StrainNet-3D: Real-time and robust 3-dimensional speckle image correlation using deep learning

This paper proposes a methodology of applying convolutional neural network (CNN) in solving 3D-DIC tasks. First, a multi-configuration stereo speckle dataset generation algorithm is designed with labels to train the networks. Then, an affine-transformation-based disparity calculation method and a light-weight CNN used for subpixel correlation are proposed. The three-dimensional displacement is calculated using the disparities and time-wise optical flow calculated by CNN, guided by stereo-vision theory and through an optional refiner network. After training, numerical experiments are carried out to verify the accuracy and the speed. Finally, real time high -resolution film bulging experiments are carried out which indicates the CNN-based method can achieve real-time and high-precision calculation with a comparable accuracy to DIC and an excellent robustness to intensity changes, assisted by the proposed gray adjustment technique. This method, named StrainNet-3D, may play an important role in experimental measurement tasks requiring real-time calculation.

Automated summary

This paper proposes a methodology of applying convolutional neural network (CNN) in solving 3D-DIC tasks. The method utilizes a multi-configuration stereo speckle dataset generation algorithm, an affine-transformation-based disparity calculation method, and a light-weight CNN for subpixel correlation. The experiments demonstrate that the CNN-based method achieves real-time and high-precision calculation with a comparable accuracy to DIC and excellent robustness to intensity changes.