Art`s colour retrieval dataset & incremental learning colour palette generator
Date
2024-04-17
Embargo
Advisor
Coadvisor
Journal Title
Journal ISSN
Volume Title
Publisher
Language
English
Alternative Title
Abstract
Este projeto apresenta uma ferramenta de combinac¸ao de cores baseada no framework ˜
GAN, com o objetivo de desmistificar o processo de selecionar cores. Este projecto
foi impulsionado por novos desenvolvimentos e avanc¸os na tecnologia de inteligencia ˆ
artificial. Apesar da existencia de alguns modelos online, ainda s ˆ ao escassos os pro- ˜
jetos sistematicamente documentados na literatura cientifica que abordem os desafios ˆ
associados a combinac¸ ` ao de cores. ˜
Para treinar o modelo generativo, foi criado um conjunto de dados composto por paletes
de cores derivadas de imagens de filmes de animac¸ao. As cores no conjunto de dados ˜
foram descritas usando o sistema perceptualmente uniforme de cores CIELUV. A ferra menta esta publicada numa p ´ agina web, o que possiblita o seu uso em v ´ arios tipos de ´
dispositivos.
As interac¸oes do utilizador com a p ˜ agina web geram um novo conjunto de dados, que ´
pode ser utilizado para o treino continuado do modelo generativo ou aplicado em out ros projetos semelhantes. Foi realizada uma analise detalhada das paletes de cores ´
provenientes de imagens, com foco na distribuic¸ao dos componentes de CIE Lch. Al ˜ em´
disso, a estabilidade da GAN foi examinada ao longo das varias fases do treino. ´
Os algoritmos introduzidos, bem como os dois conjuntos de dados, servem como base
para futuros projetos experimentais no campo da teoria das cores. A ferramenta web
nao s ˜ o simplifica o processo de combinac¸ ´ ao de cores, mas tamb ˜ em contribui com um ´
conjunto de dados em constante expansao para explorac¸ ˜ ao futura em projetos relacion- ˜
ados.
This project introduces a colour-combining tool based on the GAN framework, aiming to enhance the creative workflow when selecting colours. This was spurred by the promising avenues opened up by advancements in technology and artificial intelligence. Despite the existence of a few machine learning models online, there is still a notable scarcity of thorough documented projects in the published literature that effectively ad dress the challenges associated with colour combination solutions. To train the generative model, a dataset was curated, comprising colour palettes derived from frames of animated films. The colours in the dataset were described using the perceptually uniform CIELUV colour system. The tool is accessible via a web page, promoting cross-platform usability. User interactions with the web page generate a new dataset, which can be utilised for on going training of the generative model or applied in similar projects. An in-depth analysis of colour palettes in the frames dataset was conducted, focusing on hue, chroma, and luminance distribution. Additionally, the stability of the GAN was examined by studying its losses across multiple training epochs. The algorithms introduced, along with the two datasets serve as a foundation for future experimental research in the realm of colour theory. The web tool not only enhances colour combination solutions but also contributes with a constantly expanding dataset for further exploration in related projects.
This project introduces a colour-combining tool based on the GAN framework, aiming to enhance the creative workflow when selecting colours. This was spurred by the promising avenues opened up by advancements in technology and artificial intelligence. Despite the existence of a few machine learning models online, there is still a notable scarcity of thorough documented projects in the published literature that effectively ad dress the challenges associated with colour combination solutions. To train the generative model, a dataset was curated, comprising colour palettes derived from frames of animated films. The colours in the dataset were described using the perceptually uniform CIELUV colour system. The tool is accessible via a web page, promoting cross-platform usability. User interactions with the web page generate a new dataset, which can be utilised for on going training of the generative model or applied in similar projects. An in-depth analysis of colour palettes in the frames dataset was conducted, focusing on hue, chroma, and luminance distribution. Additionally, the stability of the GAN was examined by studying its losses across multiple training epochs. The algorithms introduced, along with the two datasets serve as a foundation for future experimental research in the realm of colour theory. The web tool not only enhances colour combination solutions but also contributes with a constantly expanding dataset for further exploration in related projects.
Keywords
Teoria de Cor, Redes adversativas generativas, Quantizacão de Cor, Espaço uniforme de Cor
Document Type
Master thesis
Publisher Version
Dataset
Citation
Cruz, T. V. D. A. (2024). Art`s colour retrieval dataset & incremental learning colour palette generator [Dissertação de Mestrado em Ciência de Dados, Universidade Portucalense]. Repositório Institucional UPT. http://hdl.handle.net/11328/5589
Identifiers
TID
203608933
Designation
Mestrado em Ciência de Dados
Access Type
Open Access