Visual Content Characterization Based on Encoding Rate-Distortion Analysis
MetadataShow full item record
Visual content characterization is a fundamentally important but under exploited step in dataset construction, which is essential in solving many image processing and computer vision problems. In the era of machine learning, this has become ever more important, because with the explosion of image and video content nowadays, scrutinizing all potential content is impossible and source content selection has become increasingly difficult. In particular, in the area of image/video coding and quality assessment, it is highly desirable to characterize/select source content and subsequently construct image/video datasets that demonstrate strong representativeness and diversity of the visual world, such that the visual coding and quality assessment methods developed from and validated using such datasets exhibit strong generalizability. 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. However, encoding RD analysis has not been well investigated in the context of visual content characterization. This thesis focuses on applying encoding RD analysis as a visual source content characterization method with image/video coding and quality assessment applications in mind. We first conduct a video quality subjective evaluation experiment for state-of-the-art video encoder performance analysis and comparison, where our observations reveal severe problems that motivate the needs of better source content characterization and selection methods. Then the effectiveness of RD analysis in visual source content characterization is demonstrated through a proposed quality control mechanism for video coding by eigen analysis in the space of General Quality Parameter (GQP) functions. Finally, by combining encoding RD analysis with submodular set function optimization, we propose a novel method for automating the process of representative source content selection, which helps boost the RD performance of visual encoders trained with the selected visual contents.
Cite this version of the work
Zhuoran Li (2023). Visual Content Characterization Based on Encoding Rate-Distortion Analysis. UWSpace. http://hdl.handle.net/10012/19185