Non-adversarial Image Generative Algorithm

Non-adversarial Image Generative Algorithm based on Random Convex Combination in Clustered Features

The present work aims to offer a methodology for generating new images from a given dataset without training any Generative Adversarial Network (GAN). The idea is to extract the dataset's features using the Principal Component Analysis (PCA) algorithm looking for a linear transformation between the image pixel space and the features space, perform dimensionality reduction, and cluster the resulting set. Then, new feature vectors are generated through random convex combinations, which are subsequently mapped back to the image pixel space by applying the inverse of the corresponding transformation.