recolor.rscript

# Uses k-means clustering to change the colors of blobs inside images.
#
# Take an image, separate the pixels into k cluster buckets, and recolor
# one or more of those buckets.
#
# Mohit Cheppudia 2011 <mohit@muthanna.com>
#
# Prerequisites:
#   - Package 'ReadImages'
#   - Package 'cclust'
#
# Examples:
#
# Take the image 'baby.jpeg', convert all but the most common color (largest
# cluster) to grayscale and save to 'baby_recolor.jpeg'. By default it finds
# 10 clusters.
#
#   > source("~/path/to/this/recolor.rscript")
#   > save_recolor("baby.jpeg", "baby_recolor.jpeg", replacements=1:9)
#
# Find 20 clusters and gray out the three largest clusters.
#
#   > save_recolor("baby.jpeg", "baby_recolor.jpeg", replacements=17:20,
#                  palette_size=20)
#
# Change the color of the largest cluster to green:
#
#   > save_recolor("baby.jpeg", "baby_recolor.jpeg", replacements=10,
#                  rgb_factor=(10/100, 80/100, 10/100))

library(ReadImages)
library(cclust)

# Recolor the imagematrix 'im' using the given parameters:
#
# palette_size: Number of clusters to find.
# replacements: A vector of clusters to replace (smallest to biggest)
# rgb_factors:  Factors for distributing intensity among RGB channels. Defaults
#               to greyscale (i.e., even intensities acrossl all channels.)
# iterations:   Number of iterations for k-means clustering.
#
# Returns a new imagematrix with colors replaced.
recolor <- function(im, palette_size=10, replacements=1:palette_size,
                    rgb_factors=c(1/3,1/3,1/3), iterations=20) {
  # 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 -- sort clusters by size, and replace
  # the requested clusters in 'replacements' with new colors.
  top <- 1
  for (i in order(k$size)) {
    # Replace requested clusters with greyscale counterparts
    if (any(top == replacements)) {
      cluster <- observations[k$cluster == i,]

      # Calculate the total intensity of the pixel and redistribute
      # to the RGB channels based on their respective requested factors.
      intensities <- rowSums(cluster * cluster)
      intensities_r <- sqrt(intensities * rgb_factors[1])
      intensities_g <- sqrt(intensities * rgb_factors[2])
      intensities_b <- sqrt(intensities * rgb_factors[3])

      # Update observations with new colors
      observations[k$cluster == i,] <- cbind(intensities_r,
                                             intensities_g,
                                             intensities_b)
    }

    top <- top + 1
  }

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

# Take a source image from 'image_path', recolor it with the specified
# parameters, and save to 'save_path'.
save_recolor <- function(image_path, save_path, palette_size=10,
                         replacements=(1:palette_size - 1),
                         rgb_factors=(c(1/3,1/3,1/3)),
                         iterations=20) {
  im <- read.jpeg(image_path)
  fm <- recolor(im, palette_size, replacements, rgb_factors, iterations)

  dims = dim(im)
  jpeg(save_path, width = dims[2], height=dims[1], quality=90)

  # Options to remove margin (prevents rescaling of the image)
  par(xpd=NA, mai=c(0,0,0,0))
  plot(fm)
  dev.off()
}