Hamming distance is returning {:ok, 0} always.

[jhonathas]

I have 3 images to test, image 1 and 2 are similar, and I expected the result to be less than or equal to 10. Image 3 is completely different. I'm always comparing image 1 to image 2 or image 2.

Image 1

Image 2

Image 3

My code is:

  {:ok, vix_image_1} = Vix.Vips.Image.new_from_file("./tmp/ap-images/image_1.png")
  {:ok, vix_image_2} = Vix.Vips.Image.new_from_file("./tmp/ap-images/image_2.png")
  {:ok, vix_image_3} = Vix.Vips.Image.new_from_file("./tmp/ap-images/image_3.png")

  {:ok, hash_image_1} = Image.dhash(vix_image_1)
  {:ok, hash_image_2} = Image.dhash(vix_image_2)
  {:ok, hash_image_3} = Image.dhash(vix_image_3)

  Image.hamming_distance(hash_image_1, hash_image_2) |> IO.inspect(label: :compare_image_1_and_image_2)
  Image.hamming_distance(hash_image_1, hash_image_3) |> IO.inspect(label: :compare_image_1_and_image_3)

The result:

compare_image_1_and_image_2: {:ok, 0}
compare_image_1_and_image_3: {:ok, 0}

I've also tried comparing the variables with vix_image_1 and vix_image_2, and vix_image_1 and vix_image_3, but the result is the same.

Does anyone have any idea what could be the problem?

[kipcole9]

While I believe that the implementation of Image.dhash/2 is basically the same as the Rubygem dash-vips gem, the fact that the image is reduce to 8px by 8px means that there isn't much information to work on. I think thats the primary reason, but it doesn't sit easily with me.

Thats partially why I implemented Image.compare/3. There are some examples of its usage on Elixir forum an no doubt more tutorial information is required. I hope you might find Image.compare/3 a better way to compare images for similarity.

However if the idea is to store an image hash in a database so you can do similarity comparisons in SQL then thats not going to work.

I would welcome any help on an improved image hash mechanism and I will revalidate the implementation as best I can.

Let me know if Image.compare/3 does meet - or does not meet - your needs. If not, I will aim to revise Image.hash/2 over the next few days.

[jhonathas]

I even tested Image.compare too, but, unless the image was exactly the same, it always returned something around 0.99... and the strangest thing was when I just took a screenshot of one of the images, but in smaller proportions, and when comparing, the number was closer to 1.0 than another version with the image from another angle but from the same location. Furthermore, I will need to store the hashes in the database so that I can generate reports of similar images.

[kipcole9]

Arrgghhh, it does look like there is an aggegious error in Image.dhash/2 which I will investigate ASAP. I have opened an issue.

[jhonathas]

Ahh, okay. I'm hoping that everything goes well. Thank you for your attention.

[kipcole9]

I've published Image version 0.38.4 that fixes this issue.

Image hashes for your test images

Each image now correctly generates a dhash. However note the limitations of image hashing. Whilst the human eye can detect the similarities between the first two images, hashing does sees quite a big difference.

# Open images
iex> i1 = Image.open!("/Users/kip/Desktop/image_1.png")
%Vix.Vips.Image{ref: #Reference<0.513719039.295043093.140240>}
iex> i2 = Image.open!("/Users/kip/Desktop/image_2.png")
%Vix.Vips.Image{ref: #Reference<0.513719039.295043093.140245>}
iex> i3 = Image.open!("/Users/kip/Desktop/image_3.png")
%Vix.Vips.Image{ref: #Reference<0.513719039.295043094.141349>}

# Calculate image hashes
iex> Image.dhash i1
{:ok, <<143, 155, 120, 254, 187, 185, 205, 102>>}
iex> Image.dhash i2
{:ok, <<252, 248, 239, 210, 146, 96, 96, 152>>}
iex> Image.dhash i3
{:ok, <<3, 31, 135, 142, 55, 135, 7, 7>>}

# Calculate hamming distance. Hamming distance is just a count
# of the bits that are different between two hashes. This means that
# of the 64 bits in the hash, 37 of them are different.
iex> Image.hamming_distance i1, i2
{:ok, 37}

[jhonathas]

Thanks, now it works.