Accepted author manuscript, 233 KB, PDF document
Final published version
Research output: Contribution in Book/Report/Proceedings - With ISBN/ISSN › Conference contribution/Paper › peer-review
Research output: Contribution in Book/Report/Proceedings - With ISBN/ISSN › Conference contribution/Paper › peer-review
}
TY - GEN
T1 - Perceptual video quality estimation by regression with myopic experts
AU - Giotsas, Vasileios
AU - Deligiannis, Nikos
AU - Fisher, Pam
AU - Andreopoulos, Yiannis
PY - 2015/7/2
Y1 - 2015/7/2
N2 - Objective video quality metrics can be viewed as 'myopic' expert systems that focus on particular aspects of visual information in video, such as image edges or motion parameters. We conjecture that the combination of many such high-level metrics leads to statistically-significant improvement in the prediction of reference-based perceptual video quality in comparison to each individual metric. To examine this hypothesis in a systematic and rigorous manner, we use: (i) the LIVE and the EPFL/PoliMi databases that provide the difference mean opinion scores (DMOS) for several video sequences under encoding and packet-loss errors; (ii) ten well-known metrics that range from mean-squared error based criteria to sophisticated visual quality estimators; (iii) five variants of regression-based supervised learning. For 400 experimental trials with random (non-overlapping) estimation and prediction subsets taken from both databases, we show that the best of our regression methods: (i) leads to statistically-significant improvement against the best individual metrics for DMOS prediction for more than 97% of the experimental trials; (ii) is statistically-equivalent to the performance of humans rating the video quality for 36.75% of the experiments with the EPFL/PoliMi database. On the contrary, no single metric achieves such statistical equivalence to human raters in any of the experimental trials.
AB - Objective video quality metrics can be viewed as 'myopic' expert systems that focus on particular aspects of visual information in video, such as image edges or motion parameters. We conjecture that the combination of many such high-level metrics leads to statistically-significant improvement in the prediction of reference-based perceptual video quality in comparison to each individual metric. To examine this hypothesis in a systematic and rigorous manner, we use: (i) the LIVE and the EPFL/PoliMi databases that provide the difference mean opinion scores (DMOS) for several video sequences under encoding and packet-loss errors; (ii) ten well-known metrics that range from mean-squared error based criteria to sophisticated visual quality estimators; (iii) five variants of regression-based supervised learning. For 400 experimental trials with random (non-overlapping) estimation and prediction subsets taken from both databases, we show that the best of our regression methods: (i) leads to statistically-significant improvement against the best individual metrics for DMOS prediction for more than 97% of the experimental trials; (ii) is statistically-equivalent to the performance of humans rating the video quality for 36.75% of the experiments with the EPFL/PoliMi database. On the contrary, no single metric achieves such statistical equivalence to human raters in any of the experimental trials.
KW - objective metrics
KW - perceptual video quality
KW - supervised learning
U2 - 10.1109/QoMEX.2015.7148115
DO - 10.1109/QoMEX.2015.7148115
M3 - Conference contribution/Paper
AN - SCOPUS:84939524131
SP - 1
EP - 6
BT - 2015 7th International Workshop on Quality of Multimedia Experience, QoMEX 2015
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 2015 7th International Workshop on Quality of Multimedia Experience, QoMEX 2015
Y2 - 26 May 2015 through 29 May 2015
ER -