<title>Practical applications that require some of the more advanced features of current visual models</title>

Abstract

While the use of visual models for assessing all aspects of the imaging chain is steadily increasing, one hindrance is the complexity of these models. This has impact in two ways - not only does it take longer to run the more complex visual model, making it difficult to place into optimization loops, but it also takes longer to code, test, and calibrate the model. As a result, a number of shortcut models have been proposed and used. Some of the shortcuts involve more efficient frequency transforms, such as using a Cartesian separable wavelet, while other types of shortcuts involve omitting the steps required to simulate certain visual mechanisms, such as masking. A key example of the latter is spatial CIELAB, which only models the opponent color CSFs and does not model the spatial frequency channels. Watson's recent analysis of the Modelfest data showed that while a multi-channel model did give the best performance, versions dispensing with the complex frequency bank and just using frequency attenuation did nearly as well. Of course, the Modelfest data addressed detection of a signal on a uniform field, so no masking properties were probed. On the other end of complexity is the model by D'Zmura, which not only includes radial and orientation channels, but also the interactions between the channels in both luminance and color. This talk will dissect several types of practical distortions that require more advanced visual models. One of these will be the need for orientation channels to predict edge jaggies due to aliasing. Other visual mechanisms in search of an exigent application that we will explore include cross luminance-chrominance masking and facilitation, local contrast, and cross-channel masking.