Defect Transfer GAN: Diverse Defect Synthesis for Data AugmentationDownload PDF

Published: 28 Jan 2022, Last Modified: 13 Feb 2023ICLR 2022 SubmittedReaders: Everyone
Keywords: Defect synthesis, Generative Adversarial Networks, Content transfer, Automated visual inspection, Data augmentation
Abstract: Large amounts of data are a common requirement for many deep learning approaches. However, data is not always equally available at large scale for all classes. For example, on highly optimized production lines, defective samples are hardly acquired while non-defective samples come almost for free. The defects however often seem to resemble each other, e.g., scratches on different products may only differ in few characteristics. In this work, we propose to make use of the shared characteristics by transferring a stylized defect-specific content from one type of background product to another. Moreover, the stochastic variations of the shared characteristics are captured, which also allows generating novel defects from random noise. These synthetic defective samples enlarge the dataset and increase the diversity of defects on the target product. Experiments demonstrate that our model is able to disentangle the defect-specific content from the background of an image without pixel-level labels. We present convincing results on images from real industrial production lines. Furthermore, we show consistent gains of using our method to enlarge training sets in classification tasks.
26 Replies

Loading