collate chars w/ slight doctop variations about pdfplumber

[jsvine]

Great question / conundrum!

Not too concerned about the value-alteration. This can be handled easily by — as you allude to — making a .copy() of chars. And I think snap_to_y_grid handles most cases like these.

But how should something like this handle edge-cases, e.g., a word gently-sloping-down? For the sake of a concrete-if-lame example:

letter	doctop
C	20
O	21
M	22
P	23
U	24
T	25
E	26
R	27

You'd have to set a pretty aggressive snap_to_grid value to clump all those letters together. But the letter-to-letter variation in doctop is actually pretty small — just one pixel at a time. I wonder:

• Is there a clever way to algorithmically identify strings of near-to-each-other characters? (And one that isn't too computationally expensive?)
• Should we even be worrying about edge cases like these?

Eager for your thoughts, and thanks again for raising the issue!

[jsvine]

Alternative attempt at solving this, using doctop-clustering:

• Code: https://github.com/jsvine/pdfplumber/blob/doctop-variations/pdfplumber/utils.py
• Diff: fd79cc8

Essentially, it looks at all doctops, sorts them, then clusters any doctops within y_tolerance of the next-smallest doctop. Then, it treats all doctops in the same cluster as being on the same line. Would seem to handle your situation — does it? — and my fake example above. What do you think?

And do you have any PDF (or CSV char-output) of previously-failing examples?

[jsfenfen]

I think this would work--lemme dig around for the one that was failing. This is actually a problem for me in just coalescing words for merging with some fonts--all cap fonts sometimes have larger first letters which also makes trouble. I may try to adapt this to that.

The question of sloping letters is also a really good one--and pretty gnarly to deal with once the delta y approaches line height. I've mucked around with that a little, and I think you need a different approach than clustering. For text that's just slanted I did an experiment with linear regression, and choosing words that had a shared y-intercept that sorta worked. But the other problem one gets (much more with scanned text) is of non-linear curve (think of putting a book on a copy machine). But that's an even worse problem...

[jsvine]

all cap fonts sometimes have larger first letters which also makes trouble

Interesting! Hadn't encountered that before, but makes total sense.

For text that's just slanted I did an experiment with linear regression, and choosing words that had a shared y-intercept that sorta worked.

Clever! I'm a bit worried about the computational expense of that, at least for the main collate_chars function. Maybe worth breaking out more-advanced chars-in-a-line detection into its own method?

But the other problem one gets (much more with scanned text) is of non-linear curve (think of putting a book on a copy machine). But that's an even worse problem...

Oy, yeah. I'm thinking handling super-crazy text layouts is outside the scope of this library. But definitely open to being convinced otherwise!

[jsfenfen]

Yeah, wrt angle detection that's definitely something you wouldn't want there. If you assume a constant angle (not a great assumption) you can do some detection stuff and then apply your discovered angle elsewhere.

[jsvine]

The new strategy for handling doctop-variations is now in master and v0.2.0 — pushed and published this morning.

[jsfenfen]

Awesome!