Augmenting image data using generative adversarial networks (GAN)

• Main Author: Liu, Xinchi
• Other Authors: Wang Lipo
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Generative adversarial networks
• Online Access: https://hdl.handle.net/10356/173947
• Institution: Nanyang Technological University

Deep learning is proposed to employ algorithms to replace laborious human operations with more automation, improving system performance on specific tasks. The researchers have become interested in machine learning in the past decades. There are many applications for machine learning in various fields. Data augmentation is designed to create the data and improve the deep learning model accuracy when data is rare. The generative adversarial network is a common approach to augment the datasets. In this report, the principles and architectures of several GAN methods including DCGAN, conditional GAN and cycle GAN are introduced. The functions and relationship of the generators and discriminators are discussed. By operating the simulation in Jupiter notebook with same dataset, the influence of epoch on the performance for the same GAN method is explored. The different simulation results using the DCGAN and CGAN are compared in order to analyze the difference between them. In addition, the future development of GAN is discussed. The test accuracies comparisons between original data and augmented data are done to analyze the function of data augmentation.