The ToonOut Model and Mathematical Theory behind It

The rapid evolution of artificial intelligence in the last period has led to new ways of using this technology, which until recently appeared unfeasible. In this thesis, we focus on the application of Machine Learning to computer vision, specifically in the areas of image generation and segmentation. First we discuss the Diffusion Models topic. This class of models is grounded in mathematical and probabilistic theory: they learn the training data distribution while gradually adding noise to training images and use theoretical tools with the purpose to obtain new samples similar to those the models are trained on, starting from pure noise. Subsequently, a computer vision task that can be performed on generated images is addressed. This is called Dichotomous Image Segmentation and aims to segment the main object/objects represented in the image, to distinguish them from the background. On this topic, state-of-the-art models are introduced, as well as the metrics used to evaluate the quality of the results and their mathematical structures. Finally, the ToonOut model is explained. This is a background removal project developed by us in collaboration with the Kartoon company. We also designed a new metric for Dichotomous Image Segmentation, which better deals with the background removal context where ToonOut is applied.