Reverse engineering the Sony a7RII long exposure spatial filtering

We have seen here and here that the Sony a7RII running firmware 3.30 spatially filters raw images prior to writing them to the flash card at shutter speeds slower than 3.2 seconds. The camera did similar filtering for bulb exposures in earlier releases.

A DPR poster earlier proposed that the algorithm Sony used was the one described here.

He has since amended his algorithm to read as follows (if you look at the post above, the pseudo-code below will make sense):

If (V0 > Max(V1, V2, ... , V8))  
   V0 = Max(V1, V2, ... , V8)
Else If (V0 < Min(V1, V2, ... , V8))
   V0 = Min(V1, V2, ... , V8) )
Endif

Another DPR poster, the same one who brought the FW 3.30 behavior to my attention, kindly prepared for me two images. The first was 1000×1000 pixels in size, and filled with Gaussian noise with mean 128 and standard deviation 8. The second was the first image as processed by the algorithm above.

I ran both images through my spectral analysis program.

This precisely replicates the behavior that I’m seeing on the actual camera.

To refresh your memory:

It is certainly possible that it is just a coincidence that the proposed “star-eater” algorithm produces the same spectra as whatever Sony is doing in-camera, but I don’t think so. This is probably the real deal, or close enough not to matter.

You could certainly see why Sony would want to tame the hot pixels that often crop up in long exposures, but IMHO they should have provided a way to turn it off. The camera is too good for astro work to do this to it.

And, of, course, they should have told us what they were doing, but that would be un-Sony-like.