Closing the data gap: Few-shot roadbed health assessment with self-supervised visual representations

As the core load-bearing structure of the road, the condition of the roadbed surface is a key indicator for evaluating road service performance. However, current deep learning-based algorithms heavily rely on large amounts of training data, and the scarcity of samples significantly hinders the automation of roadbed inspection. Facing the challenge of scarce training samples, this paper takes enhancing the network’s scene semantic understanding and category discrimination capabilities as the core, and proposes the Few-Shot Road Roadbed Defect Classification Network (FRNet). First, gradual knowledge transfer paradigm is proposed: self-encoding learning endows the network with scene semantic understanding, followed by task-oriented fine-tuning to enhance the model’s sample discrimination capability, effectively reducing reliance on large-scale annotated data. Second, a class-guided feature space structuring mechanism is embedded within the network architecture to optimize the spatial clustering distribution of image features, significantly improving classification accuracy in few-shot scenarios. Finally, through comparative experiments, ablation studies, and cross-domain transferability tests, the proposed network’s excellent generalization ability and classification performance are systematically validated, while the underlying mechanisms of its performance improvement are thoroughly analyzed. On the constructed roadbed defect dataset, the proposed FRNet achieves outstanding classification accuracies of 84.46% on the 1-shot task and 90.66% on the 6-shot task, significantly outperforming existing mainstream networks.