a7x, a9 PDAF stripe noise technical analysis, part deux

Started 2 months ago | Discussions thread
Horshack Veteran Member • Posts: 6,138
Evidence leads back to original reflection theory

The original theory was reflections, a second theory was based on presumption of G1 PDAF pixels from the magic raw. I'm now very confident the original theory was actually correct (with revision to PDAF layout), based on new evidence over the past two days and examining all previous evidence.

Here is the revised PDAF layout for the reflection theory (vs original layout) - the only change is that PDAF are now exclusively on B pixels, vs B/G2 previously. I've also drawn out each of the various reflection-striping conditions, including one which includes red channel striping for the first time.

This change is based on the continuing lack of evidence for any interpolation on the G2/G channels. It's also based on finally being able to get striping on the red channel, the absence of which was an impediment to the reflection theory. I did this via a slight variation of the no-lens light reflection setup, where I angled the camera on both axis. Here is an image of the striping - it is extremely faint but it's there:

Red channel striping

Here is a full statistical examination of all Blue PDAF pixels on the A7rII, including correlation checks between columns:

Blue PDAF stats

The above analysis proves the following:

  • For each repeating three-column set, column 0 has normal stdev while columns 1,2 have are much lower stdev. This demonstrates that each Blue PDAF row has a pattern of N-P-P (Normal Pixel, PDAF pixel, PDAF pixe, ...l)
  • There is no stdev correlation between pairs of Blue PDAF pixels, which means they are likely not interpolated with each other. They are instead likely interpolated from neighboring non-PDAF blue pixels and neighboring G/G2 pixels

PDAF striping theory

Directional light rays (sample) around a bright light source partially reflects off Blue PDAF pixels and onto neighboring cells. For typical on-axis lighting the amount of brightening from the PDAF reflections averages between 2.7% and 3.5% (linear measurement in ADUs of affected pixels vs their unaffected peers). For off-axis lighting such as that caused by lens flare/reflections the PDAF reflection brightness can reach 20% or more - for off-axis lighting in no-lens experiments the brightness can reach 70%.

The PDAF blue pixels themselves are also brightened by the PDAF reflections but due to their values being interpolated from the neighboring pixels that are being brightened off their mask reflections (vs being brightened directly). The blue pixels are brightened less vs their reflected peers likely due to the algorithmic color mixing ratios of interpolating luminance from their reflected green peers.

I haven't yet established what CFA the blue PDAF pixels are using - the patent describes both a clear CFA and color-specific. I have a set of color filters on the way which should allow this to be determined from experiments. I also haven't yet being able to reverse-engineer the blue PDAF pixel interpolation.

There is evidence for and against both main theories but right now the balance is tilted toward the reflection theory.

Special thanks to Hank for reproducing the striping on his A7 - seeing his red-channel stripes lead me to the bi-axis experiment to finally find striping on the red channels for the A7rII as well. And ongoing thanks to Jim for his striping plots, measurements, and sensor knowledge. And Bill for his discovery of the blue PDAF rows.


More data is needed to determine if the A7III is more prone to visible striping vs other Sony bodies. The original dpreview samples had large lens reflections, which yielded striping intensities of 20%. Either the A7III and/or 85mm F/1.8 FE lens used is more prone to lens flare reflections, or it was just bad luck. More A7III samples will be needed to determine which.

Post (hide subjects) Posted by
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow