Ph.D Defence Notice: "Visual Content Characterization Based on Encoding Rate-Distortion Analysis" by Zhuoran Li

Friday, February 3, 2023 10:00 am - 10:00 am EST (GMT -05:00)


Database is the fundamental building block of research. But how to construct a representative database depends on the effectiveness of data characterization method. For compression quality research, visual content characterization is necessary in that it can depict each content from quality perspective and replace the unreliable human judgement.

In the era of machine learning, just like other research areas, visual quality research community is hungry for large dataset as well. However, given the unbeatable time and cost limitations of data collection and evaluation process, the only possible way for researchers is to pick representative source visual content so that the real-world data can be approximated. With the explosion of image and video contents nowadays, the source content selection is increasingly difficult, since one cannot scrutinize all visual contents in his or her lifetime. Therefore, source contents should be characterized first using quality related prior knowledge so that a following automated process can be designed to select source content in a representative manner.

Encoding Rate-Distortion (RD) analysis is essential for many multimedia applications. Examples of applications that explicitly use RD analysis include image encoder RD optimization, video quality assessment (VQA), and quality-of-experience (QoE) optimization of streaming videos etc. Though it has been used in many optimization tasks, encoding RD analysis' role as a source signal characteristic is less visited. The thesis focuses on applying encoding RD analysis as a visual source content characterization method for compression applications. We will reveal the problem using a practical video quality subjective evaluation experiment, which focuses on the encoder performance comparison. Then the effectiveness of RD analysis on visual source content characterization will be demonstrated through a precise encoded video quality control model. Lastly, an RD domain source content selection model is proposed. Combining encoding RD analysis with submodular set function optimization, we are the first to propose the novel method on automating the process of representative source content selection for compression applications.