Efficient unsupervised domain adaptation for crack segmentation with interpretable Fourier– Morphology blending and Uncertainty‐guided self‐training
Saheli Bhattacharya, Chen Zhang, Dhanada K. Mishra, Matthew M. F. Yuen, Jize Zhang
Abstract
Automated crack segmentation models are vital for infrastructure monitoring but fail when deployed in new domains. Overcoming this domain shift without costly re-annotation is vital. This paper presents a novel unsupervised domain adaptation framework that uniquely integrates Fourier-based style transfer with targeted morphological operators and a robust Uncertainty-guided self-training scheme. Specifically, its Fourier–Morphology blending aligns visual styles and crack geometries between domains through controllable image processing operations governed by two intuitive parameters. This is paired with an Uncertainty-guided dual-network training scheme that safely leverages unlabeled target data for robust self-training. Experiments on public and industrial data sets show state-of-the-art performance, improving the F 1 $F1$ score by up to 18.5% over competitive baselines in challenging cross-domain scenarios.