Understanding perceived quality through visual representations

Abstract

The formatting of images can be considered as an optimization problem, whose cost function is a quality assessment algorithm. There is a trade-off between bit budget per pixel and quality. To maximize the quality and minimize the bit budget, we need to measure the perceived quality. In this thesis, we focus on understanding perceived quality through visual representations that are based on visual system characteristics and color perception mechanisms. Specifically, we use the contrast sensitivity mechanisms in retinal ganglion cells and the suppression mechanisms in cortical neurons. We utilize color difference equations and color name distances to mimic pixel-wise color perception and a bio-inspired model to formulate center surround effects. Based on these formulations, we introduce two novel image quality estimators PerSIM and CSV, and a new image quality-assistance method BLeSS. We combine our findings from visual system and color perception with data-driven methods to generate visual representations and measure their quality. The majority of existing data-driven methods require subjective scores or degraded images. In contrast, we follow an unsupervised approach that only utilizes generic images. We introduce a novel unsupervised image quality estimator UNIQUE, and extend it with multiple models and layers to obtain MS-UNIQUE and DMS-UNIQUE. In addition to introducing quality estimators, we analyze the role of spatial pooling and boosting in image quality assessment.