|
Title |
Crack Detection Framework with Domain Adaptation Using Weak Labels
|
|
Authors |
이근석(Geunseok Lee) ; 김방현(Banghyeon Kim) ; 조수진(Soojin Cho) |
|
DOI |
https://doi.org/10.11112/jksmi.2026.30.1.72 |
|
Keywords |
균열 탐지; 도메인 적응; 약한 라벨; 이미지 라벨; 포인트 라벨; 적대적 학습; 딥러닝 Crack detection; Domain adaptation; Weak label; Image label; Point label; Adversarial learning; Deep learning |
|
Abstract |
This study addresses the problem of performance degradation in deep learning-based crack detection models when applied across different environments due to inherent domain gaps between datasets. To counteract this, a crack detection framework was developed by implementing Domain Adaptation (DA) technique using weak labels, which can be made in both image-level and point-level. The methodology introduced three key improvements to existing DA methods: the application of dilation-erosion techniques, the integration of the Dice Coefficient Loss function, and the removal of Output Alignment Loss. Experiments conducted across three domain adaptation scenarios using four public datasets demonstrated the effectiveness of the proposed DA approach, achieving a maximum F1-score increase of 16.79%, 26.66%, and 18.78% across the tested scenarios.
Findings indicated that image-level pseudo-labels were highly effective when the domain gap was small, while point-level labels yielded greater benefits when the domain gap was large. Ultimately, this research confirms that employing weak labels enables efficient domain adaptation, resulting in high crack detection performance robust to various domain differences.
|