CrackUDA: Incremental Unsupervised Domain Adaptation for Improved Crack Segmentation in Civil Structures

TL;DR

This work tackles crack segmentation under cross-domain shifts by introducing CrackUDA, a two-step incremental unsupervised domain adaptation framework that preserves source-domain accuracy while learning from unlabeled target data. The method leverages an encoder with domain-invariant features and domain-specific decoders, coupled with adversarial training via a gradient reversal layer and a KL-divergence based consistency term to mitigate forgetting. It also introduces BuildCrack, a new drone-captured dataset; together with CrackSeg9K sub-datasets, the approach demonstrates improved generalization, achieving gains of 0.65 mIoU on the source and 2.7 mIoU on the target relative to strong UDA baselines. The results highlight CrackUDA's potential to enhance automated crack localization and structural health monitoring in civil engineering, offering a robust benchmark for future unsupervised domain adaptation in semantic segmentation.

Abstract

Crack segmentation plays a crucial role in ensuring the structural integrity and seismic safety of civil structures. However, existing crack segmentation algorithms encounter challenges in maintaining accuracy with domain shifts across datasets. To address this issue, we propose a novel deep network that employs incremental training with unsupervised domain adaptation (UDA) using adversarial learning, without a significant drop in accuracy in the source domain. Our approach leverages an encoder-decoder architecture, consisting of both domain-invariant and domain-specific parameters. The encoder learns shared crack features across all domains, ensuring robustness to domain variations. Simultaneously, the decoder's domain-specific parameters capture domain-specific features unique to each domain. By combining these components, our model achieves improved crack segmentation performance. Furthermore, we introduce BuildCrack, a new crack dataset comparable to sub-datasets of the well-established CrackSeg9K dataset in terms of image count and crack percentage. We evaluate our proposed approach against state-of-the-art UDA methods using different sub-datasets of CrackSeg9K and our custom dataset. Our experimental results demonstrate a significant improvement in crack segmentation accuracy and generalization across target domains compared to other UDA methods - specifically, an improvement of 0.65 and 2.7 mIoU on source and target domains respectively.

Figures (5)

• Figure 1: BuildCrack dataset was captured by imaging building facades using a drone-mounted camera from different angles and distances. BuildCrack has images with low contrast, occlusions, and shadows, which challenge the model's robustness. Sample images from our building crack dataset are shown. This dataset will be made public.
• Figure 2: Overview of our proposed architecture (Section \ref{['sec:implmentation']}). In step 1 we train our network, $M_1$, using the labeled source dataset $S$ for binary segmentation. In step 2, decoder $D_1$ and $\phi_{s_1}$ are frozen, and a new set of domain-specific parameters $\phi_{s_2}$ are added and we call this model $M_2$. An alternating training strategy is followed in which we first train for binary segmentation on the source domain followed by adversarial training on both source and target domains.
• Figure 3: Qualitative results for CrackSeg9K validation set for CrackUDA and FADA fada.
• Figure 4: Qualitative results for BuildCrack for our network and FADA fada.
• Figure 5: Some cases in which our approach does not perform well in CrackSeg9K and BuildCrack.