ScholarWorks@성균관대학교

초록

Accurate crack segmentation is vital to civil infrastructure quality and safety, yet its automated in-service monitoring is severely hindered by the high cost of pixel-level data annotation. While generative models are often explored for data augmentation, this approach struggles to efficiently produce high-quality segmentation labels. More critically, Denoising Diffusion Probabilistic Models (DDPM) exhibit a training instability on monotonous crack images, manifesting as severe color-shift artifacts that contaminate features and cause catastrophic segmentation failures. To address these challenges, this paper introduces DiffEncoderCrack, a novel semi-supervised framework that pioneers a "diffusion-as-encoder" paradigm. Instead of generating images, this approach repurposes a pre-trained DDPM as a powerful feature encoder to learn intrinsic geometric representations directly from unlabeled data, thereby achieving high label efficiency. The framework's success is underpinned by a spatial decomposition prior, implemented via a dual-branch architecture. This mechanism resolves the training instability by decoupling the learning of geometry from background statistics, thus purifying the feature representations. Furthermore, a systematic representation analysis across timesteps and network depths is conducted to identify the most discriminative features for the downstream task. Experiments demonstrate the framework's exceptional label efficiency. With a small fraction of labeled data, the proposed method significantly outperforms existing semi-supervised approaches and achieves performance comparable to, or even exceeding, fully-supervised baselines trained on the entire dataset. This work establishes a robust and label-efficient methodology, offering a scalable solution for automated infrastructure assessment. The code is publicly available at: https://github.com/LiXY-Civil/DiffCrack.git.

키워드