Super-robust digital image correlation based on learning template

Abstract

Digital image correlation (DIC) is a well-established noncontact displacement measurement method. Nevertheless, there are still three aspects that can be further improved: accuracy, efficiency, and robustness. Compared with accuracy and efficiency, there are few studies on the robustness of DIC method. However, there are many factors that make the DIC algorithm no longer robust, such as cracks, discontinuous displacements, large inter-frame deformation, partial occlusion or contamination, perspective changes, image noise, and illumination variations. This work proposes a learning template based super-robust DIC method which mitigates the influence of all these disadvantage factors in a straightforward and unified approach, significantly improving DIC measurement robustness compared to classic approaches. During the implementation of this method, the online learning and offline learning methods are adopted to keep the correlation at a high level. Interestingly, the learning template can be seen as a deep learning method with only one layer, so this work also explicitly explains why deep learning can be applied to the DIC field. Uniaxial tension tests were conducted and showed that the matching error can meet the accuracy requirements of the initial value. Experiments on landslide images also show that our method significantly improves the matching robustness.