Image augmentation using GAN models in Computer Vision

Abstract

The efficiency of deep learning algorithms will increase when it is trained on a large size of data. Over fitting problems will also be solved working on a large dataset. To collect a huge quantity of data for training a model is more challenging job. Collecting data will take more time as well as resources. The data augmentation technique will increase the diversity of data that to be trained on deep learning algorithms. This will also help in not collecting new data. This led to the need for generative models. When the set of training data is given to this generative model, it will learn to produce the same statistical data as the set of training data. GANs are used in many fields including computer vision, medical, agriculture etc. In this paper along with the traditional GANs two architecture variant and two loss-variant GANs are reviewed and experiments are conducted to generate images. Detailed reviews on the performance metrics of GAN models are also discussed and observed that FID plot doesn’t give much insight to compare the generated images. Despite of the significant success achieved till date, applying GAN model to real time dataset has some challenging problems. Challenges faced while training GANs including mode collapse, image quality and vanishing gradients are discussed.