Reduce for resize #4273

homm · 2019-12-17T01:27:34Z

This PR is a logical conclusion of #4251 and #4231 (and also the old one #2254) and shouldn't be merged first. All the new code in the single commit: fa9e85c

With all these tools (.reduce() method and box argument for .resize() method) we finally have an extremely precise image resize approximation which works in times faster on a very large image reduce.

Performance

In [1]: from PIL import Image 
   ...: im = Image.open('../imgs/space.jpeg') 
   ...: im.load() 
   ...: 'Size: {}x{}'.format(*im.size)
Out[1]: 'Size: 4928x3280'

In [2]: %timeit im.resize((411, 273), Image.BICUBIC, reducing_gap=None)
75.6 ms ± 331 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [3]: %timeit im.resize((411, 273), Image.BICUBIC, reducing_gap=5.0)
36.7 ms ± 295 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [4]: %timeit im.resize((411, 273), Image.BICUBIC, reducing_gap=4.0)
42 ms ± 192 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [5]: %timeit im.resize((411, 273), Image.BICUBIC, reducing_gap=3.0)
27.4 ms ± 343 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [6]: %timeit im.resize((411, 273), Image.BICUBIC, reducing_gap=2.0)
18 ms ± 39.1 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [7]: %timeit im.resize((411, 273), Image.BICUBIC, reducing_gap=1.0)
11.4 ms ± 127 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

With a very realistic value (reducing_gap=2.0) this method speed ups resizing in 4.2 times. This like Pillow-SIMD but without SIMD :-)

It performs even better with more heavy LANCZOS filter (5.3 times faster):

In [10]: %timeit im.resize((411, 273), Image.LANCZOS, reducing_gap=None)                                                          
109 ms ± 239 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [11]: %timeit im.resize((411, 273), Image.LANCZOS, reducing_gap=5.0)                                                           
48.5 ms ± 253 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [12]: %timeit im.resize((411, 273), Image.LANCZOS, reducing_gap=4.0)                                                           
50 ms ± 238 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [13]: %timeit im.resize((411, 273), Image.LANCZOS, reducing_gap=3.0)                                                           
30.7 ms ± 280 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [14]: %timeit im.resize((411, 273), Image.LANCZOS, reducing_gap=2.0)                                                           
20.3 ms ± 281 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [15]: %timeit im.resize((411, 273), Image.LANCZOS, reducing_gap=1.0)                                                           
12 ms ± 84 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

Accuracy

try_reduce.py

from PIL import Image

im = Image.open('../imgs/space.jpeg')
im.load()
print('Size: {}x{}'.format(*im.size))

out_size = (320, 213)

im.resize(out_size, Image.BICUBIC, reducing_gap=None).save('_bic.ref.png')
for reducing_gap in range(1, 4):
    im.resize(out_size, Image.BICUBIC, reducing_gap=reducing_gap).save(f'_bic.{reducing_gap}.png')
    
im.resize(out_size, Image.LANCZOS, reducing_gap=None).save('_lzs.ref.png')
for reducing_gap in range(1, 4):
    im.resize(out_size, Image.LANCZOS, reducing_gap=reducing_gap).save(f'_lzs.{reducing_gap}.png')

Bicubic

Reference (reducing_gap=None):

reducing_gap=3.0 and inverted difference (yes, there is the second image):

reducing_gap=2.0 and inverted difference (wipe your screen):

reducing_gap=1.0 and inverted difference:

Lanczos

Reference (reducing_gap=None):

reducing_gap=3.0 and inverted difference:

reducing_gap=2.0 and inverted difference:

reducing_gap=1.0 and inverted difference:

So the difference between fair resampling and reducing_gap=2.0 is indistinguishable by eye. reducing_gap more than 3.0 is simply meaningless.

Bonus track: the inverted difference between BICUBIC and LANCZOS:

TODO list

tests
use box argument from reduce() method to reduce image region
release notes

homm · 2019-12-17T02:01:09Z

src/PIL/Image.py

        return m_im

-    def resize(self, size, resample=NEAREST, box=None):
+    def resize(self, size, resample=NEAREST, box=None, max_reduce=None):


Maybe the better name is reduce_gap since this is actually minimal distance (in terms of scale) between the reduce() method result size and the requested size.

Renamed to reducing_gap

homm · 2019-12-17T02:01:27Z

src/PIL/Image.py

-        if res is not None:
-            box = res[1]
+        if max_reduce is not None:
+            res = self.draft(None, (size[0] * max_reduce, size[1] * max_reduce))


On the one hand, thumbnail() now takes max_reduce into account when working with JPEGs and achieves indistinguishable quality like when it works with any other images.

homm · 2019-12-17T02:01:39Z

src/PIL/Image.py


        if self.size != size:
-            im = self.resize(size, resample, box=box)
+            im = self.resize(size, resample, box=box, max_reduce=max_reduce)


On the other hand, thumbnail() takes max_reduce into account when working with any other images and achieves compared speed as when working with JPEGs.

Huge win!

src/PIL/Image.py

# Conflicts: # docs/releasenotes/7.0.0.rst # src/PIL/Image.py

# Conflicts: # docs/releasenotes/7.0.0.rst

docs/releasenotes/7.0.0.rst

src/PIL/Image.py

hugovk

A few minor suggestions.

hugovk · 2019-12-29T16:53:42Z

docs/releasenotes/7.0.0.rst

+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Speeds up resizing by resizing the image in two steps. The bigger ``reducing_gap``,
+the closer the result to the fair resampling. The smaller ``reducing_gap``,


What is meant by "fair resampling"?

This means resampling in one operation, as before, without .reduce() operation.

docs/releasenotes/7.0.0.rst

src/PIL/Image.py

# Conflicts: # docs/releasenotes/7.0.0.rst

homm added 2 commits December 17, 2019 02:27

Merge branch 'box-in-thumbnail' into reduce-in-resize-2

2db5406

max_reduce arguments in resize and thumbnail

fa9e85c

homm added the Do Not Merge label Dec 17, 2019

homm commented Dec 17, 2019

View reviewed changes

radarhere reviewed Dec 18, 2019

View reviewed changes

src/PIL/Image.py Outdated Show resolved Hide resolved

homm added 3 commits December 20, 2019 14:54

Use box for reduce

b85c610

rename max_reduce to reducing_gap

b48996c

TestReducingGapResize

a9d7a2b

homm marked this pull request as ready for review December 20, 2019 17:20

homm added 2 commits December 20, 2019 20:27

Merge branch 'reduce' into reduce-in-resize

8181aed

# Conflicts: # docs/releasenotes/7.0.0.rst # src/PIL/Image.py

fix formatting

fc28182

homm force-pushed the reduce-in-resize branch from c02d48c to fc28182 Compare December 20, 2019 17:42

homm added 2 commits December 21, 2019 03:16

add release notes

59f8eeb

Merge branch 'reduce' into reduce-in-resize

d719aaf

# Conflicts: # docs/releasenotes/7.0.0.rst

v-atamanenko reviewed Dec 27, 2019

View reviewed changes

homm removed the Do Not Merge label Dec 27, 2019

homm added 2 commits December 27, 2019 15:13

Merge branch 'master' into reduce-in-resize

222c2f6

wording

1bd9b16

hugovk reviewed Dec 29, 2019

View reviewed changes

hugovk added the Enhancement label Dec 29, 2019

homm added 2 commits December 30, 2019 03:23

texts fixes

34029e9

Merge branch 'master' into reduce-in-resize

c74c20e

# Conflicts: # docs/releasenotes/7.0.0.rst

homm merged commit c3232e5 into python-pillow:master Dec 30, 2019

homm deleted the reduce-in-resize branch December 30, 2019 14:30

Uh oh!

Reduce for resize #4273

Reduce for resize #4273

Uh oh!

Conversation

homm commented Dec 17, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Performance

Accuracy

Bicubic

Lanczos

TODO list

Uh oh!

homm Dec 17, 2019

Choose a reason for hiding this comment

Uh oh!

homm Dec 20, 2019

Choose a reason for hiding this comment

Uh oh!

homm Dec 17, 2019

Choose a reason for hiding this comment

Uh oh!

homm Dec 17, 2019

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

hugovk left a comment

Choose a reason for hiding this comment

Uh oh!

hugovk Dec 29, 2019

Choose a reason for hiding this comment

Uh oh!

homm Dec 30, 2019

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

homm commented Dec 17, 2019 •

edited

Loading