nicolashahn / diffimg

Differentiate images in python - get a ratio or percentage difference, and generate a diff image

diffimg

Get the % difference in images using PIL's histogram + generate a diff image. Images should have the same color channels (for example, RGB vs RGBA). If the image dimensions differ, the 2nd image will be resized to match the first before calculating the diff.

Installation

Now available from PyPi: pip install diffimg

Usage

>>> from diffimg import diff
>>> diff('mario-circle-cs.png', 'mario-circle-node.png')
0.007319618135968298

The very simple diff function returns a raw ratio instead of a % by default.

diff(im1_file, 
     im2_file, 
     delete_diff_file=False, 
     diff_img_file=DIFF_IMG_FILE
     ignore_alpha=False)

im1_file, im2_file: filenames of images to diff.

delete_diff_file: a file showing the differing areas of the two images is generated in order to measure the diff ratio with the same dimensions as the first image. Setting this to True removes it after calculating the ratio.

diff_img_file: filename for the diff image file. Defaults to diff_img.png (regardless of inputed file's types).

ignore_alpha: ignore the alpha channel for the ratio and if applicable, sets the alpha of the diff image to fully opaque.

As command line tool

python -m diffimg image1 image2 [-r/--ratio] [-d/--delete] [-f/--filename DIFF_IMG_FILE]

--ratio outputs a number between 0 and 1 instead of the default Images differ by X%.

--delete removes the diff file after retrieving ratio/percentage.

--filename specifies a filename to save the diff image under. Must use a valid extension.

--ignore-alpha ignore the alpha channel.

Tests

$ ./test.py
......
----------------------------------------------------------------------
Ran 6 tests in 0.320s

OK

Formula

The difference is defined by the average % difference between each of the channels (R,G,B,A?) at each pair of pixels A_xy, B_xy at the same coordinates in each of the two images (why they need to be the same size), averaged over all pairs of pixels.

For example, compare two 1x1 images A and B (a trivial example, >1 pixels would have another step to find the average of all pixels):

A_1,1 = RGB(255,0,0) (pure red)

B_1,1 = RGB(100,0,0) (dark red)

((255-100)/255 + (0/0)/255 + (0/0)/255))/3 = (155/255)/3 = 0.202614379

So these two 1x1 images differ by 20.2614379% according to this formula.

Sample image 1

Sample image 2

Resulting diff image

Difference percentage output

Images differ by 0.731961813597%

diffimg's Issues

Is there anyway to output a color-image of differ instead of black-white?

For example,
A_1,1 is (255,64,0)
B_1,1 is (80,160,0)
so output differ is
Out_1,1 (|255-80|,|64-160|,|0-0|)=(175,96,0)

Add another diff function that can take PIL.Images instead of filenames

So we can do diffs on in-memory images instead of having to read from disk

test_different_image_sizes: very slightly different ratio on python 3.12

the least significant digit of the ratio result in that test changes when using python 3.12

(it was pointed out here (also shows my fix) that it may be a good idea to consolidate that info here)

Difference ratio calculated incorrectly

Hey! Thanks for this repo, it pointed me in the direction of PIL which I could use to solve a similar sort of problem 😄.

One thing I noticed was that the difference was calculated based on a factor of 100 as the maximum difference:

diffimg/diffimg/__init__.py

Lines 46 to 51 in 5a5293d

	# Calculate difference as a ratio.
	stat = ImageStat.Stat(diff_img)
	# Can be [r,g,b] or [r,g,b,a].
	sum_channel_values = sum(stat.mean)
	max_all_channels = len(stat.mean) * 100
	diff_ratio = sum_channel_values/max_all_channels

But for normal RGB the values are in the range 0-255, so the difference could be up to 255. You can repro this with a script that generates black and white images and diffs them (which should print 1.0):

>>> import diffimg
>>> import PIL
>>> PIL.Image.new('RGB', (10,10), (0,0,0)).save('a.png')
>>> PIL.Image.new('RGB', (10,10), (255,255,255)).save('b.png')
>>> diffimg.diff('a.png', 'b.png')
2.55

Different size not taken into account. Is it possible to scale one to the format or size of the other?

Hello,

I came accross your module and find it quite useful. For my usecase I would need a feature to compare even images with different sizes because I receive pictures from different sources showing the same scene. I would like to see these images to be considered "similar" (lower percentage or ratio).

Would that be possible?

Kind regards

Jens

difference image is not convenient for png

I am using this tool to measure the difference between two png files. The alpha of each pixel of both images is 255. Consequently the diff image produced by ImageChops.difference() is completely transparent. This is not convenient to locate differences between the two images.