An Adaptable Human Vision Model for Subjective Video Quality Rating Prediction among Cif, Sd, Hd and E-cinema

A new highly adaptable model for predicting human vision response is presented for enabling an improved method of predicting subjective video quality. The ability to adapt enables comparison of video with dissimilar image sizes, viewing environments, frame rates, video quality classes, etc. (for example HD vs. SD vs. CIF). Model test results are compared with human response. Responses are from stimuli covering both JND (just noticeable differences) and supra-threshold (extending to near the opposite extreme). Supra-threshold responses compared include adaptation behavior exemplified by nonlinear response responsible for significant sensitivity changes, masking and visual illusions. Given the prediction of visible impairments, simulation of the contextual adaptation that occurs during the training portion of ITU-R BT.500 (subjective assessment methodology) is used for predicting DMOS. 1. 0 Introduction Previously developed models for predicting subjective video quality have one or both of the following: a) objective impairment measurements, each weighted by estimates of subjective annoyance, and summed to produce a single metric per frame and/or sequence [1] [2] and/or b) an attempt to at least partially mimic human vision system response [3] [4] [5] [6] [7] [8] [9]. However, vast stimulus-response data from vision science literature has remained largely unexploited. Models that have been developed within the vision science community typically only address isolated human vision system behaviors, or are parameterized as to not be readily adaptable to the application of video quality as exemplified by [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22]. As a result, current standards and practices for subjective video quality measurement ignore substantial changes in human vision sensitivities. For example, ITU-T J.144 models such as VQM [1] do not take into account differences in viewing conditions, frame rates, resolutions or nominal quality context. And since J.144 was developed for standard definition (SD), using it for predicting subjective video quality ratings of other formats such as CIF or HD is problematic [23]. There have been reports that use of J.144 methods can be problematic even with some SD video if content and/or impairments are quite different from that used for development and verification of the algorithms. In addition to human vision perceptual response issues, display device differences and the nominal quality of each format also play important roles [23]. Even for a given format, display and viewing conditions, application dependent nominal quality range (for example, set by bit rate or program content nominal complexity) can vary significantly. This has made necessary a) normalizing subjects to the range of quality of video in ITU-R BT.500 [24] and b) the bit rate and frame rate matrix of categories of VQEG impairment ranges and types respectively [25]. As such, the context of nominal quality range affects the sensitivity and quality rating scales such as MOS and DMOS. Using the simulation approach shown in the system diagram of Figure 1, video quality rating prediction requires a human vision model that can adapt as the human vision system does, to different displays, viewing conditions, video context and formats, along with a method for adapting quality scales as humans do in subjective quality rating training and conditioning. Such a human vision model has been developed and patented [26] [27] [28] [29]. An overview of its specification, components, calibration and validation are covered in the remaining sections of this paper. Fig. 1. Simulation system diagram for subjective video quality rating prediction. 2.0 Defining Human Vision Model Response Specifications The specifications of the model may be formulated from the data available from thousands of published results of experiments in vision science. Stimulus-response pairs become the numeric framework for describing the input-output behavior desired from the human vision model. A good introductory overview of experiments and associated stimulusresponse sensitivity analysis is given in [30]. For brevity, the scope of the remaining discussion will be limited to achromatic light (luminance) only. Analysis for chromatic light follows a similar path.