Abstract
Deep learning enables a computer to construct complex concepts from simpler ones by learning data representation with abstraction. However, deep learning requires a large number of samples for training to prevent over-fitting. Generative adversarial network (GAN) is a kind of unsupervised learning that can be used to synthesize images. Generating higher quality images using GAN is conducive to improving the training efficiency of deep learning. How to generate high quality and high-diversity samples using GAN to expand the sample size for deep learning training is quite appealing. In this paper, the concept of generative adversarial network including network structure and four popular loss functions is introduced. Furthermore, three typical GAN-based models to improve the image quality or diversity have been presented. Finally, the paper discusses the shortcomings of GANs to steer future research in the right direction. Generation of high quality images using GAN has great potential to help deep learning models to reduce the over-fitting problem in the near future.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have