Unsupervised vision-based structural anomaly detection and localization with reverse knowledge distillation

Abstract

Most of vision-based methods for structural damage detection rely on supervised learning, requiring a substantial number of labeled images for model training, which is labor-intensive and time-consuming. To address these challenges, this study introduces a vision-based structural anomaly detection and localization approach using unsupervised learning and reverse knowledge distillation. The proposed model incorporates a teacher model, a student model, and a trainable one-class bottleneck embedding module. The asymmetrical architecture of the teacher and student models forms an encoder-decoder structure for parameter transfer and feature extraction. The student network receives a specific embedding from the teacher network as input and target, facilitating the recovery of multiscale information from the teacher. Training images only contain the undamaged structures, and the teacher model, a pretrained model, instructs the student model to remember their undamaged features to detect and localize damages in unseen testing images. Through experiments, including a comparison among five candidate backbones for pretrained teacher models based on the residual network and testing across various structural damage types, the optimal model is identified, demonstrating good performance in both anomaly detection and localization. Furthermore, the model’s generalization performance is thoroughly validated, confirming its efficacy across diverse scenarios.