瑕疵資料生成與檢測模型建置技術

本技術針對現有工件表面檢測方法存在人工誤差高與自動化光學檢測（Automation Optical Inspection, AOI）系統彈性不足等問題，提出一套結合生成式AI與虛實整合平台的瑕疵資料生成與檢測模型建置技術。系統整合機械手臂與影像擷取模組，透過收集正常與異常表面影像，經標註與預處理後作為生成模型訓練資料，運用擴散模型與生成對抗網路技術擴充數據，並結合虛擬環境中光源與相機位置隨機生成機制，模擬實際光照與視角條件，建立多樣化標註資料。檢測模型採卷積神經網路進行訓練與驗證，經由不同樣品測試，結果顯示系統具備高準確率、良好推論效能及部署彈性。本系統不僅提供智慧製造場域自我擴充與高度可用的檢測方案，亦具備應用於其他多變製程的潛力。

This study addresses the issues of high human error in existing workpiece surface inspection methods and the limited flexibility of AOI systems, and proposes a defect data generation and detection model development technique that integrates generative AI with a cyber-physical platform. The system incorporates a robotic arm and an image acquisition module to collect both normal and defective surface images. These images, after annotation and preprocessing, are used as training data for generative models. Data augmentation is performed using diffusion models and generative adversarial networks, while a virtual environment equipped with randomized lighting and camera configurations simulates diverse illumination and viewing conditions, thereby generating enriched labeled datasets. The inspection model is trained and validated with convolutional neural networks, and evaluation on different samples demonstrates high accuracy, strong inference performance, and deployment flexibility. Beyond providing a self-expanding and highly reliable inspection solution for smart manufacturing environments, the system also shows strong potential for application in other complex production processes.