Diverse Image Generation with Very Low Resolution Conditioning
|Authors:||Abid, Mohamed Abderrahmen|
|Abstract:||Traditionally, when it comes to generating high-resolution (HR) images from a low-resolution(LR) images, the methods proposed so far have mainly focused on super-resolution techniques that aim at recovering the most probable image from low-quality image. Doing so ignores the fact that there are usually many valid versions of HR images that match a given LR image. The objective of this work is to obtain different versions of HR images from the same LR imageusing a generative adversarial model. We approach this problem from two different angles. First, we use super-resolution methods, where in addition to the LR image, the generator can be parameterized by a latent variable to produce different potential variations of the image. Such a conditioning allows to modulate the generator between retrieving the closest image to the ground truth and a variety of possible images. The results demonstrate our superiority in terms of reconstruction and variety of plausible hallucinated images compared to other state-of-the-art methods. The second approach builds on the work of image-to-image translation, by proposing a new approach where the model is conditioned on a LR version of the target. More precisely, our approach aims at transferring the fine details of an HR source image to fit the general structure, according to the LR version of it. We therefore generate HR images that share the distinctive features of the HR image and match the LR image of the target duringdownscaling. This method differs from previous methods that focus instead on translatinga given image style into target content. Qualitative and quantitative results demonstrate improvements in visual quality, diversity, and coverage over state-of-the-art methods such asStargan-v2.|
|Document Type:||Mémoire de maîtrise|
|Open Access Date:||20 September 2021|
|Collection:||Thèses et mémoires|
All documents in CorpusUL are protected by Copyright Act of Canada.