In-Service High-Voltage Cable Defect Detection Using Computed Tomography Based on Deep Learning

Fig. 4. Comparison of image preprocessing results in datasets (from left to right, methods are window width and level adjustment, cropping, and denoising in sequence)

Download full size

Fig. 5. Architecture of Cascade R-CNN

Download full size

Fig. 6. Improved CBAM structure diagram

Download full size

Fig. 7. Overall detection network structure

Download full size

Fig. 8. L-STCT defect images

Download full size

Fig. 9. Interference features in L-STCT images. (a) Metal artifact; (b) stripe artifact; (c) missing edge information; (d) precipitated crystal

Download full size

Fig. 10. Network detection results and their locally enlarged images

Download full size

Table 1. Parameters of STCT

View table

Table 1. Parameters of STCT

Parameter	Value
SOD /mm	208.3, 119, 160
DOD /mm	60, 54, 55
Number of detector arrays /pixel	1024×1024
Distance of source /mm	240
Voltage /kV	60, 0, 70
Current /μA	70, 70, 140
Pixel size /mm	0.127
Number of scanning points	1000, 1000, 500

Table 2. Training parameter list

View table

Table 2. Training parameter list

Parameter	Value
Epochs	200
Batch_size	1
Early_stop	10
Img_scale	[900, 900]
Cascade Head IoU	[0.3, 0.4, 0.5]
Anchor_scale	[16, 32, 64, 128, 256]
Anchor_ratio	[0.5, 1, 2]
Original learning rate	0.001
Momentum	0.9
Weight _decay	0.0001
Pre_model	ResNet50-FPN
Num_proposals	100

Table 3. Results of ablation experiments
View table
Table 3. Results of ablation experiments
Cascade CBAM DCNv2 EFPN F-EIoU AP^all AP^small AP^medium AR
√ 0.821 0.788 0.843 0.876
√ √ 0.853 0.811 0.878 0.903
√ √ √ 0.877 0.839 0.892 0.917
√ √ √ √ 0.878 0.851 0.885 0.919
√ √ √ √ √ 0.884 0.859 0.897 0.927

Table 4. Results of algorithm comparison experiments
View table
Table 4. Results of algorithm comparison experiments
Method AP AR FPS
Cascade R-CNN 0.821 0.876 13.986
Libra R-CNN 0.801 0.857 13.521
DETR 0.820 0.883 9.875
VIT 0.876 0.917 5.108
Our model 0.884 0.927 9.351

Table 5. Results of knowledge transfer experiments
View table
Table 5. Results of knowledge transfer experiments
Method Dataset AP AR
Cascade R-CNN L-STCT 0.821 0.876
Cascade R-CNN RCT 0.927 0.949
Our model L-STCT 0.884 0.927
Our model L-STCT+RCT 0.869 0.912
Our model PreRCT+L-STCT 0.901 0.959

Tools

Get Citation

Copy Citation Text

Chaoliang He, Ting Yan, Tianyu Ma, Xiaojiao Duan. In-Service High-Voltage Cable Defect Detection Using Computed Tomography Based on Deep Learning[J]. Acta Optica Sinica, 2025, 45(2): 0215001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites