Image Correlation CNNs for Large-Deformation Strain Measurement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing digital image correlation (DIC) methods struggle with accurately predicting full-field strain fields, especially at large deformations, and are computationally inefficient, particularly when speckle patterns deteriorate, and previous deep learning applications have not demonstrated significant performance improvements over traditional DIC.
Innovation Solution
An end-to-end deep learning-based DIC approach using separate convolutional neural networks (DisplacementNet and StrainNet) directly predicts displacement and strain fields from image pairs, with a modified encoder-decoder structure and synthetic dataset generation to enhance robustness and accuracy, and adaptively updates the region of interest to track large deformations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional DIC methods are used to measure full-field displacement and strain, then measurement coverage is improved, but computational efficiency deteriorates especially at large deformations
Solution Approach 1:
The patent replaces traditional mechanical correlation algorithms with a deep learning-based system that uses convolutional neural networks to directly predict displacement and strain fields from image pairs, eliminating the need for iterative correlation calculations and significantly improving computational efficiency while maintaining measurement accuracy
Solution Approach 2:
The patent changes the fundamental approach by training neural networks on synthetic datasets with known displacement and strain fields, allowing the system to learn optimal feature representations and prediction models that can accurately capture large deformations without the computational burden of traditional methods
2Reliability
If traditional DIC methods are used, then measurement robustness is maintained, but accuracy deteriorates at large deformations
Solution Approach 1:
The patent performs preliminary training on synthetic datasets that include various deformation scenarios before applying to real measurements, allowing the neural network to learn robust feature relationships and deformation patterns in advance, which then enables accurate prediction of large deformations in actual experiments
Solution Approach 2:
The patent uses synthetic datasets as copies of real measurement scenarios to train the neural network, where the synthetic images and ground truth displacement/strain fields serve as training examples that replicate the complexity of real large deformation cases without requiring actual experimental data for training
3Productivity
If deep learning approaches are applied to DIC, then computational efficiency is improved, but performance improvement over traditional DIC is insufficient
Solution Approach 1:
The patent segments the deep learning architecture into separate convolutional neural networks for displacement prediction and strain prediction, with each network specialized in its specific task, allowing optimized feature extraction and prediction for each field while maintaining overall computational efficiency
Solution Approach 2:
The patent transitions from 2D image input to multi-dimensional output by predicting both displacement vectors and strain tensors simultaneously from image pairs, adding the dimension of strain information directly to the prediction output and enabling comprehensive mechanical analysis
Data Source
AI summary
A system for correlating image data includes a memory configured to store a sequence of images of a sample. The system also includes a processor operatively coupled to the memory and configured to crop a first pair of images to specify a region of interest in the first pair of images, where at least one image in the pair of images is from the sequence of images. The processor is also configured to calculate, using a first convolutional neural network, a displacement field for the first pair of images. The processor is also configured to calculate, using a second convolutional neural network, a strain field for the first pair of images. The processor is further configured to determine an amount of displacement or deformation of the sample based at least in part on the displacement field and the strain field.