本研究針對工業瑕疵檢測中標記成本過高與少樣本學習性能不佳的問題，提出SAM-VPG (Few-Shot Segmentation of Industrial Defects via Visual Prompt Guidance)方法。SAM-VPG以SAM(Segment Anything)架構為基礎，融合高效微調與精細視覺提示，並使用瑕疵知識遷移提升少樣本瑕疵分割的性能。SAM-VPG採用四項技術策略：(1)設計精細視覺提示，包括框提示與群中心點提示，提供更豐富的瑕疵空間與特徵資訊；(2)引入低秩適應(LoRA)高效微調策略增強模型泛化能力；(3)建立瑕疵遷移學習策略，提升少樣本分割的準確度；(4)整合輕量化架構以滿足實際部署需求。本研究在MVTec AD標準資料集與自建SPTD工業瑕疵資料集上進行0-shot至20-shot的效能評估，實驗結果顯示SAM-VPG優異的表現：在MVTec AD的0-shot測試中達到89.1% AUROC，展現良好的跨域泛化能力；在SPTD資料集上相較於基準方法可將IoU從50.5%提升至58.9%，驗證其在複雜工業場景的適應性。;To address the challenges of high labeling costs and poor few-shot performance in industrial defect detection, this study proposes SAM-VPG (Few-Shot Segmentation of Industrial Defects via Visual Prompt Guidance). Built upon the Segment Anything (SAM) architecture, SAM-VPG integrates efficient fine-tuning with refined visual prompts and leverages defect knowledge transfer to enhance few-shot segmentation performance. It adopts four technical strategies: (1) refined visual prompts, including box prompts and cluster-centroid point prompts, to provide richer spatial and feature information of defects; (2) incorporation of LoRA, a low-rank adaptation strategy, for efficient fine-tuning and enhanced model generalization; (3) construction of a defect transfer learning strategy to improve few-shot segmentation IoU; (4) integration of a lightweight architecture to meet practical deployment demands. The proposed method is evaluated on both the standard MVTec AD dataset and the custom SPTD (Spingence Tiny Defect) dataset across 0-shot to 20-shot scenarios. Experimental results demonstrate the superior performance of SAM-VPG: achieving 89.1% AUROC in 0-shot testing on MVTec AD, demonstrating strong cross-domain generalization; and improving IoU from 50.5% to 58.9% compared to baseline methods on the SPTD dataset, validating its adaptability to complex industrial environments.
