Dual-Branch Network via Multiple Illumination-Aware Representation Learning for Steel Surface Defect Classification


Yong Seok Oh (Dongguk University), Min Geol Kim (Sogang University), Bogyeong Kim (Dongguk University), Jun Young Kim (Dongguk University), Hyeongseob Jo (Dongguk University), Jae Hyeon Park (Dongguk University), Gyoomin Lee (Dongguk University), Sung In Cho (Sogang University)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)Supplementary

Abstract

On highly reflective metal surfaces, strong lighting-induced reflections can obscure or distort defect regions, making accurate defect detection challenging. To mitigate this, various classification methods using multi-light source images captured under diverse lighting conditions have been proposed. However, such approaches struggle with two critical issues: (i) failing to clearly separate defect-like reflections from meaningful defect cues, leading to visually ambiguous representations, and (ii) incapable of capturing lighting-dependent variations in defect appearance. In this work, we propose a Dual-branch network that disentangles illumination-invariant and lighting-dependent representations via two dedicated encoders: a view-consistent encoder and a view-specific encoder. The view-consistent encoder extracts defect features that are robust to misleading reflections and invariant to changes in illumination, whereas the view-specific encoder focuses on capturing fine-grained, lighting-dependent visual cues around defect regions. This design enables the model to learn simultaneously generalized and view-adaptive representations of defects. Experimental results demonstrate that our method significantly outperforms existing multi-light source approaches in terms of average classification accuracy.

Citation

@inproceedings{Oh_2025_BMVC,
author    = {Yong Seok Oh and Min Geol Kim and Bogyeong Kim and Jun Young Kim and Hyeongseob Jo and Jae Hyeon Park and Gyoomin Lee and Sung In Cho},
title     = {Dual-Branch Network via Multiple Illumination-Aware Representation Learning for Steel Surface Defect Classification},
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_354/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