基於結構對齊優化的免訓練工業異常生成框架

Abstract

在工業缺陷檢測任務中，異常樣本生成(Anomaly Generation)對於克服缺陷 樣本稀缺的挑戰仍相當關鍵。儘管近期已有諸多進展，現有方法在面對複雜且多 樣的缺陷時仍有挑戰，尤其是在每種類型僅有單一異常樣本可用的情境下。為此， 本研究提出一種創新的免訓練異常生成架構——TF-IDG(Training-Free Industrial Defect Generation)，能於僅有單一樣本時產生多樣的異常樣本。我們引入「特徵 對齊策略」(Feature Alignment Strategy)，以縮小生成缺陷與真實缺陷之間的分佈 差距，提升外觀精準度;透過「自適應異常遮罩模組」(Adaptive Anomaly Mask Module)，有效解決小型異常易被忽略的問題，強化合成缺陷與遮罩的一致性; 同時，藉由「紋理保留模組」(Texture Preservation Module)，從正常影像中提取 背景線索，確保異常樣本與背景融合自然、逼真。評估結果指出，在 VisA dataset 中，本方法能夠產出更加準確且多樣的異常樣本，在圖像質量指標 Local IS 中達 到 3.90，異常分類準確度 (Acc) 達到 79.84%，突破現有先進方法的性能。同時 TF-IDG 顯著提升下游異常檢測模型的效能，提升先進異常偵測模型 5.5%。

Anomaly generation is crucial to address the scarcity of defective samples in industrial anomaly inspection. Despite progress, existing methods struggle with complex, multi-type defects, especially under one-shot settings. We propose TF-IDG, a novel training-free framework that synthesizes diverse anomalies from a single sample. A Feature Alignment strategy minimizes distributional gaps between generated and real defects to refine appearance. The Adaptive Anomaly Mask module mitigates the omission of small defects, improving consistency between synthetic anomalies and masks. To enhance realism, the Texture Preservation module extracts background cues from anomaly-free images. Experiments on the VisA dataset show that TF-IDG generates more accurate and diverse samples, achieving a Local IS score of 3.90 and anomaly classification accuracy of 79.84%, outperforming prior methods. Furthermore, TF-IDG improves downstream anomaly detection, with a 5.5% gain in advanced frameworks.