With the rapid advancement of display technologies in recent years, display resolutions have evolved from UHD (4K) toward 8K and even 16K
. Consequently, the demand for high-throughput data transmission and computationally intensive processing in visu...
With the rapid advancement of display technologies in recent years, display resolutions have evolved from UHD (4K) toward 8K and even 16K
. Consequently, the demand for high-throughput data transmission and computationally intensive processing in visual display systems has increased significantly. Although display interface standards such as DisplayPort (DP) have evolved to support higher bandwidth, the explosive growth of ultra-high-resolution content, particularly on over-the-top (OTT) platforms such as YouTube and Netflix, has led to substantial increases in transmitted data volume and decoding, scaling, and rendering complexity. These requirements pose critical technical challenges, especially for systems with stringent physical and power constraints, such as automotive and portable displays.
To mitigate these issues, modern display processing pipelines increasingly rely on data-efficient transmission and reconstruction techniques, such as super-resolution, as well as visually lossless compression schemes like VESA’s Display Stream Compression (DSC). However, evaluation methodologies for display systems, specifically those that assess whether visual data are correctly generated and transmitted at the display interface level and transmitted without transmission errors or buffer overflows, remain insufficiently explored. As a result, most systems depend on distortion-based metrics such as PSNR, which fail to adequately capture human perceptual quality. Moreover, conventional compression and upscaling approaches that rely on fixed and discrete scaling factors exhibit inherent limitations in maintaining visual quality in today’s environment, where web-based content of highly diverse resolutions must be rendered on displays with fixed native resolutions.
This dissertation addresses the aforementioned challenges in visual display systems by pursuing three complementary research directions centered on perceptual image quality assessment and resolution-adaptive arbitrary-scale image generation. First, in the domain of perceptual quality evaluation, we investigate Full-Reference Image Quality Assessment (FR-IQA) and propose a multi-task feature integration framework that enhances quality prediction performance by aggregating representations from multiple computer vision tasks. Second, for No-Reference IQA (NR-IQA), we introduce a surrogate full-reference evaluation framework that leverages restored images, effectively reformulating the blind IQA problem through the integration of image restoration and a Dual-Attention Task Interaction mechanism. Finally, to enable high-quality rendering of content with diverse resolutions on fixed-resolution displays, we develop an arbitrary super-resolution approach based on Self-Cascaded Diffusion Models, enabling resolution-adaptive and perceptually faithful reconstruction under varying display conditions.
Collectively, these contributions advance the perceptual evaluation and adaptive generation of high-resolution visual content, providing principled and practically applicable solutions to emerging challenges in next-generation display systems.