Strip Steel Defect Classification Using the Improved GAN and EfficientNet

In recent years, deep-learning detection algorithms based on automatic feature extraction have become the focus of defect detection. However, limited by industrial field conditions, the insufficient number of images in the collected dataset restricts the detection effect of deep learning. In this paper, an algorithm of strip steel defect classification using the improved GAN and EfficientNet was proposed. First, the label deconvolution network is constructed, where the image labels were deconvolved layer by layer to obtain the conditional masks that were superimposed into the generator and discriminator to form Mask-CGAN. Then, the mode-seeking generative adversarial networks (MSGAN) were improved and used to solve the problem of mode collapse. Finally, the EfficientNet was improved and trained on the dataset expanded by Mask-CGAN, which achieved the classification of strip steel defects. Experiments showed that Mask-CGAN proposed in this paper can generate true-to-life images and solve the problem of insufficient samples in deep learning. The improved EfficientNet with fewer parameters can accurately and efficiently classify strip steel defects.