Large Deformation Measurement Method of Speckle Images Based on Deep Learning

Hong Xiao; Chengnan Li; Mingchi Feng

doi:10.3788/AOS222084

Acta Optica Sinica, Volume. 43, Issue 14, 1412001(2023)

Large Deformation Measurement Method of Speckle Images Based on Deep Learning

Hong Xiao, Chengnan Li, and Mingchi Feng^*

School of Advanced Manufacturing Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(12)

Fig. 1. Four kinds of typical speckle images. (a) Speckle image Ⅰ; (b) speckle image Ⅱ; (c) speckle image Ⅲ; (d) speckle image Ⅳ

Download full size

Fig. 2. Partial samples of speckle displacement field image dataset

Download full size

Fig. 3. Convolutional block attention module. (a) Channel attention module; (b) sapatial attention module; (c) convolutional block attention module

Download full size

Fig. 4. Two different convolution modules. (a) General convolution; (b) depthwise separable convolution

Download full size

Fig. 5. Structure diagram of DICNet

Download full size

Fig. 6. Loss decline curves of FlowNetS, FlowNetC, and UNet during training. (a) Loss of training set; (b) loss of validation set

Download full size

Fig. 7. Performance of algorithms and models on test set. (a) Sample 1; (b) sample 2; (c) sample 3; (d) sample 4

Download full size

Fig. 8. Performance of algorithms and models on DIC challenge dataset. (a) Sample 8; (b) sample 10; (c) sample 12; (d) sample 17

Download full size

Table 1. Generation parameters of classⅡ and Ⅳ speckle images
View table
Table 1. Generation parameters of classⅡ and Ⅳ speckle images
Class R_speckle /pixel N_speckle Γ T_speckle N_image
Ⅱ 0.5 150000 0.6 E 200
0.7 150000 0.5 E 200
1 10000 0.5 P 200
3 6000 0.6 P 200
Ⅳ 0.7 150000 0.5 U 100
1.5 50000 0.6 L 100
1 50000 0.7 E 100
2 20000 0.8 P 100

Table 2. Model information and training results of FlowNetS, FlowNetC, and UNet
View table
Table 2. Model information and training results of FlowNetS, FlowNetC, and UNet
Model N_params /M T_train /min $L_{M S E}^{t r}$ /pixel² $L_{M S E}^{v a l}$ /pixel² T_inference /ms
FlowNetC 39.175 147 0.358 0.477 8.369
FlowNetS 38.675 80 0.426 0.293 5.272
UNet 7.760 99 0.056 0.073 5.532

Table 3. Performance of DICNet on self-built training set and validation set

View table

Table 3. Performance of DICNet on self-built training set and validation set

Algorithm and model	RMSE on training set $L_{R M S E}^{t r}$ /pixel				RMSE on validation set $L_{R M S E}^{v a l}$ /pixel				Mean time T_inference /ms
Algorithm and model	Mean	Maximum	Minimum	Standard deviation	Mean	Maximum	Minimum	Standard deviation	Mean time T_inference /ms
Ncorr	0.512	1.309	0.056	0.241	0.519	1.245	0.046	0.239	8089.898
OpenCorr-CPU	0.307	2.897	0.004	0.193	0.317	1.619	0.036	0.196	16492.512
OpenCorr-GPU	0.270	2.877	0.003	0.171	0.275	1.276	0.013	0.163	2383.299
DeepDIC	0.174	0.325	0.097	0.019	0.174	0.363	0.089	0.019	15.800
StrainNet	0.187	0.784	0.055	0.071	0.183	0.746	0.054	0.066	3.727
UNet	0.092	0.259	0.049	0.026	0.091	0.241	0.047	0.026	4.015
DICNet	0.056	0.209	0.032	0.019	0.055	0.189	0.034	0.018	4.919

Table 4. Performance of DICNet on self-built test set

View table

Table 4. Performance of DICNet on self-built test set

Algorithm and model	RMSE on test set $L_{R M S E}^{t e}$ /pixel				Number of model parameter N_params /M
Algorithm and model	Mean	Maximum	Minimum	Standard deviation	Number of model parameter N_params /M
Ncorr	0.509	1.127	0.071	0.223	—
OpenCorr-CPU	0.165	0.596	0.029	0.091	—
OpenCorr-GPU	0.176	0.732	0.026	0.109	—
DeepDIC	0.183	0.250	0.140	0.019	49.970
StrainNet	0.206	0.557	0.067	0.072	38.663
UNet	0.174	0.413	0.070	0.050	7.760
DICNet	0.158	0.267	0.078	0.039	6.398

Tools

Get Citation

Copy Citation Text

Hong Xiao, Chengnan Li, Mingchi Feng. Large Deformation Measurement Method of Speckle Images Based on Deep Learning[J]. Acta Optica Sinica, 2023, 43(14): 1412001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites