Stereo-DICNet: An efficient and unified speckle matching network for stereo digital image correlation measurement

Abstract

Digital image correlation (DIC) is an image-based, non-contact optical measurement technique for measuring the deformation of objects. Recently, deep learning has been preliminarily used in monocular-DIC measurement and has brought significant advantages in efficiency and robustness. However, the deep learning-based stereo-DIC methods is still underexplored, although stereo-DIC can measure three-dimensional deformation and shape compared to monocular-DIC. Different from monocular-DIC, stereo-DIC not only requires temporal matching (inter-frame) but also stereo matching (inter-camera), and the difference in image deformation scale between the two matching tasks make it difficult to perform them using a unified network. To solve this problem, we propose a unified speckle matching network that can be used for deformation measurement at different scales, referred to as Stereo-DICNet. Stereo-DICNet consists of three modules: feature extraction, cost volume generation, and displacement map prediction. Firstly, the feature extraction module composed of the residual layer and the spatial pyramid pooling layer is designed to extract the features of different receptive fields. Then, an attention connect volume was proposed to represent the similarity information between pixels at different spatial scales. Finally, the displacement map is predicted by a stacked hourglass-shaped sub-net. Simulation and physical experiments verify that Stereo-DICNet can perform both temporal matching and stereo matching. Moreover, material tensile experiments show that the MAE of Stereo-DICNet in the U, V, and W directions are no more than 0.005 mm, 0.015 mm, 0.01 mm respectively.