CODEWITHSUNDEEP
rimage
rosscova
rosscova

Reputation: 5590

Read image data into R piece-by-piece

I'm working with some fairly large image files (aerial survey mosaics, generally > 1 billion pixels), such that loading an entire image into memory would be a problem on my system. I would like to bring them into R piece-by-piece, such that I can process them in "grid-wise" sections.

NOTE: I'm not tied to a particular image format, so tiff, png, bmp etc would all be fine as inputs.

I can do something along these lines with readJPEG, but this requires loading the entire file into memory first, so it doesn't really solve my problem, but hopefully shows what I'm trying to achieve.

image.file <- "~/Desktop/penguins.jpg"

grid.size <- 100
v <- 3
h <- 1

library( jpeg )
image <- readJPEG( image.file )[ seq.int( (v-1)*grid.size+1, (v)*grid.size, 1 ),
                                 seq.int( (h-1)*grid.size+1, h*grid.size, 1 ), ]

The above loads in only a sample of the image, designated by grid.size, v, and h, such that it would be easy to build this into a loop to analyse an image in sections.

Is it possible to achieve this without loading the entire image into memory? Something like read.csv, making use of the skip and n parameters would be reasonable (it would at least only load the vertical sections one at a time, so much less memory needed than readJPEG).

Upvotes: 0

Views: 891

Answers (2)

aoles
aoles

Reputation: 1545

You can easily achieve this entirely in R for almost any image format with the help of RBioFormats, which can be obtained from GitHub.

devtools::install_github("aoles/RBioFormats")

The chunk size can be specified in the subset argument to read.image(). The following example illustrates how to process an image piece-wise without ever loading the whole file into memory.

library(RBioFormats)

filename <- system.file("images", "sample-color.png", package="EBImage")

## first, get image dimensions from metadata
meta <- coreMetadata(read.metadata(filename))

xdim <- meta$sizeX
ydim <- meta$sizeY

## set chunk size
chunksize <- 300  

## itarate over image chunks row-wise
for(i in 1:ceiling(ydim/chunksize)) {
  for(j in 1:ceiling(xdim/chunksize)) {
    x1 <- (j-1) * chunksize + 1
    x2 <- min( j * chunksize, xdim )
    y1 <- (i-1) * chunksize + 1
    y2 <- min( i * chunksize, ydim )

    cat(sprintf("[%d:%d, %d:%d] ", x1, x2, y1, y2))

    img <- read.image(filename, subset = list(X=x1:x2, Y=y1:y2))

    ## perform the actual image processing
    ## here we just print the min and max pixel intensities
    cat(range(img), "\n")        
  }
}

You might also want to check out EBImage, an image processing toolbox for R. It provides functionality to view images and perform various transformations and filtering.

Upvotes: 4

Weihuang Wong
Weihuang Wong

Reputation: 13118

If you have ImageMagick installed, you can crop the image from the command line before reading it into R. An example using this image: http://www.worldatlas.com/worldmaps/worldpoliticallarge.jpg

To create the cropped image:

x <- 800  ## x and y are offsets
y <- 400
w <- 200  ## width and height of cropped image
h <- 100
filename <- "worldpoliticallarge.jpg"
outname <- "crop.jpg"
cmd <- sprintf("jpegtran -crop %dx%d+%d+%d -copy none %s > %s", w, h, x, y, filename, outname)
system(cmd)

Check to see if the new image contains the region we want:

library(jpeg)
original <- readJPEG(filename)
cropped <- readJPEG(outname)
all.equal(original[(y+1):(y+h), (x+1):(x+w), ], cropped)
# [1] TRUE

Upvotes: 3

Related Questions