MOSp human visual system (hvs) no reference vs. full reference model profiles as a calibration method OUTLINE1.
The human visual system (hvs) is sophisticated. In fact, the hvs detects digital visualartifacts in video without prior knowledge of “good”, or original/source video.Moreover, the hvs can make value judgments on whether quality is artificiallyimpaired a little, or a lot.With this theory as the basis for modeling the perceived quality of distribution video,a properly trained software analysis engine should provide accurate measurement ofvideo without source formats.v.Cortex works like the hvs:It knows MPEG impairments and video.It knows quality.It does not rely on the original video.It is a powerful No Reference Model1. hvs1. Model licensed by BT2.
source encode encoded 1 −1 2 = ( ) =0 (10 log10 (255 /D(n)) vs.1001010MPEG1001001011100010101111000010 Perceptual quality video models attempt to achieve a perfect correlation between Mean Opinion Scores(MOS) and Perceived Mean Opinion Scores(MOSp). In a full reference model, the source material image). Within the bit stream, v.Cortex focuses must be available to the analyzer. A popular, and on the Quantizer Step-size. This parameter is simple, calculation is PSNR; though, this significant in an MPEG encoder’s capacity to measurement provides little information reduce picture data. regarding viewer perception. It serves a industry accepted quantitative difference that does not Also of concern, the human visual system may take encoder type into consideration. not see significant impairments if the video content masks these impaired regions. To ensure Other full reference models seek to determine proper interpretation of MPEG impairment, impairment types between source and encoded v.Cortex applies pixel domain methods to assess material. Models weight these impairment types the levels of masking and combines these to provide a prediction of viewer quality estimates with its bit stream calculations. assessment. continued… v.Cortex utilizes a no reference model that full reference vs. takes advantage of the MPEG (H.264 and MPEG-2) bit stream and pixel domain (decoded 3. no reference
MOSp MOSFigure 1 | MOS vs. MOSp continued from p.3 v.Cortex produces a MOSp that is closely While v.Cortex accurately and quickly assesses the correlated, at .91, to MOS (fig. 2). The model quality of MPEG video, network operators objectively provides consistent and accurate require methods, or filters, to quickly provide results. In fig.1, the data indicates the close pass/fail status to videos that have been relationship between v.Cortex’s MOSp and actual evaluated. The next section evaluates the use of MOS values for a variety of quality levels. It is filters with MOSp to save significant time in important to emphasize that this data reflects video quality assurance. distribution quality MPEG video via standard consumer TV, as perceived by consumers. ITU MOS 5 Excellent The research also indicates that future v.Cortex analyzers may be calibrated to closely predicate 4 Good MOS values of H.264 video delivered to mobile 3 Fair devices, such as iPad, Blackberry and Android 2 Poor handheld machines. Some test metrics suggest that the high resolution and small screens of 1 Bad mobile devices may cause a near linear upshift of Figure 2 | ITU MOS MOS performance. Simply stated, a 2Mbps full reference vs. standard definition video may appear of higher quality on a smaller, but full resolution, screen. 4. no reference
v.Cortex produces accurate MOS results. Operators can calibrate v.Cortex by analyzing aHowever, video network operators must few good and bad files from their libraries, thencalibrate their customer perceptions with their apply profiles that meet these expectations.network capabilities and transport delivery v.Cortex compliments an operator’s perspectivemodel. Many cable operators seek to deliver and identifies quality problems without anthree MPEG-2 HD videos per 6Mhz 256-QAM operator viewing each video.channel. However, with an average of 12Mbpsper video, this may be insufficient to reach For example, broadcast graphics, which aredesired quality levels. Operators must have a difficult to encode, may create a trough, whichmethod to easily find adverse MPEG effects to can be identified by v.Cortex, and either ignoredmake a subjective decision about quality. or caught for closer review. In another case, an operator may tolerate movies at an averagev.Cortex incorporates operator-controlled filters MOSp of 3.2, but will want to ensure that nofor: bad scenes, troughs, and black/white more than 5% of the scenes are below 2.5. Withintensity detection. Each combination of filters a calibrated profile, the operator can ensure thatcan be saved as a profile, and applied to different all movies are of an appropriate MOS value andvideo types (sports, news, movies, etc.). Bad operator quality standard, without viewing eachscenes are collections of frames that result in video. Finally, v.Cortex can be customized toundesirable MOSp results over a specified ensure operator intervention is minimized whileduration. Troughs are extreme, but short, low customer quality is achieved.points in MOSp. Black and white intensityidentifies areas where video may be missing dueto improper production problems. calibration 5. through profiles