Author: Zeng, Hui
Title: Learning to evaluate and enhance image quality
Advisors: Zhang, Lei (COMP)
Degree: Ph.D.
Year: 2021
Subject: Image processing -- Digital techniques
Neural networks (Computer science)
Imaging systems -- Image quality
Hong Kong Polytechnic University -- Dissertations
Department: Department of Computing
Pages: xx, 171 pages : color illustrations
Language: English
Abstract: The explosive growth of digital images and their ubiquitous usage in our daily life raise two fundamental challenges to the image processing community and digital camera manufacturers. The first one is how to automatically and reliably evaluate the perceptual quality of the huge amounts of images without additional information. The second one is how to design more intelligent and efficient algorithms to further enhance the image quality. In this thesis, we embrace the recent advance of artificial intelligence, especially the powerful deep convolutional neural network (CNN), to tackle these two challenges. Regarding the first challenge, we make in-depth analyses on the limitations of existing blind image quality assessment (BIQA) and image aesthetic assessment (IAA) methods, and propose a new solution to each of the two problems. In the first work, we propose a probabilistic quality representation (PQR) approach to describe the image subjective score distribution, whereby a more robust loss function can be employed to train more effective deep BIQA models. In the second work, we reveal, for the first time to our best knowledge, the inherent relationship of the three IAA tasks, and propose a unified probabilistic formulation to handle all of them. We also propose a method to fit the aesthetic scores to a more stable and discriminative score distribution, which contributes to learning more competitive IAA models on all the three tasks. To address the second challenge, we propose a new formulation to improve the image composition and a novel method to enhance the color and tone of images. Specifically, in the third work, we revisit the image cropping task and propose a more reasonable formulation, namely grid anchor based image cropping (GAIC). Under the new formulation, we construct a new cropping benchmark and define more reliable evaluation metrics. We also design an effective and lightweight cropping model by considering the special properties of image cropping, and achieve robust image cropping performance with high efficiency. In the fourth work, we propose to learn image-adaptive 3D lookup tables (LUTs) to achieve high-performance and real-time photo enhancement. In addition to obtaining clear state-of-the-art performance on two general-purpose photo enhancement benchmarks, we also successfully apply the image-adaptive 3D LUT model to the portrait photo retouching (PPR) task with two dedicated improvements. In summary, in this thesis we introduce human domain knowledge into the formulation and learning of task-specific CNN models, providing new perspectives on designing more effective image processing and computational photography models. The proposed image quality and aesthetic assessment methods can be widely used to monitor the huge amounts of images uploaded to internet and recommend high-quality ones to users, while the proposed enhancement methods can be deployed in digital camera devices and image processing software to automatically generate more visually pleasing images.
Rights: All rights reserved
Access: open access

Files in This Item:
File Description SizeFormat 
5842.pdfFor All Users31.88 MBAdobe PDFView/Open

Copyright Undertaking

As a bona fide Library user, I declare that:

  1. I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
  2. I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
  3. I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.

By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.

Show full item record

Please use this identifier to cite or link to this item: