lmmx/page-dewarp

ZeroDivisionError: float division by zero

Closed this issue · 7 comments

I'm using page-dewarp that was installed from pip into venv on Ubuntu 24.04.

venv/bin/pip list
Package         Version
--------------- -----------
contourpy       1.2.1
cycler          0.12.1
fonttools       4.51.0
kiwisolver      1.4.5
matplotlib      3.8.4
mpmath          1.3.0
numpy           1.26.4
opencv-python   4.9.0.80
packaging       24.0
page-dewarp     0.1.5
pillow          10.3.0
pip             24.0
pyparsing       3.1.2
python-dateutil 2.9.0.post0
scipy           1.13.0
six             1.16.0
sympy           1.12
toml            0.10.2
tomlkit         0.12.5

It reports:

venv/bin/page-dewarp -d 3 ep.jpg 
Loaded ep.jpg at size='723x854' --> resized='362x427'
Traceback (most recent call last):
  File "/tmp/venv/bin/page-dewarp", line 8, in <module>
    sys.exit(main())
             ^^^^^^
  File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/__main__.py", line 21, in main
    processed_img = WarpedImage(imgfile)
                    ^^^^^^^^^^^^^^^^^^^^
  File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/image.py", line 55, in __init__
    self.contour_list = self.contour_info(text=True)
                        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/image.py", line 155, in contour_info
    return mask.contours()
           ^^^^^^^^^^^^^^^
  File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/mask.py", line 59, in contours
    return get_contours(self.name, self.small, self.value)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/contours.py", line 107, in get_contours
    contours_out.append(ContourInfo(contour, rect, tight_mask))
                        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/contours.py", line 60, in __init__
    self.center, self.tangent = blob_mean_and_tangent(contour)
                                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/contours.py", line 40, in blob_mean_and_tangent
    mean_x = moments["m10"] / area
             ~~~~~~~~~~~~~~~^~~~~~
ZeroDivisionError: float division by zero

ep

lmmx commented

Thanks for including the image, I reproduced this.

I've added test instrumentation (if you run uv sync you'll get development dependencies now, equivalent to uv pip install pysnooper, or if not using uv pip install pysnooper).

It arises from an all-zero moments

    14:41:36.505573 call        32 def blob_mean_and_tangent(contour):
    14:41:36.505844 line        39     moments = cv2_moments(contour)
    New var:....... moments = {'m00': 0.0, 'm10': 0.0, 'm01': 0.0, 'm20': 0.0,...u30': 0.0, 'nu21': 0.0, 'nu12': 0.0, 'nu03': 0.0}
    14:41:36.506094 line        40     area = moments["m00"]
    New var:....... area = 0.0
    14:41:36.506349 line        41     mean_x = moments["m10"] / area
    14:41:36.506605 exception   41     mean_x = moments["m10"] / area
    Exception:..... ZeroDivisionError: float division by zero

so we can breakpoint the bug with

    area = moments["m00"]
    if not area:
        breakpoint()

Then in PDB we can pprint the moments, which are all zero

(Pdb) pp moments
{'m00': 0.0,
 'm01': 0.0,
 'm02': 0.0,
 'm03': 0.0,
 'm10': 0.0,
 'm11': 0.0,
 'm12': 0.0,
 'm20': 0.0,
 'm21': 0.0,
 'm30': 0.0,
 'mu02': 0.0,
 'mu03': 0.0,
 'mu11': 0.0,
 'mu12': 0.0,
 'mu20': 0.0,
 'mu21': 0.0,
 'mu30': 0.0,
 'nu02': 0.0,
 'nu03': 0.0,
 'nu11': 0.0,
 'nu12': 0.0,
 'nu20': 0.0,
 'nu21': 0.0,
 'nu30': 0.0}
Click to show the value of contours giving rise to this zero moment
(Pdb) p contour
array([[[279, 351]],

       [[278, 352]],

       [[277, 352]],

       [[276, 352]],

       [[275, 352]],

       [[274, 352]],

       [[273, 352]],

       [[272, 352]],

       [[271, 352]],

       [[270, 352]],

       [[269, 352]],

       [[268, 352]],

       [[267, 352]],

       [[266, 352]],

       [[265, 352]],

       [[266, 352]],

       [[267, 352]],

       [[268, 352]],

       [[269, 352]],

       [[270, 352]],

       [[271, 352]],

       [[272, 352]],

       [[273, 352]],

       [[274, 352]],

       [[275, 352]],

       [[276, 352]],

       [[277, 352]],

       [[278, 352]],

       [[279, 351]],

       [[280, 351]],

       [[281, 351]],

       [[282, 351]],

       [[283, 351]],

       [[284, 351]],

       [[285, 351]],

       [[286, 351]],

       [[287, 351]],

       [[288, 351]],

       [[289, 351]],

       [[290, 351]],

       [[291, 351]],

       [[290, 351]],

       [[289, 351]],

       [[288, 351]],

       [[287, 351]],

       [[286, 351]],

       [[285, 351]],

       [[284, 351]],

       [[283, 351]],

       [[282, 351]],

       [[281, 351]],

       [[280, 351]]], dtype=int32)

If we do the opposite, and breakpoint on a regular (truthy) area with if area: breakpoint() (giving area = 3.0):

Click to show a regular contour
(Pdb) pp contour
array([[[ 98, 405]],

       [[ 97, 406]],

       [[ 96, 406]],

       [[ 95, 406]],

       [[ 94, 406]],

       [[ 93, 407]],

       [[ 92, 407]],

       [[ 91, 407]],

       [[ 90, 407]],

       [[ 89, 407]],

       [[ 88, 407]],

       [[ 87, 407]],

       [[ 86, 407]],

       [[ 85, 407]],

       [[ 84, 407]],

       [[ 83, 407]],

       [[ 82, 407]],

       [[ 81, 407]],

       [[ 80, 407]],

       [[ 79, 407]],

       [[ 78, 407]],

       [[ 77, 407]],

       [[ 78, 407]],

       [[ 79, 407]],

       [[ 80, 407]],

       [[ 81, 407]],

       [[ 82, 407]],

       [[ 83, 407]],

       [[ 84, 407]],

       [[ 85, 407]],

       [[ 86, 407]],

       [[ 87, 407]],

       [[ 88, 407]],

       [[ 89, 407]],

       [[ 90, 407]],

       [[ 91, 407]],

       [[ 92, 407]],

       [[ 93, 407]],

       [[ 94, 406]],

       [[ 95, 406]],

       [[ 96, 406]],

       [[ 97, 406]],

       [[ 98, 406]],

       [[ 99, 406]],

       [[100, 406]],

       [[101, 405]],

       [[102, 405]],

       [[103, 405]],

       [[104, 405]],

       [[105, 405]],

       [[106, 405]],

       [[105, 405]],

       [[104, 405]],

       [[103, 405]],

       [[102, 405]],

       [[101, 405]],

       [[100, 405]],

       [[ 99, 405]]], dtype=int32)

I'm using page-dewarp that was installed from pip into venv on Ubuntu 24.04.

venv/bin/pip list Package Version

contourpy 1.2.1 cycler 0.12.1 fonttools 4.51.0 kiwisolver 1.4.5 matplotlib 3.8.4 mpmath 1.3.0 numpy 1.26.4 opencv-python 4.9.0.80 packaging 24.0 page-dewarp 0.1.5 pillow 10.3.0 pip 24.0 pyparsing 3.1.2 python-dateutil 2.9.0.post0 scipy 1.13.0 six 1.16.0 sympy 1.12 toml 0.10.2 tomlkit 0.12.5

It reports:

venv/bin/page-dewarp -d 3 ep.jpg Loaded ep.jpg at size='723x854' --> resized='362x427' Traceback (most recent call last): File "/tmp/venv/bin/page-dewarp", line 8, in sys.exit(main()) ^^^^^^ File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/main.py", line 21, in main processed_img = WarpedImage(imgfile) ^^^^^^^^^^^^^^^^^^^^ File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/image.py", line 55, in init self.contour_list = self.contour_info(text=True) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/image.py", line 155, in contour_info return mask.contours() ^^^^^^^^^^^^^^^ File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/mask.py", line 59, in contours return get_contours(self.name, self.small, self.value) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/contours.py", line 107, in get_contours contours_out.append(ContourInfo(contour, rect, tight_mask)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/contours.py", line 60, in init self.center, self.tangent = blob_mean_and_tangent(contour) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/contours.py", line 40, in blob_mean_and_tangent mean_x = moments["m10"] / area ~~~~~~~~~~~~~~~^~~~~~ ZeroDivisionError: float division by zero

lmmx commented

I looked it up and found something that might be informative/similar:

The issue is that cv2.moments() has a bug and OpenCV contours are weird.

lmmx commented

I think the underlying issue is that a single bad contour should be omitted rather than spoiling the entire operation. We are not writing an OCR algorithm, only acquiring contours to provide to the dewarping routine: not comprehensively defining all text line contours is not going to harm this overall operation.

As the traceback shows:

  File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/image.py", line 55, in __init__
    self.contour_list = self.contour_info(text=True)
                        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/image.py", line 155, in contour_info
    return mask.contours()
           ^^^^^^^^^^^^^^^
  File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/mask.py", line 59, in contours
    return get_contours(self.name, self.small, self.value)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/contours.py", line 107, in get_contours
    contours_out.append(ContourInfo(contour, rect, tight_mask))
                        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/tmp/venv/lib/python3.12/site-packages/page_dewarp/contours.py", line 60, in __init__
    self.center, self.tangent = blob_mean_and_tangent(contour)
                                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

We are:

  • [in image.py] building a contour_list attribute on the WarpedImage object (at init time) using its contour_info() method, with contour type text=True, which is just a wrapper on...
  • [in mask.py] converting a Mask to contours via its contours() method, which in turn just passes along to...
  • [in contours.py] running the get_contours() function, which uses cv2.findContours() and then postprocesses the contours received to drop any unsuitable ones before appending to the contours_out variable which ends up going back to the WarpedImage.contour_list

This last step is where the fix seems most suitable:

  • We are already looping over a contour list and skipping 'bad eggs' with continue
  • The simplest solution is to try to append and pass on the errors in the else block (rather than halt the entire program)

We could also do it without side effects (i.e. without allow an error to throw), by checking for zero area and returning some sentinel value like (None, None) which we then check for in get_contours, but catching the ZeroDivisionError seems proactive and clear enough

lmmx commented

I ran it after allowing those errors to be suppressed (and their contours dropped) and the result is pretty poor.

zerodivisionerror_image_thresh

I suspect that the error is a hint that you may want to use a larger input image, as then the contours will be detected by OpenCV, and you'll get more accurate text line angles and subsequent dewarping. I expect this result would be unusable.

Perhaps a better approach would be to print a message to user when skipping these zero-moment contours suggesting to use a larger image (or perhaps sample the image more densely? I forget which parameters do this, I think there's a setting for that).

lmmx commented

The result with a 2x enlarged input is better, it still crops off one side though

input:

zerodivisionerror_image_2x_enlarged

output:

zerodivisionerror_image_2x_enlarged_thresh

lmmx commented

Ah, I see the problem: the left margin is not being included! If you run the program with the -d (debug level) flag set to 3 you will produce all intermediate steps for debugging:

zerodivisionerror_image_debug_4_keypoints_before.png

zerodivisionerror_image_debug_4_keypoints_before

zerodivisionerror_image_debug_5_keypoints_after.png

zerodivisionerror_image_debug_5_keypoints_after

That explains why enlarging the image does not resolve the issue (I ran it again on the enlarged one to confirm):

zerodivisionerror_image_2x_enlarged_debug_4_keypoints_before.png

zerodivisionerror_image_2x_enlarged_debug_4_keypoints_before

zerodivisionerror_image_2x_enlarged_debug_5_keypoints_after.png

zerodivisionerror_image_2x_enlarged_debug_5_keypoints_after

lmmx commented

The assumption of this program is that you will be supplying a page with a margin, which in this case is a suitable assumption for the right side but not the left.

The flag for this is -x (see page-dewarp --help for all flags, I just refactored these to use defaults from this module)

In this case my intuition that the small image was too small seems to be correct, the "keypoints" (i.e. samples taken from the text line contours whose detection we were debugging above) are notably sparse and irregular when the program is run with -x 0 (no left/right page margin)

zerodivisionerror_image_debug_4_keypoints_before (with -x 0 flags)

zerodivisionerror_image_debug_4_keypoints_before

...resulting in another poor result:

zerodivisionerror_image_debug_6_output

If I use the enlarged version it works nicely:

zerodivisionerror_image_2x_enlarged_debug_4_keypoints_before.png

zerodivisionerror_image_2x_enlarged_debug_4_keypoints_before

zerodivisionerror_image_2x_enlarged_debug_5_keypoints_after.png

zerodivisionerror_image_2x_enlarged_debug_5_keypoints_after

...leading to a nice result:

zerodivisionerror_image_2x_enlarged_debug_6_output.png

zerodivisionerror_image_2x_enlarged_debug_6_output