Data Augmentation Techniques For Expanding The Dataset In The Task Of Image Processing

In this paper we've analyzed and elaborated the main problem within the field of Machine Learning (ML) which is the lack of data used for model training also referred to as data augmentation and how your neural network model generalizes data and makes sure it provides high-quality results. Throughout the paper it is mentioned and analyzed the various methods used during data augmentation in the field of image classification ranging from zooming, rotating, cropping and then presenting an approach that generates prime quality images as a combination of basic image and other image style transfer. Initially, a graph is presented that describes signs of overfitting and desired convergence, showing a model with the desired relationship between training and testing error. Then, images that show the different possibilities of stylizing a single image. and these images that are created are used for the training of the given neural network in order to improve the training process. The example of a drone is taken and the images transferred from the original one are presented. Then, it is shown why one of the techniques is more suitable than the others, how it works and what are its features, all the while representing the benefits and downsides of these methods being analyzed.