A Mixture of Experts Approach to Multi-strategy Image Quality Assessment

The success of some recently proposed multi-strategy image quality metrics supports the hypothesis that the Human Visual System (HVS) uses multiple strategies when assessing image quality, where the effect from each strategy on the final quality prediction is conditioned on the quality level of the test image. To date, how to optimally combine multiple strategies into a final quality prediction remains an unsolved problem, especially when more than two strategies are involved. In this paper, we present a data-driven combination method based on a conditional Bayesian Mixture of Experts (BME) model. This method provides an effective way to model the interaction of a flexible number of strategies. Extensive evaluation on three publicly-available image quality databases demonstrates the potential of our method.