Psychophysical measurement and prediction of digital video quality

Digital signal processing methods are used to transform video data for many purposes. Innovation in this area has accelerated during the last decade as hardware resources have become more capable and less expensive. Much of the research in video processing is concerned with producing video that looks good. Compression algorithms attempt to transmit good video with minimum resources. Enhancement algorithms attempt to improve degraded video. In the end, every engineer in this area faces a common problem: How do you judge the success of your work?; The normal approach compares the algorithm inputs and outputs. Several comparison methods, or fidelity metrics, are in common use. The most common metrics are computationally simple but relatively inaccurate. Efforts to improve the simple metrics have been underway for more than thirty years. Several complex methods based on human visual system models have been proposed. These metrics promise more accurate fidelity measurements at the cost of computational complexity.; The main goal of this research was to test the assumptions made in recent fidelity metrics and to develop models with similar accuracy but reduced computational complexity. First, several unique psychophysical experiments were performed to accumulate subjective data sets. Afterwards, fidelity models were developed.; The psychophysical experiments measured several subjective attributes of short, spatially-limited defects inserted into otherwise normal videos. The experiments measured subjective values such as detection probability, annoyance value, and importance. These values were examined for their most significant influences and compared to find their inter-relationships.; Two models were developed. A full-reference model was created using methods common in other complex fidelity metrics. Several implementations of each model stage were tested to determine the trade-off between computational complexity and accuracy. The final model predicted both detection and impairment results.; Finally, a reduced-reference model was designed. The model combined objective detection thresholds and simple impairment measures. The resulting model reduced reference information by 1/400 and, more importantly, processing time by 1/10,000. This implementation gain was achieved while maintaining the model accuracy.