ABSTRACT
Vision-based displacements can be used to measure response of structures subjected to dynamic loading. The accuracy of these measurements can be dependent on a number of factors, with image resolution being one of the most critical factors. This study focuses on improving the accuracy of displacement measurements via tests of a dynamically loaded, reduced scale, cantilevered beam using a consumer-grade camera. The effectiveness with which a deep learning based Super-Resolution approach improved image quality was studied by training an Enhanced Super Resolution Generative Adversarial Network (ESRGAN). ESRGAN aims to construct artificial high-resolution frames by upscaling original frames to twice their dimensions with sharper details, leading to enhanced measurement accuracy. ESRGAN was selected for study because of its reported ability to generate detailed and realistic images. To examine SR model accuracy, displacement measurements from the original frames and from SR-extracted frames were compared to reference displacements measurements taken using a contact sensor. Results showed that root mean square errors (RMSEs) of the measured dynamic displacements were smaller when SR image enhancement was used.
