Image Diversification via Deep Learning based Generative Models

Abstract

Machine learning driven pattern recognition from imagery such as object detection has been prevalenting among society due to the high demand for autonomy and the recent remarkable advances in such technology. The machine learning technologies acquire the abstraction of the existing data and enable inference of the pattern of the future inputs. However, such technologies require a sheer amount of images as a training dataset which well covers the distribution of the future inputs in order to predict the proper patterns whereas it is impracticable to prepare enough variety of images in many cases.
To address this problem, this thesis pursues to discover the method to diversify image datasets for fully enabling the capability of machine learning driven applications.
Focusing on the plausible image synthesis ability of generative models, we investigate a number of approaches to expand the variety of the output images using image-to-image translation, mixup and diffusion models along with the technique to enable a computation and training dataset efficient diffusion approach. First, we propose the combined use of unpaired image-to-image translation and mixup for data augmentation on limited non-visible imagery. Second, we propose diffusion image-to-image translation that generates greater quality images than other previous adversarial training based translation methods. Third, we propose a patch-wise and discrete conditional training of diffusion method enabling the reduction of the computation and the robustness on small training datasets.
Subsequently, we discuss a remaining open challenge about evaluation and the direction of future work. Lastly, we make an overall conclusion after stating social impact of this research field.