工業異常檢測已成為電腦視覺與智慧製造領域中的重要研究議題，其中自動化檢測系統被廣泛應用於產品品質控管與缺陷偵測。然而，現有的異常檢測方法仍面臨數項挑戰，包括難以辨識細微的局部缺陷以及逐類別訓練方法在可擴展性上的限制。為了解決上述挑戰，本論文提出 AceGPT，一個少樣本工業異常檢測框架，結合了文字與視覺表徵的動態對齊機制，以及多種推論階段的增強技術。首先，本研究引入 Text Prompt Adapter，用於優化由大型語言模型生成的類別特定描述，進而提升文字與視覺表徵之間的對齊效果。此外，本研究採用 Hybrid Reference Selection 策略，在推論階段選取具有多樣性的正常參考影像。進一步地，本研究使用 Geometric Test-Time Augmentation，將多個幾何視角所產生的異常圖進行聚合，以提升定位結果的一致性與穩健性。該框架亦整合語言模型，提供包含異常判斷、空間定位與缺陷描述在內的文字輸出。實驗結果顯示，AceGPT 在 MVTec AD 與 VisA 兩個基準資料集上皆展現出良好的效能。在 2-shot 設定下，AceGPT 於 MVTec AD 達到 97.1% 的 imagelevel AUROC，於 VisA 達到 92.4%，優於多個近期方法，包括 AnomalyGPT、PromptAD 與 IADGPT。此外，AceGPT 僅需使用 3，594 張訓練影像進行單一訓練階段，並且能在不同 shot settings 下泛化而無需重新訓練，提供了一種有效且具訓練效率的少樣本工業異常檢測解決方案。;Industrial anomaly detection has become an important research topic in computer vision and smart manufacturing， where automated inspection systems are widely used for product quality control and defect detection. Nevertheless， existing anomaly detection methods still face several challenges， including the difficulty of identifying subtle local defects and the limited scalability of per-class training schemes. To address these challenges， this thesis proposes AceGPT， a few-shot industrial anomaly detection framework that combines adaptive text-visual alignment with multiple inference-time enhancement techniques. A Text Prompt Adapter is introduced to refine class-specific descriptions generated by large language models， thereby improving the alignment between textual and visual representations. In addition， a Hybrid Reference Selection strategy is adopted to select diverse normal reference images during inference. Furthermore， Geometric Test-Time Augmentation is employed to aggregate anomaly maps generated from multiple geometric views， improving localization consistency and robustness. The framework further integrates a language model to provide textual outputs including anomaly judgment， spatial localization， and defect description. Experimental results demonstrate that AceGPT achieves strong performance on both the MVTec AD and VisA benchmark datasets. Under the 2-shot setting， AceGPT achieves 97.1\% image-level AUROC on MVTec AD and 92.4\% on VisA， outperforming several recent approaches including AnomalyGPT， PromptAD， and IADGPT. Moreover， AceGPT requires only a single training stage using 3，594 training images and generalizes across different shot settings without retraining， providing an effective and training-efficient solution for few-shot industrial anomaly detection.
