基於視覺分析之工業瑕疵檢測能判斷畫面中的產品是否存在異常狀況，並進一步定位瑕疵區域，對於自動化製造、生產與品質管制具有重要價值。一般而言，工業產品中的異常樣本數量較為稀少且型態多變，產品開發階段或生產初期亦不易取得大量正常樣本，使得傳統仰賴目標類別資料進行建模的方法在實用性與部署彈性上受到限制。近年來，CLIP 等視覺語言模型因具備跨模態語意表示與零樣本泛化能力，逐漸被應用於工業產品異常檢測任務。然而，原始 CLIP 主要學習自然影像中的全域語意與物件層級資訊，對於工業產品中常見的細微刮痕、裂紋、污染與局部紋理破壞等局部線索較不敏感；當正常與異常區域的外觀相近時，也可能產生特徵混淆，進而影響異常圖（anomaly map）的品質。本研究提出 ROSE-CLIP，一種基於局部不一致性建模之零樣本工業瑕疵檢測方法。ROSE-CLIP 引入 DPAM / V--V attention 局部視覺分支，以補充 CLIP visual tokens 中的局部視覺資訊，並進一步提出多尺度中心與周圍差異聚合（Multi-scale Center--Surround Discrepancy Aggregation， CSD-Agg），透過多尺度方式比較中心區域與周圍區域的特徵差異，以強化局部不一致性的表示。本研究同時設計雙向 Patch 分離損失（Bidirectional Patch Separation Loss， BPS Loss），從正常與異常兩個方向建立特徵分離機制，以降低局部特徵混淆並提升正常與異常局部特徵的可分性。此外，ROSE-CLIP 使用多組通用的正常與異常文字描述，並根據不同 visual tokens 與文字特徵之間的相似度選取較合適的語意參考，使不同局部區域能更彈性地利用正常與異常文字語意作為判斷依據。實驗採用 MVTec AD、VisA、DTD-Synthetic、BTAD、DAGM 與 MPDD 等六個資料集進行零樣本評估。實驗結果顯示，ROSE-CLIP 在影像層級（image-level）異常分類上，於平均 AUROC 與 AP 兩項指標皆取得最佳結果；在像素層級（pixel-level）異常定位上，亦於平均 AUROC 與 PRO 兩項指標取得最佳表現。可視化結果進一步顯示，ROSE-CLIP 所產生的 anomaly map 能更聚焦於真實瑕疵位置，並降低非瑕疵區域的異常反應。整體而言，本研究所提出的方法在保留 CLIP 跨模態語意泛化能力的同時，能強化局部異常感知、提升正常與異常特徵的可分性，並提升不同局部區域使用文字語意參考的彈性，對零樣本工業瑕疵檢測具有實質助益。;Vision-based industrial defect detection aims to determine whether products in images contain defects or anomalies and to further localize defective regions， which is of significant value for automated manufacturing， production， and quality control. In general， anomalous samples in industrial products are scarce and diverse， while a large number of normal samples may also be difficult to collect during product development or the early stages of production. As a result， conventional methods that rely on target-category data for model training may be limited in terms of practical applicability and deployment flexibility. In recent years， vision-language models such as CLIP have been increasingly applied to industrial anomaly detection due to their cross-modal semantic representation and zero-shot generalization capabilities. However， the original CLIP primarily learns global semantics and object-level information from natural images， making it less sensitive to subtle local cues commonly found in industrial defects， such as scratches， cracks， contamination， and local texture distortions. When normal and anomalous regions exhibit similar appearances， feature confusion may also occur， thereby degrading the quality of the anomaly map. To address these issues， this study proposes ROSE-CLIP， a zero-shot industrial defect detection method based on local inconsistency modeling. ROSE-CLIP incorporates a DPAM / V--V attention local visual branch to supplement the local visual information in CLIP visual tokens， and further introduces Multi-scale Center--Surround Discrepancy Aggregation (CSD-Agg)， which compares feature differences between center and surrounding regions across multiple scales to enhance the representation of local inconsistencies. In addition， Bidirectional Patch Separation Loss (BPS Loss) is designed to establish feature separation constraints from both normal and anomalous directions， thereby reducing local feature confusion and improving the separability between normal and anomalous local features. Furthermore， ROSE-CLIP employs multiple generic normal and anomaly text descriptions and selects more appropriate semantic references according to the similarity between individual visual tokens and text features， allowing different local regions to utilize normal and anomaly textual semantics more flexibly as semantic references for anomaly detection. Zero-shot evaluations are conducted on six datasets， including MVTec AD， VisA， DTD-Synthetic， BTAD， DAGM， and MPDD. Experimental results show that ROSE-CLIP achieves the best average AUROC and AP for image-level anomaly classification， as well as the best average AUROC and PRO for pixel-level anomaly localization. Visualization results further demonstrate that the anomaly maps generated by ROSE-CLIP are more concentrated on actual defective regions while reducing anomalous responses in non-defective areas. Overall， the proposed method preserves CLIP′s cross-modal semantic generalization capability while strengthening local anomaly perception， improving the separability between normal and anomalous features， and enhancing the flexibility of text semantic reference usage across different local regions， thereby providing substantial benefits for zero-shot industrial defect detection.
