Advanced image obfuscation techniques

[DonaldTsang]

Would like to ask if imagehash can handle these types of edited images

Goal: Update on https://github.com/pippy360/transformationInvariantImageSearch

1. Brightening/Darkening
2. Stretched/Smushed and Rotated
3. Upscaled/Downscaled and Blurred
4. Hue Rotations and Color Inversion
5. Greyscaling and Saturation manipulation

Comparison: https://github.com/kennethrapp/phasher

[Lucassifoni]

Hi @DonaldTsang,
From my experience with this library :

• Brightening/Darkening : It really depends on the levels
• Stretched/Smushed : Not much degradation
• Rotated : Intense degradation depending on the angle
• Upscaled/Downscaled and Blurred : Almost no degradation except at really strong blur levels
• Hue Rotations : Not too much degradation
• Color Inversion : Intense degradation
• Greyscaling and Saturation manipulation : Not too much degradation

Since the underlying algorithm works with luminance averages in regions, you could visualize it as looking at two images while squinting. What is "the same" image here means spatially-wise, not "featuring the same contents".

Have a nice day

[DonaldTsang]

@Lucassifoni let's see how this could be fixed

• Brightening/Darkening and Blurring: justifiable that as long as it is not too strong it can be handled
• Rotated: Affine Invariant hashing, or hashing at multiple angles, would be the best solution
• Color Inversion: Can't be helped something to be dealt with

[Lucassifoni]

I feel that the rotation/color inversion cases could be handled outside this library.

Conceptually, an inverted image's hash is the complement of the original's hash because of the luminance sampling technique, so you could test for inversion with count_bits(~hash1 ^ hash2) < n where n is your target.

Rotation, the same way, could be handled by rotating the hash itself, since the hash translates to a spatial representation :

abcd1234efgh5678 

-> 2d

a b c d
1 2 3 4
e f g h
5 6 7 8

-> rotate 90° clockwise

5 e 1 a
6 f 2 b
7 g 3 c
8 h 4 d

-> 1d

5e1a6f2b7g3c8h4d

But this feels beyond the scope of this library, since the question it tries to solve is "are those two images roughly the same", and an inverted or rotated image is visually not the same as the original...