Distribution-aware Noisy-label Crack Segmentation

TL;DR

This work tackles the problem of noisy labels and limited cross-domain generalization in road crack segmentation by augmenting the SAM-Adapter with distribution-aware semantic guidance. It introduces a Mixture of Gaussians (MoG)–based per-image semantic representation to generate robust, distribution-informed labels that supervise the SAM-Adapter during training, mitigating the impact of label noise. The method achieves state-of-the-art results on Crack500 and shows strong cross-domain performance on the CFD dataset, outperforming prior CNN/Transformer and SAM-based approaches. The proposed framework enhances crack detection reliability in real-world deployment, supporting scalable road infrastructure inspection and maintenance.

Abstract

Road crack segmentation is critical for robotic systems tasked with the inspection, maintenance, and monitoring of road infrastructures. Existing deep learning-based methods for crack segmentation are typically trained on specific datasets, which can lead to significant performance degradation when applied to unseen real-world scenarios. To address this, we introduce the SAM-Adapter, which incorporates the general knowledge of the Segment Anything Model (SAM) into crack segmentation, demonstrating enhanced performance and generalization capabilities. However, the effectiveness of the SAM-Adapter is constrained by noisy labels within small-scale training sets, including omissions and mislabeling of cracks. In this paper, we present an innovative joint learning framework that utilizes distribution-aware domain-specific semantic knowledge to guide the discriminative learning process of the SAM-Adapter. To our knowledge, this is the first approach that effectively minimizes the adverse effects of noisy labels on the supervised learning of the SAM-Adapter. Our experimental results on two public pavement crack segmentation datasets confirm that our method significantly outperforms existing state-of-the-art techniques. Furthermore, evaluations on the completely unseen CFD dataset demonstrate the high cross-domain generalization capability of our model, underscoring its potential for practical applications in crack segmentation.

Figures (5)

• Figure 1: Examples of noisy labeled samples in the Crack500 ref8 dataset.
• Figure 2: Upper row: mislabeled sample in SAM-Adapter training. Mislabeled pixels disappear as training goes on. Lower row: changes of test sample predictions corresponding to training epoches.
• Figure 3: The proposed crack segmentation framework. The network is jointly trained by both the supervised loss and the domain-specific distribution-guidance loss.
• Figure 4: Comparison on Crack500 test dataset
• Figure 5: Comparison of different methods' zero-shot ability on the CFD dataset.