Concepts from unclear textual embeddings for text-to-image synthesis

Automatically generating images based on a natural language description is a challenging problem with several key applications in the fields of retail, marketing, education and entertainment. In the last few years, some progress has been made in this direction specifically by using Generative Adversarial Networks(GANs). Although current state of the art models can generate images that roughly adhere to the textual description, there still remains a long way to go, in terms of producing high quality images that adhere to the nuances of the sentence. To this end we propose CuteGAN, our simple text-to-image generation approach that encourages the model to leverage the attribute information while also attending to more relevant words in a sentence while generating images. We perform experiments on the competitive CUB-200 and MS-COCO datasets and achieve state-of-the-art performance on standard metrics of inception score and R-precision, indicating that our method produces more photo-realistic images that are better correlated with the text.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2020

Author dc:creator

• Kumar, Maghav

Contributors dc:contributor

• Schwing, Alexander

• Computer Vision
• deep learning
• GANs
• text-to-image
• machine learning

Statement dc:rights

• Copyright 2019 Maghav Kumar

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/106392

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/106392