| Abstract |
Subjective perceptual image quality can be assessedin lab studies by human observers. Objective image qualityassessment (IQA) refers to algorithms for estimation of themean subjective quality ratings. Many such methods have beenproposed, both for blind IQA in which no original referenceimage is available as well as for the full-reference case. Wecompared 8 state-of-the-art algorithms for blind IQA and showedthat an oracle, able to predict the best performing method for anygiven input image, yields a hybrid method that could outperformeven the best single existing method by a large margin. Inthis contribution we address the research question whetherestablished methods to learn such an oracle can improve blindIQA. We applied AutoFolio, a state-of-the-art system that trainsan algorithm selector to choose a well-performing algorithm for agiven instance. We also trained deep neural networks to predictthe best method. Our results did not give a positive answer,algorithm selection did not yield a significant improvement overthe single best method. Looking into the results in depth, weobserved that the noise in images may have played a role in whyour trained classifiers could not predict the oracle. This motivatesthe consideration of noisiness in IQA methods, a property thathas so far not been observed and that opens up several interestingnew research questions and applications. |