Weld Structured Light Image Segmentation Based on Lightweight DeepLab v3+ Network

Bing Chen; Sheng He; Jian Liu; Shengfeng Chen; Enhui Lu

doi:10.3788/CJL221398

Chinese Journal of Lasers, Volume. 50, Issue 8, 0802105(2023)

Weld Structured Light Image Segmentation Based on Lightweight DeepLab v3+ Network

Bing Chen¹, Sheng He¹, Jian Liu^1、*, Shengfeng Chen^1、**, and Enhui Lu²

¹State Key Laboratory of Advanced Design and Manufacture for Vehicle Body, Hunan University, Changsha 410082,Hunan , China

²School of Mechanical Engineering, Yangzhou University, Yangzhou 225009, Jiangsu , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(12)

Fig. 1. Architecture of DeepLab v3+ semantic segmentation network^[25]

Download full size

Fig. 2. Working process of normal convolution and atrous convolution. (a) Normal convolution; (b) atrous convolution with dilation rate of 2

Download full size

Fig. 3. Architecture of lightweight DeepLab v3+ semantic segmentation network

Download full size

Fig. 4. Structured light image segmentation results. (a) Target result; (b) segmentation result 1: false negative; (c) segmentation result 2: false positive

Download full size

Fig. 5. Pixel ratio of structured light stripes and image background in structured light images

Download full size

Fig. 6. Structured light images and image labels. (a₁) Offline fillet weld image; (a₂) offline fillet weld label; (b₁) online fillet weld image; (b₂) online fillet weld label; (c₁) offline groove weld image; (c₂) offline groove weld label; (d₁) online groove weld image; (d₂) online groove weld label

Download full size

Fig. 7. Segmentation results of structured light images from test set by lightweight DeepLab v3+ semantic segmentation network. (a₁) Offline fillet weld image; (a₂) segmentation result of offline fillet weld image; (b₁) online fillet weld image; (b₂) segmentation result of online fillet weld image; (c₁) offline groove weld image; (c₂) segmentation result of offline groove weld image; (d₁) online groove weld image; (d₂) segmentation result of online groove weld image

Download full size

Fig. 8. Segmentation results of structured light images from test set by different networks. (a₁)-(a₄) Original images; (b₁)-(b₄) segmentation results of proposed network; (c₁)-(c₄) segmentation results of U-net network; (d₁)-(d₄) segmentation results of Segnet network; (e₁)-(e₄) segmentation results of FCN network

Download full size

Table 1. Detailed composition of structured light image dataset
View table
Table 1. Detailed composition of structured light image dataset
Weld type Quantity of offline image Quantity of online image Total quantity
Fillet weld 300 600 900
Groove weld 300 600 900

Table 2. Performance of DeepLab v3+ semantic segmentation network in test set structured light image segmentation with different backbone networks

View table

Table 2. Performance of DeepLab v3+ semantic segmentation network in test set structured light image segmentation with different backbone networks

Backbone network	Pixel accuracy		IOU		BF score B_L /%	Average score /%	Time /ms
Backbone network	A_L /%	A_I /%	IOU_L /%	MIOU /%	BF score B_L /%	Average score /%	Time /ms
Resnet-18	94.53	99.80	89.31	94.55	99.93	95.62	14.8
Resnet-50	95.17	99.80	89.40	94.59	99.95	95.78	23.4
Mobilenet-v2	94.90	99.79	89.00	94.39	99.86	95.59	18.1
Xception	94.44	99.78	88.78	94.28	99.86	95.43	30.2

Table 3. Performance of DeepLab v3+ semantic segmentation network in test set structured light image segmentation with different ratios of coefficients

View table

Table 3. Performance of DeepLab v3+ semantic segmentation network in test set structured light image segmentation with different ratios of coefficients

α₁/α₀	Pixel accuracy		IOU		BF score B_L /%	Average score /%
α₁/α₀	A_L /%	A_I /%	IOU_L /%	MIOU /%	BF score B_L /%	Average score /%
Original loss	94.53	99.80	89.31	94.55	99.93	95.62
1/1	99.03	99.67	84.33	91.99	99.85	94.97
1/3	98.37	99.72	86.39	93.05	99.89	95.49
1/5	98.35	99.73	86.61	93.17	99.91	95.56
1/10	98.07	99.74	87.13	93.43	99.93	95.66
1/15	96.47	99.78	89.04	94.41	99.93	95.93
1/20	95.92	99.79	89.24	94.51	99.91	95.87
1/30	95.11	99.80	89.65	94.73	99.92	95.84

Table 4. Performance of different networks in test set structured light image segmentation

View table

Table 4. Performance of different networks in test set structured light image segmentation

Network	Pixel accuracy		IOU		BF score B_L /%	Average score /%	Time /ms
Network	A_L /%	A_I /%	IOU_L /%	MIOU /%	BF score B_L /%	Average score /%	Time /ms
FCN	90.04	99.63	81.19	90.41	99.13	92.08	50.7
Segnet	91.49	99.64	81.92	90.78	99.16	92.60	82.2
U-net	92.92	99.72	85.54	92.63	97.81	93.72	51.4
DeepLab v3+	93.68	99.78	88.67	94.23	99.86	95.24	33.2
Ours	96.47	99.78	89.04	94.41	99.93	95.93	15.9

Tools

Get Citation

Copy Citation Text

Bing Chen, Sheng He, Jian Liu, Shengfeng Chen, Enhui Lu. Weld Structured Light Image Segmentation Based on Lightweight DeepLab v3+ Network[J]. Chinese Journal of Lasers, 2023, 50(8): 0802105

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Laser Forming Manufacturing

Received: Nov. 7, 2022

Accepted: Jan. 12, 2023

Published Online: Mar. 28, 2023

The Author Email: Liu Jian (liujian@hnu.edu.cn), Chen Shengfeng (shengfengc@hnu.edu.cn)

DOI:10.3788/CJL221398

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology

Table 1. Detailed composition of structured light image dataset

Table 1. Detailed composition of structured light image dataset

Table 2. Performance of DeepLab v3+ semantic segmentation network in test set structured light image segmentation with different backbone networks

Table 2. Performance of DeepLab v3+ semantic segmentation network in test set structured light image segmentation with different backbone networks

Table 3. Performance of DeepLab v3+ semantic segmentation network in test set structured light image segmentation with different ratios of coefficients

Table 3. Performance of DeepLab v3+ semantic segmentation network in test set structured light image segmentation with different ratios of coefficients

Table 4. Performance of different networks in test set structured light image segmentation

Table 4. Performance of different networks in test set structured light image segmentation