Question about Partial Decode

[william-silversmith]

Hi Randy,

This isn't exactly a feature request, I just don't have a firm grasp of the PNG format yet and was wondering if the following is possible.

I’ve been interested in getting the image formats I work with to allow for as close to random access as possible. The regular PNG format is based on deflate, and so is not random access. However, ordinary (not interlaced) Progressive PNG allows for skipping the decoding of trailing scanlines. Is it at all possible to skip decoding some leading scanlines to achieve a vertical crop? Is it possible to do this horizontally as well?

I recently introduced the spng decoder into google/neuroglancer#370 which is a scientific viewer for petascale 3d neuroscience data when I realized PNG beats gzipping raw arrays by about 25% on storage. A significant fraction of researchers in my field are paranoid about lossy compression, so I've been trying to find ways to accommodate this desire. I've been trying to optimize the ability to make rapid queries of large datasets by partially decoding images. It's not essential to do this, but it would be interesting if it is possible.

Thanks for the wonderful library Randy!

[randy408]

Image data is always encoded as a single DEFLATE stream, interlacing doesn't change that.

I have a proposed specification for subdividing image data which would allow random access: https://github.com/libspng/png-restart-marker

And to leverage that for multi-threading (#192) but all of this would need funding to implement.

[william-silversmith]

Thanks again for the quick reply Randy! I think I am confused on a point. Isn't the purpose of progressive PNG to allow dynamic loading over a slow connection? Doesn't that imply the existence of multiple DEFLATE encoded sections that can be accessed independently? Or is it that you can start decoding the single DEFLATE stream in an on-line fashion and extract each scanline?

Thanks for pointing me to the work you've been considering. If I am correct about the on-line decoding of the deflate stream, I can see the benefit of the restart markers.

I haven't applied for this myself, but I've heard good things about this grant program for essential open-source imaging software: https://chanzuckerberg.com/eoss/ I think a new funding cycle should start within a few months though I'm not tapped into their internal conversations. Check back in periodically and you might be able to apply. This is certainly a very useful library. I would suspect you'd have a competitive application.

[randy408]

Interlacing just rearranges the order of pixels to allow for "better" progressive display, you get a lower resolution preview early on as opposed to loading the full resolution image from top to bottom. It's still a single DEFLATE stream that is decompressed linearly. See the wikipedia page on Adam7 interlacing for an illustration (.GIF's don't embed here).

[william-silversmith]

Thank you so much Randy! That explanation makes it clear. Please feel free to close this issue unless it is otherwise helpful for you.