Tamer Ajaj: EEG-based assessment of perceived quality in complex natural images

BCCN Berlin / TU Berlin

 

Abstract

Assessment of visual perceptual quality is an important tool for measuring the performance of multimedia communication methods. In recent years, psychophysiological methods have been used to provide more objective estimations of perceptual quality. One important approach for psychophysiological assessment is analyzing neural signals of Electroencephalography (EEG). Steady State Visually Evoked Potential (SSVEP) framework has been used for the purpose of quality assessment using EEG signals. In the SSVEP framework, texture images have been previously used as stimuli, but natural images haven’t been tested. This motivates the application of this framework on the assessment of the quality perception of more realistic stimuli.

This study consists of the following aspects. First, it utilizes complex images extracted from the new standardized videos data-set presented by the Video Quality Experts Group (VQEG). Second, it compares the performance of two dimensionality reduction techniques of the EEG signals. Particularly, the methods of Spatio-Spectral Decomposition (SSD) and Reliable Component Analysis (RCA) are adapted for the SSVEP framework and useful neural components are extracted from the EEG recordings. Third, correlations between neural signals and Mean Opinion Score (MOS) are calculated to validate the dimensionality reduction techniques, and the predicting power of MOS from neural signals is estimated. In addition, an outlier rejection method is proposed to exclude participants that do not contribute to the quality estimation process. This study implemented an experimental paradigm that ensured synchronization and replicability for future studies.

Results show a high correlation between neural components and MOS values. In addition, subsets of harmonics are tested to achieve prediction results with SSD and RCA that are superior to the ones solely based on behavioral inputs. The outlier rejection method of participants shows promising results as it detects participants with low correlation values.

 

 

Additional Information

Master thesis defense

Organized by

Prof. Dr. Klaus-Robert Müller

Location: TU Berlin, room MAR 4.033, Marchstr. 23, 10587 Berlin

Go back