1.8.8-2 iMac drizzle weight issue?

[mschuster]

Hi Juan,

I see an unexpected artifact in drizzle weights, x1, square, drop shrink 1. The artifact affects the drizzle integrated result. The artifact does not appear if drop shrink is 0.9.

All files (~300 MB):
https://www.dropbox.com/s/4a30qbj24oqud3i/Drizzle.zip?dl=0

Edit: Process icons:
https://www.dropbox.com/s/haxy7tsljmd2e1b/DrizzleProcessIcons.xpsm?dl=0

Drizzle weights, with "block structured" artifact:


Difference image, showing that artifact affects the result (difference created by the PixelMath icon):

[mschuster]

I just added a link to the processing icons, which I forgot to include.

[Juan Conejero]

Hi Mike,

Bug confirmed. I have reproduced it with other data sets that I have just checked. At first glance this looks like a 'simple' coordinate roundoff error, but I have to investigate it in depth. I've probably caused this bug with the latest improvements I've made to my drizzle implementation (in core version 1.8.7), where I have made slight changes to the way drizzle applies coordinate systems internally, in order to make it fully compatible with regular image integration.

Workaround: set drop shrink to 0.99. This fixes the problem completely, and is a good hint that this is just a numeric roundoff issue.

Thank you for discovering this bug. I'll work to fix it as soon as possible.

By the way, this image looks terrific with drizzle x2

[mschuster]

Hi Juan,

Thank you. I have two other questions:

1) Notice that the satellite trail shows in the difference image as a subtle level change along the trail. Does this indicate a small normalization issue related to rejection in drizzle?

2) Note that in this example ImageIntegration uses bilinear interpolation. As a test, please change it to Nearest Neighbor and retry the entire process. You will see that the satellite trail is now much more prominent in the drizzle integrated result. I think this is an issue also. Note, I recall several years ago NN being an issue with drizzle, maybe it is just not supported by drizzle?

[Juan Conejero]

This bug is now fixed:

https://gitlab.com/pixinsight/PCL/commit/a1e5997d57e2feff61d98645e4edf732c6f6a711

The fixed version of the ImageIntegration module will be included in version 1.8.8-3, which I am preparing now as a cumulative update/bugfix release for all platforms.

1) Notice that the satellite trail shows in the difference image as a subtle level change along the trail. Does this indicate a small normalization issue related to rejection in drizzle?

Not at all. With a drop shrink value of 1.0, the satellite trail has pixel values close to 0.9474 on the drizzle weights image. This is approximately equal to 1 - 1/19, which corresponds to the expected weight when only one image of the set of 19 integrated images has been rejected.

2) Note that in this example ImageIntegration uses bilinear interpolation. As a test, please change it to Nearest Neighbor and retry the entire process. You will see that the satellite trail is now much more prominent in the drizzle integrated result. I think this is an issue also. Note, I recall several years ago NN being an issue with drizzle, maybe it is just not supported by drizzle?

Nearest neighbor interpolation should not be used for drizzle integration. This is because NN does not provide subpixel interpolation accuracy, which is crucial for the drizzle algorithm, even for a drizzle scale of 1. Ideally we should issue at least a warning message when this happens in drizzle, but this would require storing additional metadata in .xdrz files.