Table 10.

Classification results of the CNN model for MO images of weld defects.

Defect Types	Number of Images	Train Samples	Valid Samples	Test Samples	Recognition Result	Classification Accuracy/%
Crack	1000	800	100	100	97	97
Pit	1000	800	100	100	97	97
Lack of penetration	1000	800	100	100	100	100
Gas pore	1000	800	100	100	94	94
No defects	1000	800	100	100	98	98
Total	5000	4000	500	500	486	97.2