palette.rscript

# Uses k-means clustering to generate a palette of colors
# from an image.
#
# Mohit Cheppudia 2011 <mohit@muthanna.com>
#
# Prerequisites:
#   - Package 'ReadImages'
#   - Package 'cclust'
#
# Quick run:
#   source("~/path/to/this/palette.rscript")
#
#   plot_palette("/path/to/image") # Find 10 clusters (default)
#   plot_palette("/path/to/image", 20) # Find 20 clusters
#
# Alternate run:
#   image <- read.jpeg("/path/to/image")
#   p = palette(image, 10, 20)
#   plot(p)

library(ReadImages)
library(cclust)

palette <- function(im, palette_size, iterations) {
  # Convert the image into a list of observations, each observation
  # is a 3-dimensional vector of RGB. For cclust we represent this
  # as 2x3 matrix where each row is an observation and each column
  # is a feature (R,G,B).
  observations <- matrix(im, ncol=3)

  # Run k-means clustering, ask for 'palette_size' clusters, run
  # at most 100 iterations until convergence. This call requires the
  # 'cclust' package to be installed.
  k <- cclust(observations, palette_size, iter.max=iterations)

  # We now have the clusters -- merge the clusters back into a matrix
  # to display the final image. k$cluster is a list of cluster numbers,
  # one for each observation.
  #
  # i.e., k$cluster[1] holds the cluster number for observation 1.
  final_matrix <- NULL
  for (i in order(k$size)) {
    cluster <- observations[k$cluster == i,]
    final_matrix <- rbind(final_matrix, cluster)
  }

  # Create the final image (with the dimensions of the original image
  # and return.
  dims = dim(im)
  final_im = imagematrix(final_matrix, "rgb", dims[1], dims[2])
  final_im
}

# Display image and palette next to each other
plot_palette <- function(image, palette_size=10, iterations=20) {
  im <- read.jpeg(image)
  fm <- palette(im, palette_size, iterations)

  layout(matrix(c(1, 2, 1, 2), 2, byrow=T))
  plot(im)
  plot(fm)

  title(sprintf("Clusters: %d", palette_size))
}