Better Pixels with Fewer Defects on the Way
Ravi Velhal of Intel and the man that was involved in Intel’s development of 8K processing around the Olympics often talks about the need for better pixels as well as having more pixels. That means better colors and fewer artifacts and errors as well as higher refresh rates and dynamic range.
This week we saw two news items that highlighted the desire for even better pixels.
An 8K Pioneer in the News
The first was from another 8K pioneer, Dr Masaoka of NHK. He was the scientist that led the tests that established 8K as the level of resolution needed to bring a level of quality that was indistinguishable from reality. He has continued to work in the visual field and we bumped into him at the SID Display Week in San Jose this year. When we spoke, he talked about work that was then doing to clarify what he regarded as some incorrect, but widely accepted, color science. Now, in a letter to the Journal of the SID, he has detailed that work.
Some Background
Human color perception is a very complex topic but to try to simplify color, scientists developed a system, codified by the CIE, a standards body, to analyze color with 3 dimensions. One dimension was luminance, or the level of brightness, while the other two (x,y) identified the hue. This was standardized in 1931. The range of colors that were visible to most people were shown in the a diagram and a color space shape that is very familiar to anyone involved with color.
However, there was a problem associated with the diagram. In some areas of the chart, humans are able to differentiate very tine differences in the numbers that defined a color, while in others it needed relatively big changes. To correct for this, two additional color spaces, CIELUV (CIE 1976) and CIELAB were developed. These were intended to be ‘perceptually uniform’ color spaces, where the same movement in the chart indicates the same level of color difference. There are mathematical functions that have been defined to convert between these color spaces.
(All of the above is acknowledged as a gross simplification!)
Now, much of the development of the CIELUV and CIELAB spaces were based on work done in the 1940s by David MacAdam. This work identified the non-linearity of CIE 1931 (left in the image below) and showed how the chromaticity space used in 1976 (right in the image below) introduced less distortion.
In his letter Dr Masaoka argues that the two 1976 color spaces are not that uniform because they are relatively insensitive to luminance difference which can enhance color perception change. Masaoka also highlights weaknesses in MacAdam’s methodology and says that “Subsequent experiments have cast doubt on their reproducibility”. He also criticizes the small number of samples. His conclusion is that the u’v’ space is actually less uniform than the original 1931 space – a surprising conclusion!
Gamut Rings
Dr Masaoka is the key developer of concept of gamut rings to characterize display performance. These rings give a much more useful sense of the performance of a display at different levels of luminance than chromaticity diagrams. The gamut rings have also allowed a new color space which is more uniform.
From Theory to Practice
While the letter from Dr Masaoka is really based on color theory, Netflix has published a blog post explaining the techniques it has been using to improve the efficiency of its quality control systems for video, Up to now, the process of checking video quality for ‘stuck’ dark or light pixels has been largely manual and very time consuming.
The blog points out that to reduce the processing needed, a lot of automated vision models use down-sampled data. However, with pixel level errors, downsampling would cause the errors to disappear. This means the checking has to be done at full resolution.
Netflix developed a system based on a single GPU that can process video in real time to detect errors. The process is based on machine learning but, of course, to develop a good algorithm, you need to train it. As pixel errors are relatively rare, the team had to generate artificial errors to allow training and the article details this.
So, the quest for better pixels as well as more pixels continues!
