A transfer learning strip steel surface defect recognition network based on VGG19

The types of surface defects of strip steel are various and complex gray gradation structure. The existing image detection technology based on machine vision still has the disadvantages of low recognition efficiency and poor generalization performance in strip steel defect detection. However, image detection technology based on deep learning need large numbers of image data to train networks. For a typical multi-class and small sample data with low quality pixels, these data cannot complete a deep neural network training. For this type of data, traditional convolutional neural networks have low recognition rate for small samples and poor generalization for large samples. Combining with deep learning and transfer learning, this paper proposes a transfer learning strip steel defect recognition network based on VGG19. The frozen pre-training network layers in VGG19 are not trained, the learning rates are setting in combination with the actual use of the network layers. The convergence speed and accuracy of the model are taken into account, and the recognition rate and generalization force on small sample data are greatly improved. On the NEU surface dataset 2, the recognition accuracy of our model is 97.5%, which is much higher than the traditional machine learning algorithm. Moreover, the network model in this paper does not require data preprocessing and model parameter adjustment, nor does it need to manually participate in designing the classifier. It is a simple and effective method for identifying the surface defects of strip steel. The method of this paper has a certain practical value in the field of surface recognition of other products.