Deep Adversarial Data Augmentation for Fabric Defect Classification With Scarce Defect Data

Fabric defect classification is a crucial and challenging task for fabric production quality guarantee. In recent years, many deep neural network-based methods have been proposed and shown promising performance on this task. However, it would be laborious and time-consuming to collect enough defect images to satisfy high-quality training because that defects are too rare in factories. In this paper, we propose a deep adversarial data augmentation method named DefectTransfer to address the defect data scarcity issue. Since the defect may happen anywhere on the background texture with any size, we consider the position and size of a defect should not be fully linked to the background texture in the network training. Based on this assumption, we design a cut-paste approach to augment the defect images by cutting out defects and pasting them on defect-free images. The defects are randomly transformed with scaling, rotating, and moving before the paste operation. To make the network training more efficient, we further propose an adversarial transformation algorithm that adjusts the pasted defects targeting the weakness of the classification network. The high diversity of the adversarial synthetic defect images forces the network to learn more discriminative category features. Experimental results show that our method can achieve comparable performance with recent fabric defect classification methods with only 1% fabric defect data on the ZJU-Leaper dataset. DefectTransfer also largely surpasses traditional augmentation methods even without manually annotated masks.