Classification of strip steel surface defects based on data augmentation combined with MobileNet

Inspection of strip steel surface defects is one of the important steps in the strip production process and is closely related to the quality of the strip products. Since the strip travels at speeds of up to several hundred meters per minute on the work rolls, the efficiency of the inspection algorithm is very demanding. Lightweight deep learning networks with less number of parameters and lower computational complexity can meet practical needs in terms of efficiency. However, the relatively low accuracy of lightweight deep learning networks makes it difficult to meet the accuracy needs of enterprises. In this paper, we propose a new data augmentation method to improve the classification accuracy of MobileNet lightweight deep learning network. Experimental results on the X-SDD dataset show that the method in this paper improves the classification accuracy of MobileNet in the test set by 3.18%.