isyntax-to-tiff features

[jonasteuwen]

The following features would be useful in the isyntax-to-tiff

• Embed ICC profile
• Create our own downsampling

The first is self-explanatory; the second would mean that we create the pyramid ourselves from the tiff. Right now I'm using isyntax-to-tiff and subsequently use libvips to create dense pyramid. This solves the misalignment issue.

However, the downsampling is not really trivial, if we feel this is a good feature, does anyone have a suggestion for a lightweight downsample library (probably uses anti-aliasing too, I will check the libvips code). EDIT:
EDIT: Vips seems to be using the 'mean' operator: https://github.com/libvips/libvips/blob/159e8bea25fbc70316e32cd36add0b7c899568b7/libvips/foreign/tiffsave.c#L424C24-L424C47 that should be fine with stb_image.h

Would this be of interest? I can create it.

[Falcury]

I think this would be a great addition. My first thought also was to check how libvips does it, and try to do it that way.

The implementation of the mean shrink operator seems to be here:
https://github.com/libvips/libvips/blob/master/libvips/iofuncs/region.c#L1190

And here:
https://github.com/libvips/libvips/blob/master/libvips/iofuncs/region.c#L1149

I'm kind of surprised they don't use SIMD!

[jonasteuwen]

I've been playing around a bit with the 2x2 reduce operator (averaging), which seems to give good images (also to pathologists). That shouldn't be too hard to implement.

[Falcury]

@jonasteuwen I finally got around to playing around with resampling and reconstructing the pyramid from level 0 (in order to both solve the offsets issue and reduce recompression artefacts). For non-power-of-2 resizing the lanczos3 kernel seems to be a recommended choice, and for power-of-2 shrinking the 2x2 reduce/averaging.

I adapted code from Pilllow into Slidescape:
https://github.com/amspath/slidescape/blob/e3f5c5eb05c0562e5f08b49def02f27fd73d1280/src/core/image_resize.c

I used a recursive method to build up the pyramid without eating up too much RAM:
https://github.com/amspath/slidescape/blob/e3f5c5eb05c0562e5f08b49def02f27fd73d1280/src/tiff/tiff_write.c#L1597-L1603
https://github.com/amspath/slidescape/blob/e3f5c5eb05c0562e5f08b49def02f27fd73d1280/src/tiff/tiff_write.c#L1189

It would be a bit tricky to copy the code directly over into the isyntax-to-tiff utility though, because I'm using my own TIFF implementation instead of libtiff. (We can copy that over as well, but it would take a bit more work)