Generative Adversarial Network (GANs) for Image Generation or Data Augmentation
Ramandeep Kaur1*, Baljinder Kaur2, Amrik Singh2
Abstract
In recent years, image segmentation has become a crucial aspect in various fields, ranging from disease diagnosis to autonomous vehicle navigation. In computer vision, image segmentation is a vital task and is more complex than other vision problems due to its reliance on low-level spatial data. Deep Learning has significantly influenced the field of segmentation, resulting in several successful models today. Among these, Generative Adversarial Networks (GANs) have shown remarkable performance in image segmentation. This study provides a systematic review of recent publications on GAN models and their applications. The authors conducted a search across three databases—Embase (Scopus), WoS, and PubMed—to find relevant papers, resulting in 2,084 documents. After a two-phase screening process, 52 records were selected for the final review. Key applications of GANs identified include 3D object generation, medicine, pandemics, image processing, face detection, texture transfer, and traffic control. Prior to 2016, research in this area was limited, but its practical application has grown significantly worldwide since then. The study also highlights the challenges associated with GANs and suggests directions for future research in this field. In this paper, we introduce an adversarial approach for abstractive text summarization, where we concurrently train a generative model (G) and a discriminative model (D). Specifically, the generator G is designed as a reinforcement learning agent that takes raw text as input and produces an abstractive summary. The discriminator, on the other hand, attempts to differentiate the generated summary from the ground truth summary. Extensive experiments show that our model achieves competitive ROUGE scores compared to state-of-the-art methods on the CNN/Daily Mail dataset. Additionally, qualitative results demonstrate that our model generates summaries that are more abstractive, readable, and diverse. Generative Adversarial Networks (GANs) are a type of artificial intelligence algorithm developed to address the generative modelling problem. The objective of a generative model is to analyse a set of training examples and learn the underlying probability distribution that generated them. GANs then use this estimated distribution to generate new examples. While deep learning-based generative models are common, GANs stand out as one of the most successful, particularly due to their ability to produce realistic high-resolution images. GANs have been applied successfully to a wide range of tasks, primarily in research contexts. However, they continue to present unique challenges and research opportunities, as they are grounded in game theory, unlike most other generative modelling approaches that rely on optimization techniques. Deep learning has made significant advancements in the field of artificial intelligence, leading to the development of various deep learning models. One such model, Generative Adversarial Networks (GANs), was introduced based on zero-sum game theory and has quickly become a prominent research focus. The importance of GANs lies in their ability to learn data distributions through unsupervised learning and generate more realistic, authentic data. Due to their vast application potential, including in image and vision computing, as well as video and language processing, GANs have been widely studied. This paper provides an overview of GANs, including their theoretical foundations and various extensions that enhance or modify the original model’s structure. The paper also explores typical applications of GANs, outlines existing challenges, and discusses future directions for GAN model development.
Keywords:
GANs, Image Generation, Data Augmentation, Deep Learning, Systematic Review
