Image characterization preprocessor

[kba]

As discussed in our Open Tech Call, there is interest in the community for a pre-processor that validates certain assumptions on the images and possibly offers solutions to alleviate them. This arose from a discussion around reliable calculation of DPI (cf. #676).

Maybe we can use this thread to gather requirements before moving forward with the implementation.

Functionality such a processor should have:

• Determine the pixel density in DPI for all images
  • Parameters:
    • min-dpi: Lower threshold to reject an image
    • estimate-dpi: Whether to use heuristics on image size and page size to estimate DPI
    • scale: Whether to scale images to fake a higher DPI
• Determine the hardware capabilities of the underlying system
  • Can this image be processed with the RAM or GPU memory available?

Feedback and proposals are welcome!

[bertsky]

IIRC we also wanted to validate image formats here, i.e.

• whether the image containers/codecs themselves are permitted,
• whether the image metadata are valid and consistent (e.g. PNG header vs EXIF info),
• whether the image dimensions are permitted for a certain workflow or processing node (as lower or upper limit),
• whether the image colours are allowed (e.g. alpha channels, CMYK and others)

The idea was to get a fail-early-and-loud mechanism, but as opt-in (by using this processor as first step in the workflow). Another idea might be to conveniently repair, if possible. (This aspect applies to DPI meta-data as well).

[bertsky]

BTW, DPI estimation is non-trivial and already implies some kind of line segmentation dependency (like ocrolib) IMO.

Strategically, we could approach this as a processor in core with some base functionality (without DPI estimation, or a very crude one), but hooks for easy overriding. Other modules (like ocrd_cis) could then subclass from the image characterization in core to provide their own.