Weld Feature Extraction Based on Fully Convolutional Networks

Fig. 6. Extraction results obtained with three different algorithms. (a1)-(a4) Weld original images; (b1)-(b4) welds extracted with gray gravity algorithm; (c1)-(c4) welds extracted with Steger algorithm; (d1)-(d4) welds extracted with FCN-8s

Download full size

Fig. 7. Error curves of weld position extraction

Download full size

Table 1. Parameters of network structure

View table

Table 1. Parameters of network structure

Layer	Kernel (width×height×numbers)	Stride	Pad
Conv1	3×3×64,3×3×64	1,1	1,1
Pool1	2×2	2	0
Conv2	3×3×128,3×3×128	1,1	1,1
Pool2	2×2	2	0
Conv3	3×3×256,3×3×256,3×3×256	1,1,1	1,1,1
Pool3	2×2	2	0
Conv4	3×3×512,3×3×512,3×3×512	1,1,1	1,1,1
Pool4	2×2	2	0
Conv5	3×3×512,3×3×512,3×3×512	1,1,1	1,1,1
Pool5	2×2	2	0
Conv6-Conv8	1×1×4096,1×1×4096,1×1×2	1,1,1	0,0,0
Deconv1-Conv3	32×32×2,16×16×2,8×8×2	32,16,8	0,0,0
Deconv4-Deconv5	4×4×2,4×4×2	2,2	0,0

Table 2. Training results of three upsampling structures
View table
Table 2. Training results of three upsampling structures
Network model PACC /% MACC /% MIU /%
FCN-32s 99.2 89.9 81.3
FCN-16s 99.4 94.3 86.8
FCN-8s 99.5 97.4 87.7

Table 3. RMSE of weld position between results and label images pixel
View table
Table 3. RMSE of weld position between results and label images pixel
Algorithm RMSE forFig.6(a1) RMSE forFig.6(a2) RMSE forFig.6(a3) RMSE forFig.6(a4)
FCN-8s 4.15 4.78 5.85 3.87
Stegeralgorithm 13.78 30.56 56.61 44.42
Gray gravityalgorithm 13.53 35.01 60.18 47.96

Tools

Get Citation

Copy Citation Text

Yongshuai Zhang, Guowei Yang, Qiqi Wang, Lei Ma, Yizhong Wang. Weld Feature Extraction Based on Fully Convolutional Networks[J]. Chinese Journal of Lasers, 2019, 46(3): 0302002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites