generative AI (GAI) and added to the training set of the inspection model. This augmentation method improves defect detection recall rate from 72% to 99%, effectively reducing missed detections and enabling rapid deployment of the inspection system.