Why are 2 cropped areas producing same .extract_text() output when one is "empty"?

[cmdlineluser]

I was experimenting with the example pdf from #926

page4.pdf

Cropping 2 areas produce the same output from .extract_text()

import pdfplumber

pdf = pdfplumber.open("page4.pdf")
page = pdf.pages[0]

a = page.crop((489.0136, 510.25570000000005, 533.1974, 521.686))
b = page.crop((489.0136, 521.686, 533.1974, 529.686))

a.extract_text()
# '人民幣百萬元'

b.extract_text()
# '人民幣百萬元'

If we look at the cropped areas as images, b is "empty"

A.png

B.png

From what i can tell, matrix seems to be the unique identifier for an object, so these are all the first character:

c1 = a.chars[0]
c2 = b.chars[0]
c3 = next(c for c in page.chars if c['matrix'] == c1['matrix'])

>>> c1
{'matrix': (8.075, 0.0, 0.0, 9.5, 482.4896, 287.3948),
 'fontname': 'NKVMXY+MHeiHK-Bold',
 'adv': 1.0,
 'upright': True,
 'x0': 489.0136,
 'y0': 285.6183,
 'x1': 490.5646,
 'y1': 295.1183,
 'width': 1.5509999999999877,
 'height': 8.930299999999988,
 'size': 9.5,
 'object_type': 'char',
 'page_number': 1,
 'text': '人',
 'stroking_color': (0, 0, 0, 1),
 'non_stroking_color': (0, 0, 0, 1),
 'top': 512.7557,
 'bottom': 521.686,
 'doctop': 512.7557}

>>> c2
{'matrix': (8.075, 0.0, 0.0, 9.5, 482.4896, 287.3948),
 'fontname': 'NKVMXY+MHeiHK-Bold',
 'adv': 1.0,
 'upright': True,
 'x0': 489.0136,
 'y0': 285.6183,
 'x1': 490.5646,
 'y1': 295.1183,
 'width': 1.5509999999999877,
 'height': 0.5697000000000116,
 'size': 9.5,
 'object_type': 'char',
 'page_number': 1,
 'text': '人',
 'stroking_color': (0, 0, 0, 1),
 'non_stroking_color': (0, 0, 0, 1),
 'top': 521.686,
 'bottom': 522.2557,
 'doctop': 521.686}

>>> c3
{'matrix': (8.075, 0.0, 0.0, 9.5, 482.4896, 287.3948),
 'fontname': 'NKVMXY+MHeiHK-Bold',
 'adv': 1.0,
 'upright': True,
 'x0': 482.4896,
 'y0': 285.6183,
 'x1': 490.5646,
 'y1': 295.1183,
 'width': 8.074999999999989,
 'height': 9.5,
 'size': 9.5,
 'object_type': 'char',
 'page_number': 1,
 'text': '人',
 'stroking_color': (0, 0, 0, 1),
 'non_stroking_color': (0, 0, 0, 1),
 'top': 512.7557,
 'bottom': 522.2557,
 'doctop': 512.7557}

The only thing that seems to differ is the top/bottom/doctop values:

c1['bottom'] - c1['top']
# 8.930299999999988

c2['bottom'] - c2['top']
# 0.5697000000000116

c3['bottom'] - c3['top']
# 9.5

I'm just wondering if someone can explain to me what is happening here and how this works.

I guess the char is considered inside if any point falls within cropped area?
And then its position values are "truncated" to those that fit within the area?

I want to discard both of the crops as "false matches" (crop A only has the very edge of the first char).

It seems we can find the original matrix in page.chars and if any of the values differ then we can classify it as a false match?
Update: (Hm, no - it appears they always differ - even when fully contained?)

Thanks all.

[jsvine]

I guess the char is considered inside if any point falls within cropped area? And then its position values are "truncated" to those that fit within the area?

Yep, that's exactly it. From the readme:

Cropped pages retain objects that fall at least partly within the bounding box. If an object falls only partly within the box, its dimensions are sliced to fit the bounding box.

I acknowledge that this can be confusing at first, especially with char objects whose bounding boxes extend beyond their visual markings. On the other hand, I wanted crop to adhere to what "crop" means in other contexts/software. It also has some more "expected"/preferred results with things like line objects.

The "good" news: You can use page.within_bbox(...), which is "similar to .crop, but only retains objects that fall entirely within the bounding box."

[cmdlineluser]

Ah, thank you for the explanation @jsvine - .within_bbox is what I was missing.