alisinabh / nx_color

Toolkit for managing colorspaces in Elixir Nx to use in Machine Learning.

NxColor

Colorspace conversion in Elixir Nx.

With NxColor you can convert images (to be precise, pixels) between various colorspaces like RGB, CMYK or CIE-Lab.

Since this is implemented using Elixir Nx, It should be really fast.

Installation

NxColor can be installed by adding nx_color to your list of dependencies in mix.exs:

def deps do
  [
    {:nx_color, "~> 0.0.1-dev"}
  ]
end

Usage

Using libraries like stb_image or using an already loaded image or images in an Nx.Tensor, You can import them into NxColor using the NxColor.from_nx function.

stb_image = StbImage.read_file!("image.jpg")

image =
  stb_image
  |> StbImage.to_nx()
  |> NxColor.from_nx(colorspace: NxColor.Colorspace.RGB, channel: :last)

Then you can use NxColor.change_colorspace to actually convert the colorspace of the image.

You can see the list of all available colorspaces in the documentation.

image = NxColor.change_colorspace(image, NxColor.Colorspace.CMYK)

After you are done with the conversions, You can convert your image back to an Nx tensor.

tensor = NxColor.to_nx(image, channel: :last)

The channel option

Usually there are two ways of representing images in tensors.

Let's say we have a 32 x 32 image in the RGB colorspace. Each pixel in that colorspace is represented by 3 different number. Red Green and Blue.

The first way is if we encode the colors (the channel) as the first dimension in each image. So the shape for our image tensor would be: {3, 32, 32}

The second way is if we encode the colors (the channel) as the last dimension in each image. So the shape of our image tensor in this case would be: {32, 32, 3}

In NxColor you can choose between these two based on you input data and also your use-case. Simply set the channel option in NxColor.from_nx and NxColor.to_nx either to :first or :last.

nx_color's Issues

Add more testing cases using scikit image

We need more test cases. The best scenario is if we grab a few real images and convert their colorspaces using scikit image then save them as the results of our test cases.

Add HSL colorspace

References

• https://www.easyrgb.com/en/math.php
• http://www.brucelindbloom.com/index.html?Eqn_RGB_XYZ_Matrix.html

Add Hunter-LAB Colorspace

References

• https://www.easyrgb.com/en/math.php
• http://www.brucelindbloom.com/index.html?Eqn_RGB_XYZ_Matrix.html

Add HSV colorspace

References

• https://www.easyrgb.com/en/math.php
• http://www.brucelindbloom.com/index.html?Eqn_RGB_XYZ_Matrix.html

Add grayscale colorspace

Grayscale should be the L value from CIE Lab or CIE Lch

Add CIE-L*uv colorspace

References

• https://www.easyrgb.com/en/math.php
• http://www.brucelindbloom.com/index.html?Eqn_RGB_XYZ_Matrix.html