Abstract: Supervised learning-based visual models are effective for tunnel surface disease detection, but model training requires large datasets. Additionally, diseases evolve dynamically, and models trained on data from a specific period may perform well currently but cannot guarantee future detection accuracy. Continuously retraining models with new data incurs significant costs. To address this, a three-stage active learning method is proposed to reduce annotation costs through selective labeling and continuously select new training samples to update the model, ensuring its performance throughout its service life. In the first stage, a classification model selects disease samples based on global feature differences and removes interference from non-disease samples. In the second stage, classification probability and Intersection over Union (IoU) are used to predict sample difficulty, selecting those with higher difficulty for annotation to improve detection accuracy. In the third stage, samples are mapped to feature space, and image similarity is measured using geometric methods to obtain diverse, representative samples, upgrading disease features and reducing redundancy. Experimental results show that selecting 17% of the samples for annotation in the three-stage active learning method achieves 94.0% accuracy for the trained model when annotating all target class samples, ensuring accuracy while reducing annotation costs.

Key words: tunnel lining, disease detection, labeled data, dynamic evolution, active learning