Human Jury Assessment of Image Quality as a Measurement: Modeling with Bayes Network
Haque, Md. Asadul
Permanent address of the item is
Summary ViewShow full item record
Image quality assessment has been done previously manually by human jury assessment as reference. Due to lack of rationality in human jury voting and its high costs it is desirable to replace it with instrumental measurements that can predict jury assessment reliably. But high uncertainty in jury assessments and sensitivity of image context make it cumbersome for the instrumental measurements. Previous research has shown that modeling with a Bayesian network can resolve some of the problems. A Bayesian network is a belief network of causal model representation of multivariate probabilistic distributions that describes the relationships between the interacting nodes in the form of conditional independency. By conditioning and marginalization operations we can estimate the conditional probabilities of unmeasured elements and their uncertainty in Bayes network. In this thesis we have considered a four-layer pre-existing Bayes network consisting of both qualitative and quantitative component and we have tried to assess probabilities of quality elements assessed by jurors based on instrumental measurement values. To analyze and to quantify the relationship between perceptual quality elements and instrumental measurements, we have calculated mutual information from our provided data set. Based on mutual information calculation and Kullback-Leibler distance measure we have investigated the sensitivity of the network, and we have tried to validate a feasible network model where network parameters have been selected such a way that it minimizes the uncertainties of our chosen Bayes network.