Implementació d'un procés de transferència d'estil mitjançant una GAN

This work is based on the application of generative adversarial networks (GAN) to transfer the style of a set of images, specific to an author, to an input image. Specifically, we want to achieve models capable of generating new images, given an input real photograph as input, applying the transfer...

Descripción completa

Detalles Bibliográficos
Autor: Deza Tripiana, Ricard
Tipo de recurso: tesis de maestría
Fecha de publicación:2020
País:España
Institución:Universitat Oberta de Catalunya (UOC)
Repositorio:O2, repositorio institucional de la UOC
OAI Identifier:oai:openaccess.uoc.edu:10609/123406
Acceso en línea:http://hdl.handle.net/10609/123406
Access Level:acceso abierto
Palabra clave:generative adversarial network
style transfer
deep learning
xarxes generatives adversarials
transferència d'estil
aprenentatge profund
redes generativas adversariales
transferencia de estilo
aprendizaje profundo
Data mining -- TFM
Mineria de dades -- TFM
Minería de datos -- TFM
Descripción
Sumario:This work is based on the application of generative adversarial networks (GAN) to transfer the style of a set of images, specific to an author, to an input image. Specifically, we want to achieve models capable of generating new images, given an input real photograph as input, applying the transfer of style of paintings by van Gogh, Picasso and Pollock. This study delves into the different characteristics of the images processed by the networks and the components involved in the style transfer process. It is based on the configuration and treatment of losses described in the article "Artsy" GAN: A style transfer system with improved quality, diversity, and performance¿ by Liu et al. (2016). This paper proposes an adversarial generative approach using perceptual loss, processing images with chroma subsampling, introducing noise into generator input images, and a loss target function that encourages generating different details for the same content image. These modifications are intended to improve the performance and quality of the results obtained with previous studies, such as the use of CycleGan's.