Accurate, automated diagnosis of pavement distress is essential for effective roadway maintenance but presents considerable challenges. Supervised learning methods are constrained by limited labeled data, while existing unsupervised representation learning approaches are difficult to capture the fine-grained details needed for precise pixel-level segmentation in pavement images with similar backgrounds. To address these limitations, we propose a novel unsupervised approach for pavement distress segmentation that employs a new pretext task within Siamese networks. Our method integrates an explicit prediction head and a high-dimensional cross-entropy loss, enabling implicit class labeling and enhancing fine-grained recognition of distress patterns. Additionally, vision transformers are employed to leverage self-attention mechanisms, facilitating accurate segmentation of foreground distress regions. Experimental results demonstrate that our approach outperforms existing unsupervised representation learning and anomaly detection methods. Notably, when used to pre-train backbone networks such as ResNet-50, our method yields higher accuracy and faster convergence on downstream supervised tasks compared to pre-training on the labeled ImageNet dataset. The proposed method holds promise for advancing pavement maintenance decision-making and enhancing the performance of traditional supervised deep learning models.

Detection and repair of pavement distress in time are crucial to maximize functional performance and service life, while minimizing maintenance costs on extensive roadway networks. Manual distress detection is labor intensive and error prone. While deep learning techniques offer unparalleled capabilities for automated and accurate pixel-level pavement distress segmentation, their reliance on extensive manual annotations remains a bottleneck. To address this challenge, we propose an open-ended self-supervised framework enabling flexible integration of various pretext tasks for pavement distress segmentation without manual annotations. We introduce a spatial contexts-informed pretext task that automatically generates pseudo labels by leveraging the highly consistent semantic information inherent across continuous pavement images within localized areas. Amulti-line parallel network architecture is then employed, where each line extracts a distinct deep representation aligned with the pseudo-label generation process. These representations are jointly optimized through a shared weight update scheme augmented by momentum encoders to capture long-range dependencies. Avision transformer processes the input images during inference, utilizing self-attention to highlight distressed regions based on the learned representations for precise segmentation. Extensive evaluations validate the performance of our framework, outperforming state-of-the-art self-supervised methods by 0.075 mIoU on average, while remarkably surpassing weakly supervised techniques requiring manual image-level annotations. These results are far more promising given that our self-supervised approach avoids human labeling costs, striking a trade-off between model effectiveness and annotation efficiency for large-scale deployments. It helps transportation agencies to realize timely, proactive infrastructure maintenance through scalable, accurate distress monitoring over extensive road networks.

Accurate and fast detection of pavement distress can provide reliable and effective technical support for pavement maintenance and rehabitation. Recently, deep learning has been widely used in pavement distress detection. However, its application is still limited by the laborious and difficult annotation process due to the complex topology of pavement distress. In this study, we propose a pavement anomaly detection network (PAD Net), which is a semi-supervised learning approach based on generative adversarial networks for identifying pixel-level anomalous image segments. We build a mapping function for unpaired abnormal and normal pavement images through a framework containing two generators and three novel discriminators. The framework is capable of maintaining background pixels and modifying anomalous foreground regions with the help of multi-style discriminators that consider interrelationships of multi-scale generated images. Meanwhile, pixel-level abnormal areas are detected through an end-to-end mask channel. Experiments show that our approach is able to achieve 80.75% accuracy on our dataset without pixel-level or patch-level annotations. Quantitative comparisons with several prior semi-supervised methods demonstrate the superiority of our approach.