Transformer-based video motion magnification for vision-driven structural health monitoring: Experimental validation on lab and full-scale structures

This paper investigates the potential of deep learning-based video motion magnification (VMM) to enhance vision-based structural health monitoring (SHM) under real-world conditions. The proposed approach integrates a Swin Transformer architecture to amplify subtle structural vibrations captured in video recordings, aiming to improve the stability and reliability of non-contact dynamic response measurements. Building upon prior research on learning-based VMM, this paper evaluates whether incorporating attention-based representations can mitigate typical image artifacts and noise amplification seen in traditional approaches and enable the extraction of quantitative vibration-related indicators relevant to SHM. Experimental validation on both laboratory-scale and full-scale structures demonstrates that STB-VMM provides sharper magnified motion and more stable frequency responses than a previous convolutional model, particularly when video quality or spatial resolution is limited. The results suggest that transformer-based magnification represents a promising approach for enhancing vision-based vibration measurements, contributing to the advancement of practical, scalable, and cost-efficient SHM using standard video recordings.