The Unappreciated Test Behind the Development of PQ

Yet another in our intolerable series “No One Asked But We Answer It Anyway!” 😝

Everyone is aware that 12-bit PQ ST2084 was devised using the Barten ramp to ensure that contouring artefacts in the BT.2020 standard were below the visible threshold. But did you know that PQ as the EOTF is only applicable to monochrome images and that for color images, another metric was needed?

The Art of Better Pixels, D. G. Brooks, Dolby Laboratories Inc., USA. Conference paper at IBC 2014

EDR Colour Volume Quantization

When researchers at Dolby Laboratories set out to determine the bit depth necessary to accurately represent an EDR colour volume of 0-10,000 nits BT.2020, they realized that resolving this question would require an investigation into both luminance and chrominance elements.

Note: Video combining both HDR & WCG is defined as extended dynamic range, or EDR

Luminance Quantization

Researchers at Dolby Labs came up with the Perceptual Quantizer (PQ) which follows the Barten curve to ensure the optimal use of the bits available. It can be seen that 12-bit PQ is below the visible threshold for contouring artefacts.

Colour Quantization

“Whilst using PQ as the EOTF can provide the solution for monochrome images, an additional metric is required to ensure that colour images are also artefact free. The JND Cross test was developed for this purpose.”

“Each square in Figure 7 represents a block of pixels which have been perturbed by 1 code word away from the grey background. All possible variations of this in RGB result in 26 different colour patches.”

“If the image looks like a flat grey image, then there is no visible quantization. If some patches are visible, then some contouring or banding is likely.”

“This test is then performed for multiple different grey levels. The industry accepted metric to judge the visibility of these patches is CIEDE2000 colour difference formula.”

“Subjective tests by Nezamabadi et al. indicate that for noise free images such as animation, CGI and graphics a DE2000 threshold of ~2.5 is needed to avoid contouring artefacts. When noise is present, this threshold increases to ~5.”

“It should also be noted that the DE2000 metric starts to lose accuracy below about 3 nits; this corresponds to the point where the human visual system begins to switch between photopic to scotopic vision.”

“By combining the use of this metric with the JND colour patches allows the quantification of possible colour quantization artefacts. The results are shown in Figure 9.”

EDR Colour Volume Conclusions

“For noise free EDR images such as animation, CGI or graphics, using 12-bit PQ ensures that both monochromatic and colour banding artefacts are below the visible threshold. For captured images that contain some noise, using a 10-bit PQ representation is adequate, but care needs to be taken, especially with graphics.”