Element Defective Sample Augmentation Method Based on Improved DCGAN

The deep learning-based methods for inspecting defects of printing label are usually having problems such as difficult to collect training samples and insufficient training accuracy. In this paper, a defect sample augmentation method based on improved DCGAN network is proposed to achieve effective augmentation of a limited sample dataset, where spectral normalization and Wasserstein distance are introduced into the original DCGAN network. The generated defective sample by the original DCGAN and the improved network are compared, and the generalization of the proposed method is verified by applied to augment the sample images of different element in printing label. The influence of image resolution and defect size of the original samples on the network training are investigated respectively. The experimental results show that the improved DCGAN could effectively alleviate the phenomenon of pattern collapse and gradient vanishing that existed in the original network during the training process, and the morphology and location of defects in the generated defect samples are diversity and randomness.