DIMENSIONALITY REDUCTION BASED CLASSIFICATION USING GENERATIVE ADVERSARIAL NETWORKS DATASET GENERATION

Abstract

The term data augmentation refers to an approach that can be used to prevent overfitting in the training dataset, which is where the issue first manifests itself. This is based on the assumption that extra datasets can be improved by include new information that is of use. It is feasible to create an artificially larger training dataset by utilizing methods such as data warping and oversampling. This will allow for the creation of more accurate models. This idea is demonstrated through the application of a variety of different methods, some of which include neural style transfer, adversarial training, and erasure by random erasure, amongst others. By utilizing oversampling augmentations, it is feasible to create synthetic instances that can be incorporated into the training data. This is made possible by the generation of synthetic instances. There are numerous illustrations of this, including image merging, feature space enhancements, and generative adversarial networks, to name a few (GANs). In this paper, we aim to provide evidence that a Generative Adversarial Network can be used to convert regular images into Hyper Spectral Images (HSI). The purpose of the model is to generate data by including a certain amount of unpredictable noise.