Deep Transfer Learning-Based Pulsed Eddy Current Thermography for Crack Defect Detection

Baiqiao Hao; Yugang Fan; Zhihuan Song

doi:10.3788/AOS221532

Acta Optica Sinica, Volume. 43, Issue 4, 0415002(2023)

Deep Transfer Learning-Based Pulsed Eddy Current Thermography for Crack Defect Detection

Baiqiao Hao^1,2、affaff, Yugang Fan^1,2、aff*, and Zhihuan Song^3、aff

¹Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming 650500, Yunnan, China

²Yunnan Key Laboratory of Artificial Intelligence, Kunming University of Science and Technology, Kunming 650500, Yunnan, China

³College of Control Science and Engineering, Zhejiang University, Hangzhou 310027, Zhejiang, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(12)

Fig. 1. Illustration of ASFF approach

Download full size

Fig. 2. Network structure of ASFF-YOLO v5

Download full size

Fig. 3. Construction process of defect detection model based on deep transfer learning

Download full size

Fig. 4. Pulsed eddy current excitation device

Download full size

Fig. 5. Number of defect images in each temperature range

Download full size

Fig. 6. Comparison graph of loss function

Download full size

Fig. 7. Comparison of training results of mAP and loss. (a) mAP; (b) loss

Download full size

Fig. 8. Comparison of detection results. (a) Original images; (b) results detected by YOLO v5; (c) results detected by ASFF-YOLO v5

Download full size

Table 1. Crack defect parameters and types
View table
Table 1. Crack defect parameters and types
Crack type Crack parameter
Length /mm Width /mm Depth /mm
Crack 1 4 1.5 1，2，3，4
Crack 2 8 1，2，3，4
Crack 3 20 1，2，3，4

Table 2. Hyperparameters of model
View table
Table 2. Hyperparameters of model
Hyperparameter Value
Epoch 200
Batch size 8
Learning rate 0.01
Momentum for SGD 0.937
Weight decay 0.0005

Table 3. Comparison of model evaluation indicators
View table
Table 3. Comparison of model evaluation indicators
Method P /% R /% η_mAP /%
M_FPS /
（frame·s^-1）
YOLO v5 97.9 69.6 83.5 60
Transfer-YOLO v5 98.2 86.1 92.4 58

Table 4. Comparison of model evaluation indicators
View table
Table 4. Comparison of model evaluation indicators
Model P /% R /% η_mAP /% M_FPS /（frame·s^-1）
Transfer+YOLO v5 98.2 86.1 92.4 58
Transfer+ASFF-YOLO v5 98.4 97.8 98.6 46

Tools

Get Citation

Copy Citation Text

Baiqiao Hao, Yugang Fan, Zhihuan Song. Deep Transfer Learning-Based Pulsed Eddy Current Thermography for Crack Defect Detection[J]. Acta Optica Sinica, 2023, 43(4): 0415002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites