Self-Supervised Augmentation of Quality Data Based on Classification-Reinforced GAN

EasyChair Preprint 9396

Abstract

In deep learning, the quality of ground truth training data is crucial for the resulting performance. However, depending on applications, collecting a sufficient amount of quality data from a realistic setting is problematic. In this case, data augmentation can play an important role as long as augmentation ensures data quality and diversity for training, preferably in an unsupervised way. Recently, a number of GAN variants have been emerged for improved quality in data augmentation. Although successful, further improvement is necessary for enhancing diversity in addition to quality in data augmentation. In this paper, we propose a GAN-based approach to self-supervised augmentation of quality data based on Classification-Reinforced GAN referred to here as CLS-R GAN, to extending diversity as well as quality in data augmentation. In CLS-R GAN, a discriminator-independent classifier additionally self-trains the generator by classifying the fake data, as well as augmenting the real data in an unsupervised way. Extensive experiments were conducted, including an application to augmenting liver ultrasonic image data, to verify the effectiveness of CLS-R GAN based on standard evaluation metrics. The results indicate the effectiveness of CLS-R GAN for improved quality and diversity in augmented data.

Keyphrases: GAN, Unsupervised Data Augmentation, generator, self-training

@booklet{EasyChair:9396,
  author    = {Seunghwan Kim and Sukhan Lee},
  title     = {Self-Supervised Augmentation of Quality Data Based on Classification-Reinforced GAN},
  howpublished = {EasyChair Preprint 9396},
  year      = {EasyChair, 2022}}