基於生成式擴增與動態類別中心調整之工業瑕疵少樣本細粒度多分類研究;Few-Shot Fine-Grained Multi-Class Industrial Anomaly Classification via Generative Augmentation and Dynamic Class Center Adaptation

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/98312

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/98312

Title:	基於生成式擴增與動態類別中心調整之工業瑕疵少樣本細粒度多分類研究;Few-Shot Fine-Grained Multi-Class Industrial Anomaly Classification via Generative Augmentation and Dynamic Class Center Adaptation
Authors:	李奕緯;Lee, Yi-Wei
Contributors:	資訊工程學系
Keywords:	少樣本;細粒度工業瑕疵分類;多模態;資料擴增;類別中心;語意輔助;Few-shot learning;Fine-grained industrial defect classification;Multi-modal;Data augmentation;Class center;Semantic assistance
Date:	2025-07-22
Issue Date:	2025-10-17 12:37:26 (UTC+8)
Publisher:	國立中央大學
Abstract:	近年來，隨著深度學習與人工智慧技術的迅速發展，相關技術已廣泛應用於工業領域，特別是在智慧製造與自動化檢測方面展現出高度潛力。然而，在實際工廠場域中，高良率產線所產生的瑕疵樣本極為稀少，使得傳統深度學習模型所仰賴的大量標註資料難以取得。此外，同種類工業產品的瑕疵容易具備高度相似性的問題，類別間僅存在細微差異，如顏色、形狀或位置等，進一步加劇了少樣本下的細粒度分類困難。本研究針對上述挑戰，提出一套整合式分類架構，包含三個關鍵模組：受限條件下的多樣化特徵擴增模組、動態類別中心自適應調整模組，以及文字輔助的跨模態注意力模組。本方法透過對比式學習與多組判別器引導生成器於可控範圍內產生多樣化但語意一致的特徵，有效擴展決策邊界，同時利用特徵對齊與類別中心調整機制提升類內穩定性與類間可分性，再搭配文字語意特徵輔助分類，進一步增強模型對相似瑕疵的判別能力。在四個瑕疵資料集（X-SDD、NEU-DET、DAGM、Fabric）上，本方法在各項指標皆優於現有方法，其中於真實布料資料集中達到93.17% 的準確率，較 SemFew 方法[1] 的87.80% 更高，顯示其在工業應用場景下具有潛在的效能與泛化能力。;In recent years, deep learning and artificial intelligence technologies have shown great potential in smart manufacturing and automated inspection. However, in real-world factories, defect samples are extremely rare due to high-yield production lines, making it difficult to obtain sufficient labeled data. Moreover, defects in similar industrial products often exhibit subtle differences in color, shape, or location, posing challenges for fine-grained classification under few-shot conditions. To address these issues, this study proposes an integrated classification framework with three key modules: a Limited Diversity Feature Augmentation Module, a Dynamic Class Center Adaptation Module, and a Cross Attention Module enhanced by textual semantics. The framework uses contrastive learning and multiple discriminators to guide the generator in producing diverse yet semantically consistent features within a controllable range. It also incorporates class center alignment to improve intra-class compactness and inter-class separability, while cross-modal attention leverages text features to further enhance discrimination of similar defects. Experiments conducted on four defect datasets—X-SDD, NEU-DET, DAGM, and a realworld Fabric dataset—indicate that the proposed method achieves competitive performance compared to existing approaches. Specifically, on the real Fabric dataset, the method attains an accuracy of 93.17%, which is higher than the 87.80% attained by SemFew [1], suggesting its potential effectiveness and generalization capability in industrial applications.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	24	View/Open

社群 sharing

Loading...