benhoyt / dhash Goto Github PK

These have stopped working, as you can see from this run. On both Windows and macOS it looks like 2.7 is no longer supported (maybe we should just remove the tests on 2.7). But all the macOS tests seem to be failing, seemingly due to ImageMagick not being installed on the latest macos runners. This can probably be fixed with a brew install imagemagick.

I run pip install dhash in windows 10 OS then I get dhash version 1.3, finally when I run import dhash the AttributeError occured. Is any thing wrong?

testpictures.zip

I attached the picture in the zip file, one is icon type, and the other is png type
dhash of the icon type returns a value of 0
dhash of the icon type returns a value of 62738665417926919091179733734494077055

I tried both 1.4 and 1.3 dhash version. Also tried both PIL and wand

do you have any comment about this? Do I need to covert icon picture to png first to get a meaningful value? Thanks

I was using another image hash lib (https://pypi.org/project/ImageHash/) which does not do this.
Is this XY-128bit pattern a convention or something?

Hello I use your code since a bit while and I find it pretty awesome and useful, I'm using either to detect copies and also for a project to regenerate a picture as a mozaic of other pictures, first I was using dHash and PIL, then use the basic deltaE with Lab channels to compare overall color difference of two images and went to use OpenCV for lesser and faster comparison. And then I facepalm myself, thinking ...

Maybe it'd be a good idea to do dHash on each Lab channels but I'm not sure how to do it right, the L is the BW equivalent which values goes from 0 to 100 of dHash and a&b are two orthogonal (independant) chroma with a going from -128 to 127 -> green to red and b -128 to 127, thankfully openCV already normalize each value from 0 to 100 after conversion from RGB.

So I'd like to have your opinion

• Is Luminance is as important as color (or twice) ?
• Should we compute two separate matrix difference, one for luminance and one for colour ?
• What about to port this algorithm to opencv ? (I've started to work on it)

I've just been playing with your library and it kept telling me that images were very different when they looked the same. It seems that it was due to the color present in transparent areas.

For example; if I create two images as:

from PIL import Image, ImageDraw

# greyscale image, completely white and also completely transparent
im1 = Image.new('LA', (100, 100), (0xff, 0))

im2 = im1.copy()
# replace most of it with "transparent black"
ImageDraw.Draw(im2).rectangle((10, 10, 90, 90), (0, 0))

and then run dhash on them:

from dhash import dhash_row_col
print(dhash_row_col(im1), dhash_row_col(im2))

I get very different results, even though they'll be rendered the same as far as most viewers are concerned. They're obviously "different" images and you can use an image editor to pull the channels apart and show this. But given that this library is about perceptual differences I think it should probably do something about this.

I'm working around this by filling transparent areas with white, as suggested in:

python-pillow/Pillow#2609 (comment)

def fill_transparent(image, fill_color='white'):
  if image.mode in ('RGBA', 'LA'):
    cleaned = Image.new(image.mode[:-1], image.size, fill_color)
    cleaned.paste(image, image.split()[-1])
    return cleaned
  return image

print(*[dhash_row_col(fill_transparent(im)) for im in [im1, im2]])

Not sure if this is deliberate, or if/how it should be exposed to users of the library — hence opening an issue rather than pull request.

The function compares two INT hashes, not any hash the library calculates. I think it is very unintuitive. At least the function have to either convert the inputs to be right, or provide an explanation on what it expects.

I noticed this while running tests on a project of mine and thought I'd bring it up:

  /long/path/dhash.py:66: DeprecationWarning: ANTIALIAS is deprecated and will be removed in Pillow 10 (2023-07-01). Use Resampling.LANCZOS instead.
    small_image = gray_image.resize((width, height), PIL.Image.ANTIALIAS)

Referring to the resize operation here:

    elif PIL is not None and isinstance(image, PIL.Image.Image):
        gray_image = image.convert('L')
        small_image = gray_image.resize((width, height), PIL.Image.ANTIALIAS)
        return list(small_image.getdata())

This was found while using pillow==9.2.0 and dhash==1.3

Cheers!

the type of return value of dhash.format_hex() is str, maybe you can change the type of return value to hex. And the users needn't to convert it when they use dhash.get_num_bits_different().