recognize: skip tiny or bin-empty lines, too

[bertsky]

No description provided.

[mikegerber]

While I sympathize with these changes: Is this filter not something that would be useful in other processors, too? I.e. go to the core library?

@kba Your thoughts on this?

[mikegerber]

It could also go to a generic filter-cleaner processor.

[bertsky]

It could also go to a generic filter-cleaner processor.

Indeed. It makes sense to do that as a postprocessing step, depending on the workflow (not all segmenters will need it). I could easily add something along these lines to ocrd-segment-repair.

Until then, let's consider this a robustness feature. (ocrd_tesserocr already does the same filtering.)

[mikegerber]

Yeah, I had intended to merge it, I just need to run a quick test later

[kba]

While I sympathize with these changes: Is this filter not something that would be useful in other processors, too? I.e. go to the core library?

Such a filter would make sense to have in core but where would we add it? To the get_TextLine() method in the PAGE API?

[bertsky]

Such a filter would make sense to have in core but where would we add it? To the get_TextLine() method in the PAGE API?

If in core as part of OcrdPage, I'd rather place it on the producer side: TextLineType constructor and .set_Coords(). (Where it's easy to check the bbox dimensions, but checking for binary-empty is challenging – it would involve looking up the line's binarized AlternativeImage, or its parent's etc.) Perhaps even just CoordsType constructor and .set_points() (which we already patch for another reason).

Also, what about the GdsCollector mechanism?