DefectGPT: Towards Multi-Class Defect Detection with Limited Electrical Samples


Zhuoyi Lin (Institute for Infocomm Research, Agency for Science, Technology and Research (A*STAR)), Kaixin Xu (Institute for Infocomm Research, Agency for Science, Technology and Research (A*STAR)), Aye Phyu Phyu Aung (Institute for Infocomm Research, Agency for Science, Technology and Research (A*STAR)), Wen Qiu (Advanced Micro Devices (Singapore) Pte Ltd), Bernice Zee (Advanced Micro Devices (Singapore) Pte Ltd), Jiann Min Chin (Advanced Micro Devices (Singapore) Pte Ltd), Senthilnath Jayavelu (Institute for Infocomm Research, Agency for Science, Technology and Research (A*STAR))
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

Existing deep learning (DL) approaches for electrical defect detection typically require large quantities of labeled data. However, in real-world scenarios, defect samples are significantly rarer than normal samples, making data collection both challenging and resource-intensive. Meanwhile, the wide variability in anomaly types and the unpredictable nature of defect locations further complicate the labeling process. In this paper, we present DefectGPT, a novel approach for multi-class electrical defect detection with limited samples by leveraging multimodal large language models (MLLMs). Specifically, DefectGPT generates defect descriptions to establish a rich semantic context for subsequent reasoning and classification. To effectively utilize limited training samples and enhance classification accuracy, we introduce a novel in-context learning technique, termed hypothesis-first learning (HFL), which facilitates the DefectGPT to generate an initial hypothesis before refining its knowledge. This approach enables DefectGPT to improve defect classification progressively through self-reflection, enhancing its effectiveness and generalization with limited electrical samples. We further conduct a theoretical analysis to gain deeper insights into the underlying learning mechanism of DefectGPT. Experimental evaluations on two distinct electrical datasets demonstrate the effectiveness and generalizability of our approach, highlighting its potential to reduce data dependency while maintaining high classification accuracy.

Citation

@inproceedings{Lin_2025_BMVC,
author    = {Zhuoyi Lin and Kaixin Xu and Aye Phyu Phyu Aung and Wen Qiu and Bernice Zee and Jiann Min Chin and Senthilnath Jayavelu},
title     = {DefectGPT: Towards Multi-Class Defect Detection with Limited Electrical Samples},
booktitle = {36th British Machine Vision Conference 2025, {BMVC} 2025, Sheffield, UK, November 24-27, 2025},
publisher = {BMVA},
year      = {2025},
url       = {https://bmva-archive.org.uk/bmvc/2025/assets/papers/Paper_502/paper.pdf}
}


Copyright © 2025 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection