(This post uses new features from the R Cultural Analytics Library version 1.0.6 .)
We can divide an image into sections, analyse the R, G and B values of each of those sections and plot the results.
library(CulturalAnalytics)
library(jpeg)
library(vegan)
## http://blog.wolfram.com/2008/12/01/the-incredible-convenience-of-mathematica-image-processing/
## Load the image
imgdir<-paste(system.file(package = "CulturalAnalytics"), "images", sep = "/")
dirimgs<-paste(imgdir, dir(path = imgdir, pattern = ".jpg"), sep = "/")
img<-readJPEG(dirimgs[1])
## Divide it into sections and get a table of the median RGB values
sections<-divideImage(img, 8, 8)
## Get the median rgb values for each image
rgbs<-sapply(sections, function(img){ coords(medianRgb(imageToRgb(img))) },
USE.NAMES=FALSE)
## The list needs transposing so we have columns of r,g,b values
rgbs<-t(rgbs)
## Give the columns useful names
colnames(rgbs)<-c("r", "g", "b")
## Bubble Chart
plot(rgbs[,"r"], rgbs[,"g"], type="n", xlim=c(0,1), ylim=c(0,1),
main="Section Bubble Chart of \"Bonjour, Monsieur Corbet\"",
sub="(size is blue)", xlab="Red", ylab="Green")
images(rgbs[,"r"], rgbs[,"g"], sections, cex=rgbs[,"b"])
This shows no areas of pure, saturated colour.
Next we can cluster the sections of the image and show the resulting clusters.
<tt>
## Cluster the tiles by colour
## http://stackoverflow.com/questions/9019632/how-to-create-a-cluster-plot-in-r
## 5 is arbitrary
clusters<-kmeans(rgbs, 5)
## distance matrix
rgbs.dists<-dist(rgbs)
## Multidimensional scaling
cms<-cmdscale(rgbs.dists)
## Plot the clusters
plot(cms, type="n", xlim=range(cms[,1]), ylim=range(cms[,2]),
main="Section Clustering of \"Bonjour, Monsieur Corbet\"")
groups<-levels(factor(clusters$cluster))
images(cms[,1], cms[,2], sections, thumbnailWidth=20)
ordispider(cms, factor(clusters$cluster), label=TRUE)
ordihull(cms, factor(clusters$cluster), lty="dotted")<br></br> </tt>
If the plots didn’t have helpful titles, would you be able to recognize the image?
Despite the arbitrary number of clusters chosen the groupings make some visual sense. Improving on the number of clusters is left as an exercise for the reader.