Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor | Department of Electrical and Electronic Engineering | en_US |
| dc.creator | Qin, Hao | - |
| dc.identifier.uri | https://theses.lib.polyu.edu.hk/handle/200/13900 | - |
| dc.language | English | en_US |
| dc.publisher | Hong Kong Polytechnic University | en_US |
| dc.rights | All rights reserved | en_US |
| dc.title | Industrial defect detection and localization using large vision-language model | en_US |
| dcterms.abstract | Industrial Anomaly Detection (IAD) is crucial for identifying defects in manufacturing processes, where anomalies often signify potential failures that can lead to significant losses. However, real-world anomaly detection faces challenges, particularly the scarcity of defective samples due to the rarity of abnormal conditions in actual production. Traditional methods combining machine learning and image processing often fall short in meeting industrial requirements. In recent years, deep learning models, particularly those leveraging visual-language approaches like CLIP, have demonstrated remarkable progress in zero-shot and few-shot anomaly detection. | en_US |
| dcterms.abstract | This dissertation explores the application of a CLIP-based framework for industrial defect detection and localization. We propose a novel approach that enhances the existing Prompt Learning paradigm, addressing its limitations in detecting fine-grained anomalies. The key contributions of this work include: 1) designing a category-independent prompt template that avoids semantic interference by focusing on anomalies rather than object-specific descriptions, and 2) introducing learnable token embeddings into the text encoder to refine textual representation and improve alignment with anomaly semantics. | en_US |
| dcterms.abstract | Our model combines these improvements with advanced attention mechanisms to enhance both global and local feature representation. Experiments were conducted on two benchmark datasets, MVTec AD and VisA, under zero-shot and few-shot settings. The results demonstrate that our approach achieves state-of-the-art performance, with an AUPRO of 91.5 on MVTec AD and 83.5 on VisA, outperforming existing methods like AnomalyCLIP and WinCLIP. Visual comparisons of segmentation maps further confirm the model's superior ability to capture fine-grained and localized defect patterns. | en_US |
| dcterms.abstract | While our method represents a step forward in unsupervised IAD, further research is needed to meet the high accuracy demands of industrial applications. Future work could focus on leveraging incremental data for model fine-tuning and exploring the broader potential of CLIP-based models in industrial settings. | en_US |
| dcterms.extent | vii, 45 pages : color illustrations | en_US |
| dcterms.isPartOf | PolyU Electronic Theses | en_US |
| dcterms.issued | 2024 | en_US |
| dcterms.educationalLevel | M.Sc. | en_US |
| dcterms.educationalLevel | All Master | en_US |
| dcterms.accessRights | restricted access | en_US |
Files in This Item:
| File | Description | Size | Format | |
|---|---|---|---|---|
| 8308.pdf | For All Users (off-campus access for PolyU Staff & Students only) | 6.52 MB | Adobe PDF | View/Open |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/13900