Manipulation of cognitive image properties
Loading...
Journal Title
Journal ISSN
Volume Title
Perustieteiden korkeakoulu |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Author
Date
2022-07-29
Department
Major/Subject
Machine Learning, Data Science and Artificial Intelligence
Mcode
SCI3044
Degree programme
Master’s Programme in Computer, Communication and Information Sciences
Language
en
Pages
62 + 9
Series
Abstract
Photo retouching is an inherent part of visual content production in the profes- sional world. Although some general gPhoto retouching is an inherent part of visual content production in the professional world. Although some general guides exist, there is no universal principle of how the retouched photo should look, as there is no one strict de nition of beauty. This process is often subjective and laborious. Deep learning techniques have provided tremendous improvements to the image processing domain in recent years. Nowadays we can generate realistic images and edit them. Recent work in this eld proves, that we can enhance photos by style transfer or guide a Generative Adversarial Network to create more aesthetic images. Less focus was given to modifying arbitrary photos according to the broad notion of aesthetics. Our main question in this thesis is: Is it possible to increase the aesthetics of any photo, with no direct human supervision? We propose a simple technique for nding a tone curve mapping that increases photo aesthetics. The process is guided by a neural network that was previously trained to assess this property. We optimize the tone curve parameters using gradients backpropagated from the network. The framework does not assume anything specific about aesthetics and can be used with other cognitive properties, such as memorability or emotional valence. We investigate the properties and limitations of the algorithm. We design and run a user study to validate our results. We nd that participants prefer our enhancements from initial photos in 66.5% of cases. We analyze the probable causes of their decisions. uides exist, there is no universal principle of how the retouched photo should look, as there is no one strict de nition of beauty. This process is often subjective and laborious. Deep learning techniques have provided tremendous improvements to the image processing domain in recent years. Nowadays we can generate realistic images and edit them. Recent work in this eld proves, that we can enhance photos by style transfer or guide a Generative Adversarial Network to create more aesthetic images. Less focus was given to modifying arbitrary photos according to the broad notion of aesthetics. Our main question in this thesis is: Is it possible to increase the aesthetics of any photo, with no direct human supervision? We propose a simple technique for nding a tone curve mapping that increases photo aesthetics. The process is guided by a neural network that was previously trained to assess this property. We optimize the tone curve parameters using gradients backpropagated from the network. The framework does not assume anything specific about aesthetics and can be used with other cognitive properties, such as memorability or emotional valence. We investigate the properties and limitations of the algorithm. We design and run a user study to validate our results. We nd that participants prefer our enhancements from initial photos in 66.5% of cases. We analyze the probable causes of their decisions.Description
Supervisor
Lehtinen, JaakkoThesis advisor
Lehtinen, JaakkoKeywords
aesthetics, image editing, neural networks, tone curve mapping, cognitive image properties
