Candidate: Xinyu Guo
Title: Degraded Reference Image Quality Assessment
Date: December 8, 2021
Place: MS Teams
Supervisor(s): Wang, Zhou
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
Candidate: Xinyu Guo