A knowledge distillation framework for enhancing lightweight segmentation models with style-transferred data

Read full text Summarise with VIRAT
Authors: Jin Kim, Seungbo Shim, Jun-Beom An, Gye-Chun Cho
Cited by: 0
References: 59

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Critical infrastructure, such as tunnels and bridges, plays a vital role in improving quality of life. However, these structures inevitably deteriorate over time due to environmental influences, making accurate and efficient inspection methods essential. Traditionally, crack detection has relied on manual visual inspections, which are labor-intensive and prone to human error. While computer vision-based automation has shown promise, applying these technologies effectively requires the development of solutions that can account for diverse on-site conditions, which present significant challenges. To address this, we propose a novel framework that improves the performance of existing lightweight crack detection models by incorporating newly acquired data without the need for additional manual labeling. The proposed approach utilizes a generative adversarial network to synthesize training samples from the new data and applies knowledge distillation to transfer knowledge from the synthesized data to lightweight segmentation models. This semi-supervised learning strategy leverages both labeled and unlabeled data to enhance model accuracy without increasing annotation costs. The proposed framework was applied to six lightweight segmentation models, previously developed in other fields, achieving improvements in intersection over union (IoU) ranging from 5.313% to 8.794%. The results demonstrate that our approach consistently outperforms conventional training methods, contributing to the advancement of automated crack detection technology and enabling more efficient and reliable condition assessment of tunnels, bridges, and other critical infrastructure.

Fields of Study & Subjects

Infrastructure Maintenance and Monitoring Advanced Neural Network Applications Industrial Vision Systems and Defect Detection

Publication details

Journal / source
Engineering Structures
Volume
358
Pages
122615–122615
Publication date
21 March 2026
Language
EN