Traditional approach to collect mean opinion score (MOS) values for evaluation of full-reference image quality metrics has two serious drawbacks. The first drawback is a nonlinearity of MOS, only partially compensated by the use of rank order correlation coefficients in a further analysis. The second drawback are limitations on number of distortion types and distortion levels in image database imposed by a maximum allowed time to carry out an experiment. One of the largest of databases used for this purpose, TID2013, has almost reached these limitations, which makes an extension of TID2013 within the boundaries of this approach to be practically unfeasible. In this paper, a novel methodology to collect MOS values, with a possibility to infinitely increase a size of a database by adding new types of distortions, is proposed. For the proposed methodology, MOS values are collected for pairs of distortions, one of them being a signal dependent Gaussian noise. A technique of effective linearization and normalization of MOS is described. Extensive experiments for linearization of MOS values to extend TID2013 database are carried out.
The problem of increasing efficiency of blind image quality assessment is considered. No-reference image quality metrics both independently and as components of complex image processing systems are employed in various application areas where images are the main carriers of information. Meanwhile, existing noreference metrics have a significant drawback characterized by a low adequacy to image perception by human visual system (HVS). Many well-known no-reference metrics are analyzed in our paper for several image databases. A method of combining several noreference metrics based on artificial neural networks is proposed based on multi-database verification approach. The effectiveness of the proposed approach is confirmed by extensive experiments.