Optimización del proceso de transformación de imágenes usando Redes Generativas Adversarias Basadas en CycleGan
Loading...
Official URL
Full text at PDC
Publication date
2023
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citation
Abstract
En esta memoria se va a exponer el proceso de desarrollo de una red neuronal con una arquitectura de tipo Red Generativa Adversaria que recibe imágenes que comparten características comunes y las transforma en imágenes pertenecientes a otro conjunto con características comunes diferentes, manteniendo los atributos propios de la imagen. En concreto, se presenta cómo se ha implementado dicha red, la arquitectura de las subredes que la forman y el entrenamiento y ajuste de dichas subredes. Estas arquitecturas permiten el entrenamiento de los modelos en un computador antiguo con poca capacidad para hacer operaciones vectoriales. Ambas arquitecturas son una variación de la arquitectura original de una Red Generativa Adversaria Cíclica pero la primera reduce la resolución de las imágenes de entrada reduciendo así también el número de parámetros entrenables. La segunda arquitectura es completamente
opuesta a la primera. Aumenta la resolución de las imágenes mientras que incrementa el número de parámetros entrenables.
In this report, the development process of a neural network using a Generative Adversarial Network architecture will be outlined. This Generative Adversarial Network takes images with shared common characteristics and transforms them into images belonging to another set with different common features, while preserving the inherent attributes of the image. Specifically, how this network has been implemented, the architecture of the subnetworks that compose it, and the training and fine-tuning of these subnetworks will be described. These architectures are designed to enable model training on older computers with limited capacity for vector operations. Both architectures are variations of the original Cyclical Generative Adversarial Network architecture. The first one reduces the resolution of input images, thus decreasing the number of trainable parameters. In contrast, the second architecture increases the resolution of images while also increasing the number of trainable parameters.
In this report, the development process of a neural network using a Generative Adversarial Network architecture will be outlined. This Generative Adversarial Network takes images with shared common characteristics and transforms them into images belonging to another set with different common features, while preserving the inherent attributes of the image. Specifically, how this network has been implemented, the architecture of the subnetworks that compose it, and the training and fine-tuning of these subnetworks will be described. These architectures are designed to enable model training on older computers with limited capacity for vector operations. Both architectures are variations of the original Cyclical Generative Adversarial Network architecture. The first one reduces the resolution of input images, thus decreasing the number of trainable parameters. In contrast, the second architecture increases the resolution of images while also increasing the number of trainable parameters.
Description
Trabajo Fin de Grado en Ingeniería Informática, Facultad de Informática UCM, Departamento de Ingeniería de Software e Inteligencia Artificial, Cuso 2022/2023.