replace opencv-python with scikit-image for snowball detection

[zacharyburnett]

Closes spacetelescope/jwst#7409

This PR adds fallback functions for replaces ellipse construction using scikit-image and shapely, removing opencv-python entirely. The scikit-image + shapely methods produce slightly different contours, ellipses, and circles than opencv-python.

Checklist

• added entry in CHANGES.rst (either in Bug Fixes or Changes to API)
• updated relevant tests
• updated relevant documentation
• updated relevant milestone(s)
• added relevant label(s)

[zacharyburnett]

Contour-finding also needs to be ported from opencv-python to scikit-image:

ModuleNotFoundError: `opencv-python` required for this functionality

def find_circles(dqplane, bitmask, min_area):
        # Using an input DQ plane this routine will find the groups of pixels with at least the minimum
        # area and return a list of the minimum enclosing circle parameters.
        pixels = np.bitwise_and(dqplane, bitmask)
        if ELLIPSE_PACKAGE == 'opencv-python':
            contours, hierarchy = cv.findContours(pixels, cv.RETR_EXTERNAL, cv.CHAIN_APPROX_SIMPLE)
            bigcontours = [con for con in contours if cv.contourArea(con) >= min_area]
            circles = [cv.minEnclosingCircle(con) for con in bigcontours]
        else:
            # raise ModuleNotFoundError(ELLIPSE_PACKAGE_WARNING)
           raise ModuleNotFoundError('`opencv-python` required for this functionality')

[codecov]

Codecov Report

Attention: Patch coverage is 94.21488% with 14 lines in your changes missing coverage. Please review.

Project coverage is 85.10%. Comparing base (bd72ed5) to head (1fa5161).
Report is 2 commits behind head on main.

Files with missing lines	Patch %	Lines
src/stcal/jump/circle.py	87.23%	12 Missing ⚠️
tests/test_circle.py	96.72%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #138      +/-   ##
==========================================
+ Coverage   84.38%   85.10%   +0.71%     
==========================================
  Files          41       43       +2     
  Lines        7557     7733     +176     
==========================================
+ Hits         6377     6581     +204     
+ Misses       1180     1152      -28     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[zacharyburnett]

As described in #126 (comment), there are some differences in the methods used to construct ellipses between the two packages which result in different image masks. Perhaps a warning should make this more clear.

looking into replacing shape construction with scikit-image, it appears there is a crucial difference in the drawing method:

import cv2
import numpy
from matplotlib import pyplot
from skimage import draw

image_0 = numpy.full((10, 10), fill_value=0, dtype=numpy.uint8)

center = (4, 4)
radius = 4

disk = draw.disk(center, radius)
image_1 = image_0.copy()
image_1[disk] = 1

image_2 = cv2.circle(image_0.copy(), center, radius, 1, thickness=-1)

figure, axes = pyplot.subplots(nrows=1, ncols=2)

axes[0].imshow(image_1)
axes[0].set_title('scikit-image')

axes[1].imshow(image_2)
axes[1].set_title('opencv')

pyplot.show()

[zacharyburnett]

It looks like the contours found by scikit-image are consistently larger, by about half a unit, than those found by opencv:

>       assert circle[0][1] == pytest.approx(1.0, 1e-3)
E       assert 1.5 == 1.0 ± 1.0e-03

I assume this is a difference in methodology in thinking about pixels as blocks vs as points.

[zacharyburnett]

expanding the ellipse test case to a more oblong shape reveals a few edge cases in the scikit-image method; I'll convert this to draft for now and work on that

[zacharyburnett]

added some more complexity to the test shape and fixing issues in the ellipse fitting:

[zacharyburnett]

here is where the fit circle from the scikit-image contour differs from the opencv-python circle; the radius is 1.5 instead of 1

[zacharyburnett]

Currently, the .minAreaRect() is replaced by .minimum_rotated_rectangle from Shapely; this algorithm could be written without Shapely, like in circle.py, to remove the need for the additional dependency

[zacharyburnett]

we should also get input from INS on whether this alternative is acceptable

[zacharyburnett]

I don’t understand what I’m looking at

Sorry about that! I'll update the comments to be more descriptive

[perrygreenfield]

To use your first simple example. I find that if I give the radius for draw.disk as 4.0001 I get the same image as cv2. So it wouldn't surprise me that in one the test is an < test and the other <=. Or something like that. Adding a small epsilon may do it.

[zacharyburnett]

a comparison of the find_contours functions between opencv and skimage:

and the minimum_bounding_box analogues:

[zacharyburnett]

started regression test run against this branch at https://plwishmaster.stsci.edu:8081/blue/organizations/jenkins/RT%2FJWST/detail/JWST/2925/pipeline

[drlaw1558]

Hm, I'm testing this on jw02123001001_09101_00001_nrs1 and something doesn't seem to be working right. I started out by merging the latest main branch in to make sure everything was synced, and then tried processing jw02123001001_09101_00001_nrs1

Runtime was significantly longer for the Scikit version; 743 seconds vs 314 seconds for the current opencv

More problematically though, the Scikit version doesn't seem to be flagging snowballs properly. It's flagging lots of little things that don't need attention, but missing the big snowballs. Perhaps I'm running the code the wrong way? In both cases I'm just using strun to execute detector1 pipeline, with the default expand_large_events=True

[zacharyburnett]

Thanks for testing! Where is jw02123001001_09101_00001_nrs1? I can run detector1 and debug that; it should have similar results to opencv, so if it doesn't then something is wrong

[zacharyburnett]

oh I can search that on MAST! Nevermind I'll use that, thanks anyway! 😅