A loosely coupled serial digital image correlation method based on deep learning

Abstract

Digital image correlation (DIC) is a non-contact optical measurement method of displacement field, which has been widely used in various fields. Traditional DIC method needs to manually complete the tedious DIC settings such as speckle region selection and subset setting to cope with the background. In order to reduce tedious DIC settings, realize real-time online measurement, and handle various measurement objects with backgrounds, automatic selection of the speckle region is required. A loosely coupled serial measurement network SCorrNet is proposed to automatically segment speckle regions and measure its displacement fields, so as to eliminate the influence of background. With segmented speckle region as the global feature and constraint, the displacement computation is focused on the speckle region using the attention mechanism. To further improve the measurement accuracy and practicability, a diverse dataset with speckle images, background and large deformation is proposed. Finally, the proposed method is compared with traditional and learning-based DIC methods on self-built and public datasets, and validated in two practical experiments. The experimental results show that SCorrNet has high accuracy and speckle segmentation capability to reduce the influence of dynamic background. Therefore, our method has high practical value.