工業自動光學檢測（Automated Optical Inspection， AOI）於產線開線之冷啟動情境下，面臨真實瑕疵（NG）樣本稀缺與像素級標註成本高昂之雙重困境。傳統無監督異常檢測（Unsupervised Anomaly Detection， UAD）雖可繞過對 NG 樣本之依賴，但其決策邊界建立於正常樣本分佈之假設上，於工業背景高度變異之條件下易產生結構性誤報；生成式合成路徑（GAN / Diffusion）雖可擴增訓練樣本，然其運算成本高昂且物理保真度與標註可控性難以兼顧，於敏捷部署情境下實用性受限。 本研究以資料為中心（Data-centric）之路線提出跨域瑕疵移植（Cross-Domain Defect Transplantation）框架——自 CRACK2019 公開裂痕資料集萃取具備真實物理拓撲特徵之裂痕幾何遮罩，以高斯模糊（Gaussian Blur）進行邊界平滑後融合至目標產線之無瑕疵影像之背景，於極低運算成本下產出幾何對齊精確且邊界框標註完整之偽瑕疵(Fake NG)訓練樣本。框架進一步搭配主動背景負樣本注入（Active Background Negative Injection）與混合增強（Hybrid Augmentation）策略，將原本受限於無監督框架之冷啟動任務重新定義為標準全監督學習任務，並於 YOLOv11m 物件偵測與 EfficientNet-B0 + SVM 影像分類雙軌進行訓練與驗證。 於 KolektorSDD（金屬絕緣塑料表面）三折交叉驗證下，本研究方法以零真實 NG 之純合成配置達平均 AP = 97.68%，已超越 Božič 等[1] 使用全 33 張真實 NG 弱標籤之 N=0 基準（93.4% AP），並逼近其使用 5 張像素級標註之 N=5 基準（99.1% AP）；於 Magnetic-Tile（陶瓷磁磚）跨材質場域亦達 F1 = 97.35%。雙場域於極小樣本混合訓練後，於本研究測試規模下均達 F1 = 100% 之飽和效能，獨立實證跨材質裂痕幾何移植可行性之工作假設。本研究於工業 AOI 之冷啟動與中期混合訓練兩情境下確立監督式優勢，並提供從純合成至混合訓練之漸進式部署路徑。;Industrial Automated Optical Inspection (AOI) at the cold-start stage suffers from a fundamental dilemma: real defective (NG) samples are scarce and pixel-level annotation is costly. Unsupervised Anomaly Detection (UAD) bypasses NG dependence but is structurally prone to false positives under variable industrial backgrounds， while generative synthesis (GAN / Diffusion) is limited by heavy computational cost and the trade-off between physical fidelity and label controllability. This study proposes a data-centric framework， Cross-Domain Defect Transplantation (CDDT)， which transplants crack-geometry masks extracted from the public CRACK2019 dataset onto Real OK backgrounds via Gaussian-blur boundary smoothing， producing fully annotated Fake NG samples at negligible cost. Coupled with Active Background Negative Injection (ABNI) and Hybrid Augmentation， the framework reformulates the cold-start task into a fully supervised problem， validated on YOLOv11m and EfficientNet-B0 + SVM. On KolektorSDD under three-fold cross-validation， CDDT achieves an average AP of 97.68% using zero real NG samples， surpassing the Božič et al. (2021) N=0 baseline (93.4% AP) and approaching their N=5 pixel-supervised baseline (99.1% AP). On Magnetic- Tile， it reaches an F1-score of 97.35% under the same cold- start condition， and both domains saturate at F1 = 100% under few-shot hybrid training at the test-set scale adopted in this study—empirically validating cross-material crack-geometry transplantation. The proposed framework establishes a Supervision Efficiency Advantage and Universality Advantage， offering a Progressive Deployment Path from pure synthetic cold-start to mature hybrid training.
