DIC-Net: Upgrade the Performance of Traditional DIC with Hermite Dataset and Convolution Neural Network

Hermite dataset generation implementation (matlab) and DIC-Net CNN implementation (pytorch).

Introduction

Our method is currently the SOTA of 2D-DeepDIC method in terms of accuracy and spatial resolution(SR).

In the experimental test set with 3216 samples, DIC-Net-d achieves an absolute pixel error (MAE) score of 0.0130 pixels and 0.0126 pixels in the x direction and y direction which is only 48.5% and 47.9% of the prior best method. The spatial resolution (SR) is 17.25 pixels with a noise level of 0.0136, and the metrological performance indicator is 0.234 (lower is better) which outperforms existing traditional and non-traditional methods.

For more details, please refer to our paper: https://doi.org/10.1016/j.optlaseng.2022.107278

• Frame of DIC-Net

• Frame of Hermite Dataset

Main Results

• Prediction error statistics of three networks on 3216 samples

Displacement mode (number of samples)	Network	Average MAE/pixel		Average RMSE/pixel
		u	v	u	v
Pure translations (785)	DIC-Net-d	0.0124	0.0115	0.0155	0.0147
	DisplacementNet	0.0182	0.0154	0.0197	0.0172
	StrainNet-f	0.0265	0.0255	0.0442	0.0418
Gaussian (445)	DIC-Net-d	0.0128	0.0122	0.0161	0.0156
	DisplacementNet	0.0171	0.0164	0.0199	0.0194
	StrainNet-f	0.0256	0.0265	0.0441	0.0433
Periodic displacement (497)	DIC-Net-d	0.0157	0.0169	0.0201	0.0217
	DisplacementNet	0.1191	0.1539	0.1395	0.1783
	StrainNet-f	0.0288	0.0295	0.0467	0.0458
Strain concentrations (418)	DIC-Net-d	0.0137	0.0122	0.0175	0.0157
	DisplacementNet	0.0410	0.0252	0.0529	0.0315
	StrainNet-f	0.0301	0.0280	0.0504	0.0457
Linear displacement (357)	DIC-Net-d	0.0120	0.0117	0.0151	0.0148
	DisplacementNet	0.0169	0.0125	0.0187	0.0144
	StrainNet-f	0.0254	0.0251	0.0435	0.0408
Quadratic displacement (714)	DIC-Net-d	0.0121	0.0117	0.0153	0.0149
	DisplacementNet	0.0143	0.0152	0.0165	0.0174
	StrainNet-f	0.0250	0.0243	0.0427	0.0405
Total (3216)	DIC-Net-d	0.0130	0.0126	0.0165	0.0161
	DisplacementNet	0.0356	0.0378	0.0417	0.0440
	StrainNet-f	0.0268	0.0263	0.0450	0.0427

• The performance of spatial resolution in Star5 image set

Network	SR	Noise level	𝛼
DIC-Net-d	17.25	0.0136	0.234
StrainNet-f	26.93	0.0163	0.439
DisplacemeNet	52.66	0.0041	0.216

Demo predicted by DIC-Net

Displacement prediction results compared with other two models

• Star5 displacement and noise level predicted by three networks.

• Tensile plate with hole predicted by three networks.

Strain prediction result

• Star6 strain predicted by DIC-Net-s

Our Environment

Matlab

• R2020b
• MATLAB Support for MinGW-w64 C/C++ Compiler

Python

• python 3.0

CUDA 11.5

Pytorch 1.10

Numpy 1.20.3

Pretrained model and Dataset

The Version used in the paper is:

DIC-Net-d

• Google Drive:https://drive.google.com/file/d/12bO8bZ3cX_3K3mPxer_IHWVJHx16mWJx/view?usp=share_link
• Baidu Netdisk：链接：https://pan.baidu.com/s/1bGbVPY6piE1XHQX7FPtR_A 提取码：7cqj

DIC-Net-s

• Google Drive:https://drive.google.com/file/d/19aZQrc-mnd5W8jYfn6DZmhDSZTKxdxaP/view?usp=share_link
• Baidu Netdisk：链接：https://pan.baidu.com/s/1GfvhWms_5mLhN7zP1q0AXg 提取码：br61

Hermite Dataset

• Google Drive:
• Baidu Netdisk：链接：https://pan.baidu.com/s/14KE8gOk81ETEEtKnFoc8Zw 提取码：90w1
• The Hermite dataset provided here has a total of 85,700 image pairs, the first 42,900 pairs are used in our paper, please adjust as needed.

Note: It is possible to obtain a more powerful DIC-Net by modifying the parameters of the Hermite element for retraining, so our model parameters would be continuously updated.Refer to https://github.com/YinWang20/DIC-Net-pretrained-Models

Citation

@article{WANG2023107278,
title = {DIC-Net: Upgrade the performance of traditional DIC with Hermite dataset and convolution neural network},
journal = {Optics and Lasers in Engineering},
volume = {160},
pages = {107278},
year = {2023},
issn = {0143-8166},
doi = {https://doi.org/10.1016/j.optlaseng.2022.107278},
url = {https://www.sciencedirect.com/science/article/pii/S0143816622003311},
author = {Yin Wang and Jiaqing Zhao}
}

We're well underway to get the code in order, if you need any help, don't hesitate to contact the corresponding author (jqzhao@mail.tsinghua.edu.cn) right away!