Recognition Algorithm of Strip Steel Surface Defects Based on Attention Model

Yanuo Lu; Bingcai Chen; Degang Chen; Shixiang Yan; Shunping Li

doi:10.3788/LOP202158.1410014

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 14 >Page 1410014 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 14, 1410014 (2021)

Recognition Algorithm of Strip Steel Surface Defects Based on Attention Model

Yanuo Lu¹, Bingcai Chen^1、2、*, Degang Chen¹, Shixiang Yan¹, and Shunping Li¹

Author Affiliations

¹School of Computer Science and Technology, Xinjiang Normal University, Urumqi, Xinjiang 830054, China

²College of Computer Science and Technology, Dalian University of Technology, Dalian, Liaoning 116024, China

show less

DOI: 10.3788/LOP202158.1410014 Cite this Article Set citation alerts

Yanuo Lu, Bingcai Chen, Degang Chen, Shixiang Yan, Shunping Li. Recognition Algorithm of Strip Steel Surface Defects Based on Attention Model[J]. Laser & Optoelectronics Progress, 2021, 58(14): 1410014 Copy Citation Text

show less

Fig. 1. Deep residual network frame for strip steel surface defect recognition based on attention mechanism

Download full size

Fig. 2. Strip surface defect dataset. (a) Cr; (b) In; (c) Pa; (d) Ps; (e) Rs; (f) Sc

Download full size

Fig. 3. Data preprocessing

Download full size

Fig. 4. False color enhancement effect. (a) Before enhancement; (b) after enhancement

Download full size

Fig. 5. Softmax Loss curve

Download full size

Fig. 6. Training curves of model under different learning rates. (a) Accuracy curves; (b) loss function curves

Download full size

Fig. 7. Pictures with different signal-to-noise ratios. (a) No noise; (b) 50 dB; (c) 40 dB; (d) 30 dB; (e) 20 dB

Download full size

Fig. 8. Confusion matrix of each model. (a) ResNet50; (b) A-ResNet50

Download full size

Type of defects	Cr	In	Pa	Ps	Rs	Sc
Before expansion	300	300	300	300	300	300
After expansion	900	900	900	900	900	900

Table 1. Distribution of strip surface defect datasets

Learning rate	Recognition accuracy /%						Training accuracy /%	Training loss value	Average accuracy /%
Learning rate	Cr	In	Pa	Ps	Rs	Sc	Training accuracy /%	Training loss value	Average accuracy /%
0.01	100	95.00	100	85.00	100	98.33	99.24	0.0986	96.32
0.001	100	98.33	100	96.67	98.33	98.33	99.31	0.0205	98.61
0.0001	90.00	98.33	100	88.33	100	98.33	95.58	0.1154	95.83
0.00001	98.33	96.67	100	78.33	100	96.67	81.74	0.5037	95.00

Table 2. Recognition accuracy of each defect under different learning rates

Model	Recognition accuracy /%
Model	20 dB	30 dB	40 dB	50 dB
LBP^[7]	22.78	63.57	69.26	75.05
MB-LBP^[7]	25.76	71.35	90.24	91.83
Improved MB-LBP^[7]	30.63	78.56	95.66	97.00
ResNet50	21.65	75.56	92.22	95.56
A-ResNet50	34.44	88.33	94.44	98.61

Table 3. Anti-noise ability of each model under Gaussian noise with different signal-to-noise ratios

Defectcategory	Number of samples						P_pr	Recallrate	F₁
Defectcategory	Cr	In	Pa	Ps	Rs	Sc	P_pr	Recallrate	F₁
Cr	60	0	0	0	0	0	1.000	1.000	1.000
In	0	59	0	0	1	1	0.967	0.983	0.975
Pa	0	0	60	2	0	0	0.968	1.000	0.984
Ps	0	1	0	58	0	0	0.983	0.967	0.975
Rs	0	0	0	0	59	0	1.000	0.983	0.992
Sc	0	0	0	0	0	59	1.000	0.983	0.992

Table 4. Recognition ability of proposed model A-ResNet50

Model	Recognition accuracy /%						Average accuracy /%	Unit inference time /s
Model	Cr	In	Pa	Ps	Rs	Sc	Average accuracy /%	Unit inference time /s
Model in Ref.[5]	97.00	94.00	98.00	98.00	97.00	93.00	96.17	0.065
Model in Ref.[7]	99.00	98.00	97.00	96.00	97.00	95.00	97.00	0.058
AlexNet	100	98.33	100	75.00	100	98.33	95.28	0.016
A-AlexNet	100	95.00	100	85.00	100	98.33	96.39	0.017
GoogleNet	81.67	100	91.67	81.67	90.00	98.33	90.56	0.389
A-GoogleNet	93.33	100	96.67	90.00	90.00	100	95.00	0.041
ResNet50	93.33	95.00	100	85.00	100	100	95.56	0.074
A-ResNet50	100	98.33	100	96.67	98.33	98.33	98.61	0.078
ResNet101	100	98.33	100	88.33	98.33	100	97.50	0.121
A-ResNet101	100	96.67	100	91.67	100	100	98.05	0.130

Table 5. Recognition results of strip steel surface defects by different models

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information