Generative Adversarial Networks in Computer Vision: A Review of Variants, Applications, Advantages, and Limitations

In the area of artificial intelligence, Generative Adversarial Networks (GANs) have developed as an innovative form of generative modelling. GANs have created immense attention since their first appearance in 2014, consequently contributing to improvements in several fields. GANs are based on the principle of zero-sum game theory, wherein the generator and discriminator neural networks compete to produce and analyse data. GANs are capable of learning data distribution in an unsupervised way. GANs just need unlabelled data to train instead of labelled data, which is required for normal supervised learning. Hence, GANs become highly helpful for applications where labelled data might be expensive or difficult to get. The architectural advancements and loss functions of various GAN model modifications are significant because they have improved training stability and produced more realistic and high-quality data. These developments contributed by GANs enabled their adoption in a wide range of applications in different domains such as video processing, image processing, language processing, vision computing. This review paper presents a comprehensive overview of GANs, with an emphasis on their fundamental methods, various types of GAN, numerous applications, advantages, and present challenges.