Pytorch implementation of StrainNet-3D.

(Including training codes of SubpixelNet and RefineNet, and the affine-transformation-based 3D displacement calculation workflow)

• A CNN-based 3D displacement calculation method for stereo speckle images. (To realize real-time and high-precision 3D-DIC)
• Method work flow:
• Simulated stereo speckle images with displacement labels (to train or to evaluate this method) can be found in my another repository Here
• Core CNN used to calculate subpixel displacement: A light-weight CNN named SubpixelCorrNet (Architecture shows below).
• For algorithm details and the principles, please see Wang et al., 2022

• python38
• opencv-python (4.4.0 used)
• numpy (1.22.1 used)
• torch with cuda (torch1.9.0+cu111)

• Train the networks
  • Run Train_FlowNet.py
  • If you want to train FuseNet, displacement cache file calculated by SubpixelCorrNet should be generated first.
  • If you want to generate your own training dataset, please refer to the Stereo Speckle Generator
• 2-D displacement (and deformation) calculation
  • Run DispCalculator2D.py, remember to shift the model filepath and the image filepath to your own ones.
  • Strain calculation using gradient filter technique is provided.
• 3-D displacement (and deformation) calculation
  • Run DispCalculator3D.py, remember to shift the model filepath and the image filepath to your own ones.
  • The calibration parameters and other settings should be set correctly.
  • Simple strain calculation code using gradient filter is provided.

• Pre-trained parameter file of SubpixelCorrNet can be download from Google drive.

• Comparison of the 3D displacement calculated using StrainNet3D and 3D-DIC of one test set.

• Experimental speckle images calculation in extreme light conditions.

• Light-changing real-time displacement monitoring demo. The Realtime demo video can be found here

@article{WANG2022107184,
title = {StrainNet-3D: Real-time and robust 3-dimensional speckle image correlation using deep learning},
journal = {Optics and Lasers in Engineering},
volume = {158},
pages = {107184},
year = {2022},
issn = {0143-8166},
doi = {https://doi.org/10.1016/j.optlaseng.2022.107184},
url = {https://www.sciencedirect.com/science/article/pii/S0143816622002378},
author = {Guowen Wang and Laibin Zhang and Xuefeng Yao}