UWSpace is currently experiencing technical difficulties resulting from its recent migration to a new version of its software. These technical issues are not affecting the submission and browse features of the site. UWaterloo community members may continue submitting items to UWSpace. We apologize for the inconvenience, and are actively working to resolve these technical issues.
 

Degraded Reference Image Quality Assessment

Thumbnail Image

Files

Guo_Xinyu.pdf (2.02 MB)

Date

2021-12-23

Authors

Guo, Xinyu

Journal Title

Journal ISSN

Volume Title

Publisher

University of Waterloo

Abstract

Images/videos are playing a more and more important role in the 21st century. The perceived quality of visual content often degrades during the process of acquisition, storage, transmission, display and rendering. Since subjective evaluation of such a large amount of visual content is impossible, the development of objective evaluation methods becomes highly desirable. Traditionally, there are three well established Image Quality Assessment (IQA) paradigms. They are Full Reference (FR) IQA which needs full access to the pristine quality reference, Reduced Reference (RR) IQA which requires partial information from the pristine reference and, No Reference (NR) IQA which does not require any reference information. While the strict requirement prohibits FR IQA from wide usage in many applications, RR and NR IQA methods cannot produce comparable performance. In the thesis, we aim to address this problem by exploring the Degraded Reference (DR) paradigm which makes no requirement on pristine reference but on reference of degraded quality, and at the same time, outperforms RR/NR methods. We address this problem in three steps. Firstly, we develop an FR model built upon a Deep Neural Network (DNN) that can handle multiply distorted images. The model structure of this FR model is then utilized to design DNN-based DR IQA models. We further improve the DR DNN model by adjusting the network structure. Finally, we use a two-step framework, which utilizes an NR model and an FR model as base modules followed by a regressor to create a single DR prediction for a given image. We test our models on subject-related datasets in IQA field. The testing results show that our FR model has state-of-the-art performance when handling multiply distorted images, and meanwhile produces great performance when handling singly distorted images. Our DR model developed using the two-step framework gives better performance than RR/NR models when the reference is not pristine.

Description

Keywords

image quality assessment, degraded reference, deep neural networks, multiply distorted images

LC Keywords

Citation

URI

http://hdl.handle.net/10012/17818

Collections

Theses
Electrical and Computer Engineering