Deep DIC (Deep Learning for Digital Image Correlation)

A look at Deep Learning for predicting strain and displacement fields in material grains using computer vision and Digital Image Correlation

Digital image correlation (DIC) has become an industry standard to retrieve accurate displacement and strain measurement in tensile testing and other material characterization using computer vision and image datasets. This project proposes a new deep learning-based DIC approach – Deep DIC, in which two convolutional neural networks, DisplacementNet and StrainNet, are designed to work together for end-to-end prediction of displacements and strains [1].

A new dataset generation method is developed to synthesize a realistic and comprehensive dataset, including generation of speckle patterns and deformation of the speckle image with synthetic displacement field.

This project is an extension of the deepDIC work done by Prof. Ping Guo and Ru Yang at Northwestern University. The manuscript of this research work can be found in [1]: https://arxiv.org/abs/2110.13720

Dependencies

Deep-DIC is implemented in PyTorch and tested with Ubuntu 20.04, please install the dependencies using pip install requirements.txt .

• Python 3.7
• PyTorch (version = 1.16)
• Torchvision (version = 0.7.0)
• Pillow (version = 7.2)
• Scikit-learn==1.0.1
• numpy
• scipy
• CUDA

File Structure

  .
  ├── notebooks                             # contains ML model for DisplacementNet and notebook for testing ML model
  ├── dataset_generation                    # scripts to generate speckle image dataset
  ├── notebooks_archive                     # contains previously developed ML models to sanity check against
  ├── test                                  # test scripts and notebooks for sanity checking data
  ├── requirements.txt                      # install dependencies from this file
  └── README.md

Usage

Generate a dataset

First generate a dataset using python3 dataset_generation/generate_specklepattern.py

  .
  ├── ...
  ├── dataset_generation                             
  │     ├── generate_specklepattern.py                  # generate pairs of small & large deformation speckle patterns and their groundtruths
  └── ...

See README in ../dataset_generation ..

Train the model

Train the DisplacementNet model using the notebook DisplacementNet_train.ipynb. Ensure the correct paths to the training dataset and groundtruths is specified.

  .
  ├── ...
  ├── notebooks                             
  │     ├── DisplacementNet_train.ipynb                 # train DisplacementNet model on synthetic image dataset
  │     ├── experiment_displacementNet.ipynb            # test performance of trained D-Net model and visualize against groundtruth
  └── ...

Testing

Some testing scripts to check sanity of data and real tensile image samples

  .
  ├── ...
  ├── test                                  # test scripts and notebooks for sanity checking data
  │     ├── stats_analysis.ipynb            # compute statistical metrics of synthetic image dataset
  │     ├── log_visualize.ipynb             # visualize model training loss
  │     ├── tensile_images.ipynb            # compile dataset of real tensile test images
  │     ├── experiment_tensileTest.ipynb    # test prediction of D-Net on real tensile test images
  └── ...