Data-driven estimation and neurophysiological assessment of perceived visual quality

Sebastian Bosse, Technische Universität Berlin, Technische Universität Berlin, Thomas Wiegand
2018
Perceptual quality is one of the key aspects of modern multimedia communication systems. Nevertheless, the questions of how to reliably assess quality as perceived by humans, how to computationally estimate perceived quality and how to incorporate computational models of quality in multimedia systems can still not be answered satisfactorily, despite decades of research. However, these problems are connected, because multimedia systems are typically evaluated in quality assessment studies, the
more » ... tcome of quality assessment studies informs the design of computational quality models and computational quality models are in turn used for the optimization of multimedia systems. This dissertation contributes to the current state of research in several ways. First, a novel neural network-based end-to-end optimized model for image quality estimation is proposed. The proposed method achieves prediction performance that is superior to the state-of-the-art for no-reference as well as for full-reference quality estimation. The second contribution is a formal definition of distortion sensitivity that leads to the derivation of a computationally graceful, perception-based adaptation that can be applied to any given quality model. The proposed framework relates the functional psychometric outcome of quality assessment to a local weighting that can be used to improve the accuracy of quality estimation. A neural network-based method for estimating local weights is proposed and evaluated for the estimation of image quality. In a third contribution, the concept of distortion sensitivity is transferred to rate-distortion theory for lossy compression. A perceptual bit allocation scheme for block-based video compression is derived and experimentally evaluated for compression of still images. Significant bit rate savings are achieved compared to the state of the art, at identical perceptual quality . However, the results suggest that the performance of data-driven quality models crucially depends on the availability of labeled traini [...]
doi:10.14279/depositonce-7233 fatcat:jnsewueysnbbnex3e2kp4qdan4