a7x, a9 PDAF stripe noise technical analysis, part deux

Started 9 months ago | Discussions thread
Antisthenes Regular Member • Posts: 294
Re: First PDAF Pixel interpolation discovery

ProfHankD wrote:

Horshack wrote:

Interpolation is likely weighted and bayer influenced

I doubt that the interpolation is so complex. [..]

It would be difficult to deduce the formula, but I'd bet I could get it by running a GA (genetic algorithm) to find the most consistent stencil weightings over a set of test images with known PDAF pixel locations. [..]

Of course, this still doesn't directly explain the stripe artifacts, does it?

I think the excellent simulations posted by "Tons o Glass 0 Class" explain the striped artifacts quite well.

The idea is that light might reflect off the (shiny ?) partial metal masks covering the PDAF pixels.

Assume, for the sake of argument, that only the blue pixel locations have been allocated to PDAF pixels.
Now, imagine an incident ray of light is coming from the lower left and enters the Bayer pixel array at pixel B22.

R11 G12 R13 G14
G21 B22 G23 B24
R31 G32 R33 G34
G41 B42 G43 B44

That ray of light might get reflected by B22's metal mask, and then be reflected back by the microlens layer towards, say, the R13 pixel.

I think it's more likely that the microlenses cause the back-reflection, rather than the sensor package's cover glass. This is because the cover glass likely lies several hundred microns away from the sensor's surface.

A zig-zag involving the PDAF masks and the cover glass would thus cover a distance of several hundred microns times two.

Even a tiny variation of the ray angle would then presumably cause, at the end of the zig-zag, a difference in distance of tens of microns as to where the reflected ray lands back on the sensor.

Even a minuscule variation of ray angle should thus cause reflected rays to variously land on R, G or B pixels at the end of their zig-zag.

For each picture in which stripe artifacts have been observed, however, it seems that these artifacts are strongly confined to one particular Bayer group — e.g. only the R, G1 or G2 groups.

This suggests that the zig-zag reflection path is much shorter than several hundred microns, making it much less sensitive to ray incidence angle variations.

A scenario whereby the light bouncing off the metal PDAF mask is then reflected back by the microlens layer, which sits only a few microns above the silicon, seems more likely to cause the reflected light to consistently land on a particular Bayer group — e.g. like the ray of light coming from the lower left in my example above, and reflected by B22 and causing an artifact in pixel R13.

It might be interesting to move a white LED — emitting a mix of R, G and B light — circularly around the lens mount of a Sony camera, thereby changing the spatial direction in which light is reflected by the PDAF masks, and thus causing the the artifacts to migrate, say, from Bayer group R to Bayer group G1 or G2 as the LED travels around the mount's periphery.

Another possibility is that the light is reflected by the fairly dense and smooth metal interconnect layers sitting below the silicon (in a BSI sensor), and that the reflected lightray then bounces off the underside of the PDAF metal mask, and ultimately lands in the photodiode area of an adjacent pixel, thereby causing an artifact.

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