On the application of AI in subjective image editing

Image editing is a laborious process, where a user must manually and meticulously add, modify, or change elements in the image to improve its aesthetics. A professional image editing task can take hours or more due to the level of detail required in the tasks, such as retouching seams and changing lighting conditions. On the other hand, for average users, image editing can also be a non-trivial task, since they do not have the expertise in the technical knowledge and the artistic vision. As a result, an automatic image editing tool can benefit both professional editors and casual users by taking over the labor-intensive process and reducing costs, while providing inspiration for creativity. Nevertheless, image editing is also an inherently subjective task. This is because users can have different opinions on how to edit an image according to their preferences. For instance, one user may prefer a populated street and keep all the people, while another user may favor the scenery and thus wish to remove all pedestrians. This subjectivity poses new challenges in developing deep learning based image editing models, since there is no longer a single ground truth for the model to be trained on. In this thesis, we propose a series of tasks in an attempt to tackle the application of AI in subjective image editing. We approach the topic in two ways: discriminative and generative. For discriminative automatic image editing, we study the task of visual distractor detection in two scenarios: wire segmentation in high-resolution images, where we constrain the type of distractor in uncommon image settings; and general distractor detection, where we study the properties of distractors in general. For generative image editing, we propose a new task of visual concept recommendation, which aims to recommend several sensible objects for insertion given the image context. In each study, we provide a clear task definition and our proposed pipeline. We discuss methods of quantifying the subjectivity in each problem, and conduct comprehensive experiments to evaluate our methods.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2024

Author dc:creator

• Chiu, Mang Tik

Contributors dc:contributor

• Shi, Humphrey
• Hasegawa-Johnson, Mark A
• Hwu, Wen-mei
• Varshney, Lav R

• Computer Vision
• Ai
• Deep Learning
• Image
• Editing

Statement dc:rights

• Copyright 2024 Mang Tik Chiu

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/124495

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F124495