Cluster images by similarity, and plot

cluster-images.R

# read in the images...
library(magick)
img_paths <- list.files("../outline_P/", full.names = TRUE )

n <- 100 # just get a sample
imgs <- image_read(img_paths[1:n])

# reduce size
imgs_cropped <- image_crop(imgs, "100x100")

# rescale, all the same size
imgs_scaled <- image_scale(imgs_cropped, "5x5")

## convert to EBImage format, grayscale, matrix...
# one image per row

library(EBImage)

x2 <- vector("list", length = length(imgs_scaled))
for(i in seq_along(imgs_scaled)){
  x1 <- as_EBImage(image_convert(imgs_scaled[i], colorspace = 'gray'))
  x2[[i]] <-  as.vector(imageData(x1))
}

names(x2) <- img_paths[1:n]

# list of vectors to data frame, one row per image
library(dplyr)
x3 <- bind_rows(x2)
x4 <- t(x3)

# how many clusters?
library(optCluster)
opclust_out <- optCluster(x4, nClust = 2:10, 
                          clMethods = c("kmeans"), 
                          validation = "stability")

# do kmeans
library(fpc)
k_out <- kmeans(x4, max(optAssign(opclust_out)$cluster))
x5 <- discrcoord(x4, k_out$cluster)
plot(x5$proj[ , 1:2], col = k_out$cluster)

## plot
library("ggplot2")
library("ggimage")

# prepare data frame for plotting
d <- data.frame(x = x5$proj[ , 1],
                y = x5$proj[ , 2],
                image = img_paths[1:n],
                clus = k_out$cluster)

# draw the plot
ggplot(d, 
       aes(x, y)) + 
  geom_image(aes(image = image), 
             size=0.075) +
  geom_point(aes(colour = factor(clus))) +
  theme_minimal()