Enhanced pavement crack segmentation with minimal labeled data: a triplet attention teacher-student framework

Effective crack detection is critical for pavement maintenance, yet existing methods face significant challenges. Deep learning has become a popular solution due to its superior accuracy and automation capabilities, but supervised methods require extensive labelled data, making it costly and time-intensive. Unsupervised methods, while less data-reliant, struggle with feature extraction, especially for indistinct crack edges. Semi-supervised methods attempt to bridge this gap by using minimal labelled data, enhancing accuracy while reducing manual labelling efforts. However, accurate segmentation remains difficult, particularly with sparse labelled data or complex crack patterns. These limitations highlight the need for a more effective approach. To address these challenges, we propose a novel framework based on a teacher-student model enhanced with triplet attention. Leveraging the Unet architecture, this approach enables efficient feature extraction with minimal labelled data. The teacher model refines predictions using an exponential moving average of the student model’s weights, while the student model is trained with supervised and uncertainty losses. Evaluations on three benchmark datasets demonstrate our framework outperforms fully supervised models, achieving higher intersection-over-union and dice scores with just 2% labelled data. This approach overcomes the limitations of existing methods, enabling accurate crack segmentation with minimal labelled data.