Digital image correlation (DIC) is an image registration technique to measure finite three-dimensional shape and deformations of planar and curved surfaces. This technique requires an optimal unique pattern or set of unique localized patterns as a carrier of deformation information in order to accurately measure correlations in temporal images. Recent advances in obtaining an optimal pattern in terms of saliency and uniqueness require operators’ experience and/or prior metrics. In our study, we propose a preprocessing methodology to automatically classify the saliency and uniqueness of a localized pattern for DIC processing of a large structure for structural health monitoring. In order to ensure pattern saliency, we develop a localized multi-scale CNN classifier using an in-house dataset containing 20k unique coarse and fine patterns. This classifier ensures that the projected pattern is salient within a real world image. For ensuring uniqueness within an image and a set of images, we develop a novel uniqueness algorithm that ensures the structural similarity (SSIM) index of the pattern is above a similarity threshold in every part of an image as well as for all subsequent images. We integrate these algorithms as a preprocessing step to our in-house 3D-DIC program for an efficient study of 3D vibrations of large-sized structures. Initial experiments are performed on a large-sized (10m height) light tower, and it is observed that our methodology is capable of optimizing the size, saliency, and uniqueness of a pattern in order to perform efficient displacement measurements for vibrational study and health monitoring purposes.
|