Visuelle Qualitätsmetrik basierend auf der multivariaten Datenanalyse von H.264/AVC Bitstream-Features

This thesis presents a visual no-reference quality metric for H.264/AVC high definition video, which estimates the video quality exclusively with information directly extracted from the bitstream. Based on the H.264/AVC reference software a ’feature-extractor’ was implemented to extract 64 characteristic bitstream-features (e.g. bitrate, quantization parameters, ratio of different sliceand macroblock-types and different attributes of the motion vectors). The quality metric was built with multivariate data analysis: A partial least squares regression model was calibrated with training data from 32 video sequences and subsequently validated and optimized through cross-validation. The final model contains 48 variously weighted bitstream-features to predict the quality of unknown video.