Weld Seam Feature Extraction Based on Improved U-Net Convolutional Neural Network

Leilei Xiong; Xuejun Zhu; Huige Lai; Checao Yu; Kun Mao; Ming Yang; Da Peng

doi:10.3788/LOP250654

Laser & Optoelectronics Progress, Volume. 62, Issue 16, 1622002(2025)

Weld Seam Feature Extraction Based on Improved U-Net Convolutional Neural Network

Leilei Xiong, Xuejun Zhu^*, Huige Lai, Checao Yu, Kun Mao, Ming Yang, and Da Peng

Author Affiliations

School of Mechanical Engineering, Ningxia University, Yinchuan 750021, Ningxia , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(17)

Fig. 1. Data collection platform. (a) FV-160 weld seam tracking sensor; (b) data collection diagram

Download full size

Fig. 2. Partial dataset

Download full size

Fig. 3. Improved U-Net model structure

Download full size

Fig. 4. Transfer learning training process

Download full size

Fig. 5. U-Net model structure diagram with VGG16

Download full size

Fig. 6. Structure diagram of GGCA mechanism

Download full size

Fig. 7. GSConv structure diagram

Download full size

Fig. 8. Upsampling network structure. (a) GS bottleneck; (b) VoVGSCSP; (c) improved VoVGSCSP

Download full size

Fig. 9. Impact of transfer learning on model training loss

Download full size

Fig. 10. Comparison of segmentation results using different algorithms

Download full size

Fig. 11. Extraction result of laser stripe centerline

Download full size

Fig. 12. Error results of laser stripe centerline extraction. (a) U-Net; (b) Deeplabv3+; (c) PSP-Net; (d) proposed

Download full size

Table 1. Comparison of model performance under different weight coefficients of loss functions
View table
Table 1. Comparison of model performance under different weight coefficients of loss functions
$ω$ mPA /% mIoU /%
0 95.67 89.76
0.2 95.78 90.11
0.4 95.83 90.23
0.6 95.76 90.04
0.8 95.69 89.92
1.0 95.53 89.66

Table 2. Comparison of model performance under different group numbers
View table
Table 2. Comparison of model performance under different group numbers
G mPA /% mIoU /%
1 96.48 91.89
2 96.56 92.09
4 96.51 91.92

Table 3. Comparison of ablation experiment results (segmentation accuracy)
View table
Table 3. Comparison of ablation experiment results (segmentation accuracy)
Model VGG16 VoVGSCSP Pretrain GGCA mIoU /% mPA /% ACC /%
Baseline × × × × 89.76 95.67 99.46
1 √ × × × 90.80 95.90 99.54
2 √ √ × × 90.91 96.07 99.55
3 √ √ √ × 91.78 96.45 99.59
4 √ √ √ √ 92.09 96.56 99.61

Table 4. Comparison of ablation experiment results (model size)
View table
Table 4. Comparison of ablation experiment results (model size)
Model VGG16 VoVGSCSP GGCA Params /10⁶ FLOPs /10⁹ FPS
Baseline × × × 34.527 131.045 34.95
1 √ × × 24.891 112.918 39.21
2 √ √ × 16.306 68.693 36.89
3 √ √ √ 16.351 68.741 35.21

Table 5. Comparison of evaluation indicators of the proposed algorithm and other algorithms
View table
Table 5. Comparison of evaluation indicators of the proposed algorithm and other algorithms
Algorithm ACC /% mPA /% mIoU /%
U-Net 99.46 95.67 89.76
Deeplabv3+ 99.51 95.86 90.19
PSP-Net 99.34 94.93 87.75
Proposed 99.61 96.56 92.09

Tools

Get Citation

Copy Citation Text

Leilei Xiong, Xuejun Zhu, Huige Lai, Checao Yu, Kun Mao, Ming Yang, Da Peng. Weld Seam Feature Extraction Based on Improved U-Net Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2025, 62(16): 1622002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites