Machine Learning-Aided Spatial Adaptation for Improved Digital Image Correlation Analysis of Complex Geometries

Digital Image Correlation (DIC) is a widely used experimental technique for measuring full-field deformations. However, its applicability to samples with irregular geometries, particularly for measuring deformations near sample edges, has been limited. This is because DIC subset windows must be split when their centroids approach the sample edge, and portions of the subset outside the sample region need to be discarded, which often requires significant manual intervention. In this paper, we present a novel machine learning-aided approach that significantly improves the efficiency and speed of DIC post-processing. Our approach begins by utilizing the recently developed Segment Anything Model 2 (SAM 2) to rapidly generate initial masks that delineate regions of interest within the sample. Compared to conventional image segmentation methods, this approach reduces computation time by one to two orders of magnitude. These masks then undergo spatial and temporal refinement to improve accuracy and consistency. The refined masks serve as input for our recently developed SpatioTemporally Adaptive Quadtree mesh DIC (STAQ-DIC) method, which automatically selects regions of interest and generates adaptively refined meshes, particularly around areas with complex geometries. DIC subsets near sample edges or internal voids are automatically split, further improving the accuracy of the resolved displacement fields. Through multiple case studies, we demonstrate the effectiveness of this approach in accelerating and improving DIC analysis for complex geometries. It provides a more efficient and accurate means of measuring deformations in challenging experimental scenarios, while minimizing manual intervention and processing time. Additionally, we provide an open-source code that is freely available to use our approach.