Crop solid background colours (e.g. black or white borders)

[garygreen]

At the moment there doesn't appear to be any way to hook into imagehash to remove solid colours from the image - this causes two images with a heavy black border to look similar.

It would be great if there was a way to filter the image before the perceptual hash is generated.

Take for example these two images - notice the black border on either side. This causes the two images to look 76% similar simply due to the black borders.

GD provides a imagecropauto which I believe will auto crop out the black / white borders, maybe this could be used?

[qathom]

Hi @garygreen
I agree with you - but I am not sure if imagehash should handle this.
If you ask me, it's more an image preprocessing task.

I also faced this issue and I created this helper "remove_constant_colors()".
It will remove all constant colors, all repetitive pixel colors in both vertical and horizontal directions.
By applying this preprocessing task, the image looks like this:

When I compare both images, I get a result of hash difference value of 313 which is very high so different images.

Let me know what you think! :-)

import cv2
import numpy as np
from PIL import Image
import imagehash

def remove_constant_colors(path: str):
    image = cv2.imread(path)
    image_h, image_w, nb_col = image.shape
    
    selection_rows = np.invert(np.all(image == image[0], axis = 0))
    image_1 = image[:,selection_rows[:,0]]
    
    image_1_h, image_1_w, nb_col = image_1.shape
    
    selection_cols = np.invert(np.all(image_1 == image_1[0], axis = 1))
    image_2 = image_1[selection_cols[:,0]]

    return image_2

# Usage: hash comparison between the 2 images
path_1 = 'test.jpg'
path_2 = 'test2.jpg'
image_hash_1 = imagehash.average_hash(Image.fromarray(remove_constant_colors(path_1)), hash_size=32)
image_hash_2 = imagehash.average_hash(Image.fromarray(remove_constant_colors(path_2)), hash_size=32)

diff_hash = image_hash_1 - image_hash_2

print('Diff hash', diff_hash) # Returns 313

[omidMolaverdi]

I guess the easy way is to use imagecropauto(). This can be embedded into the hash function as well?