kmeans-colors

Calculate the k average colors in an image using k-means clustering.

k-means clustering works by starting with an initial random guess of the k number of colors in the image called centroids. For each step, every pixel in the image is traversed to find what centroid is closest to it in color. Then, the centroids calculate the average of all the colors close to them and move to that color. This process repeats until the centroids stop moving or the maximum step count is reached.

The k-means can be used to find the dominant colors or color palette of an image. Some other applications are binarization, line extraction, and "color style" transfer. The crate wraps around a generic k-means implementation.

Examples

1) Basic usage

kmeans_colors -i gfx/pink.jpg -k 2 -r 3 -o pink2

The animation is a composite of k=2 to k=9 k-means with the above command. -k is the number of colors to find in the image. -r is the amount of runs to perform, -o specifies the output. By default, the images will save as .png files. The -o option is not required.

2) Color palettes

kmeans_colors -i gfx/lanterns.jpg --no-file --palette

kmeans_colors -i gfx/pink.jpg --no-file --palette --proportional

kmeans_colors -i gfx/flowers.jpg --no-file --palette --proportional --sort

By default, palettes will be composed of equal swatches. Passing --proportional will scale the swatches proportionally to their presence in the image. The default sorting method is from darkest to lightest, passing --sort will rearrange the palette in order from most frequent to least frequent color. The --height and --width of the palette can be specified as well as output name with --op. Passing -k 1 will produce the average color of the image.

3) The find subcommand

a) Binary Ferris Example

We can use k-means to clean up this doodle and extract the line work from it. The paper this is on is folded and scribbled over with highlighter and blue pencil, the back of the sheet also has ink on it.

kmeans_colors -i gfx/ferris.jpg -k 2 -r 3 -o gfx/ferris-2color.png

The first attempt uses 2 colors, unfortunately this only picks up the color of the highlighter and the average of the ink and paper.

kmeans_colors -i gfx/ferris.jpg -k 3 -r 3 -o gfx/ferris-3color.png

Next, we try with k=3. This is a lot better and shows us that we can separate marker colors from ink too. But it's still not what we want.

kmeans_colors find -c 000000,ffffff -i gfx/ferris.jpg -o gfx/ferris-find.png

The solution to effective black and white separation is using the find subcommand with the -c option, which allows us to specify the colors black (#000000) and white (#ffffff). The k-means algorithm will only need one iteration to find the nearest colors in the image to the colors passed with -c.

b) The --replace flag

With --replace, we run the k-means calculation on an image and replace the centroids with our own custom colors. The colors we input will replace the centroids in order from darkest to lightest, and the number of colors we use will be the amount of k-means centroids we calculate; if we specify 4 colors, we would be replacing the color groups we'd calculate using -k 4 as in Example 1.

We can transfer the average colors of the left image to the lanterns on the right. Running the following command prints the 12 colors below in order from darkest to lightest in hexadecimal.

kmeans_colors -i gfx/flowers.jpg -r 3 -p -k 12 --no-file

492f38,6c363e,8d444e,ae525b,8c6779,677a9b,b87078,4b95bb,a499b0,d7969d,e3b8c0,c5c6da

Then, we can use those colors with the -c option on the right image.

kmeans_colors find -i gfx/lanterns.jpg -c 492f38,6c363e,8d444e,ae525b,8c6779,677a9b,
b87078,4b95bb,a499b0,d7969d,e3b8c0,c5c6da --replace

The top half of the image is the previous command which runs in Lab mode by default. The bottom half of the image was the previous command passed with the addition of the --rgb flag.

Passing the same colors without --replace results in the image below, which colors the pixels with the closest color found in the list we supplied.

kmeans_colors find -i gfx/lanterns.jpg -c 492f38,6c363e,8d444e,ae525b,8c6779,677a9b,
b87078,4b95bb,a499b0,d7969d,e3b8c0,c5c6da

Returning to the Ferris example, we can recolor the image using --replace.

kmeans_colors find -i gfx/ferris-find.png -c de4a18,bee0fa --replace -o gfx/ferris-replace.png

-r and -m can be used with the find --replace subcommand and flag combination. They don't do anything with find by itself, since only one iteration is needed to produce the result.

4) Print, Percentage, & Verbose

kmeans_colors -i gfx/pink.jpg -k 2 -pv --pct --no-file

The results of the iterations can be printed with the -v verbose flag. The score gets smaller as the colors converge. This can be helpful for troubleshooting if results are unexpected since the k-means may not have converged. The -p flag prints the colors in hexadecimal ordered from darkest to brightest as seen below. The --pct flag prints the percentage of each color present in the result image. Finally, --no-file can be passed in order to avoid writing out a result image.

gfx/pink.jpg
Score: 37370.32
Score: 18.397617
Score: 0.2376602
Iterations: 2
4d5f60,dd5d5b
0.6613,0.3387

Usage Notes:

k-means can get stuck in local minima which prevent it from finding the best result. To combat this, the amount of runs can be specified with -r to repeat the process and keep the best result. The -m flag can be used to specify the max amount of iterations to perform. Lastly, the convergence factor can be specified with -f. Larger image files will take longer to complete so defaults were carefully selected for each of these.

A conservative setting of -r 3 -m 40 nearly ensures convergence with Lab for smaller k values (below ~40) in a reasonable amount of time. Finding the --rgb k-means converges very quickly and tends to happen in under 10 iterations for small k.

The --transparent flag can be passed when working with transparent PNG images. The k-means will be calculated without factoring in any pixels with transparency. Otherwise, transparent pixels become matte and negatively impact the results.

Features

• create a color palette from an image
• Lab space or RGB space calculations
• find the nearest colors to input colors
• replace the colors with custom colors
• adjustable iteration count and repetition
• print the average colors
• print the percentage of each color in the image
• transparency support
• supports multiple images as input to process
• specify random seed for reproducible results

License

This crate is licensed under either

• the MIT License, or
• the Apache License (Version 2.0)

at your option.