Scala - idiomatic way to calculate sums of interleaved array?

Question

Hi all,

I'm attempting to calculate the average color of an image in Scala, where "average" is defined as the redSum/numpixels, greenSum/numpixels, blueSum/numpixels .

Here is the code I am using to calculate the average color in a rectangular region of an image (the Raster).

// A raster is an abstraction of a piece of an image and the underlying
// pixel data.
// For instance, we can get a raster than is of the upper left twenty
// pixel square of an image
def calculateColorFromRaster(raster:Raster): Color = {
  var redSum = 0
  var greenSum = 0
  var blueSum = 0

  val minX = raster.getMinX()
  val minY = raster.getMinY()

  val height = raster.getHeight()
  val width = raster.getWidth()
  val numPixels = height * width

  val numChannels = raster.getNumBands() 

  val pixelBuffer = new Array[Int](width*height*numChannels)
  val pixels = raster.getPixels(minX,minY,width,height,pixelBuffer)

  // pixelBuffer now filled with r1,g1,b1,r2,g2,b2,...
  // If there's an alpha channel, it will be r1,g1,b1,a1,r2,... but we skip the alpha
  for (i <- 0 until numPixels) {
    val redOffset = numChannels * i
    val red = pixels(redOffset)
    val green = pixels(redOffset+1)
    val blue = pixels(redOffset+2)

    redSum+=red
    greenSum+=green
    blueSum+=blue
  }
  new Color(redSum / numPixels, greenSum / numPixels, blueSum / numPixels)
}

Is there a more idiomatic Scala way of summing up over the different interleaved arrays? Some way to get a projection over the array that iterates over every 4th element? I'm interested in any expertise the Stack Overflow community can provide.

Answer 1

pixels.grouped(3) will return an Iterator[Array[Int]] of 3-element arrays. So

val pixelRGBs = pixels.grouped(3)

val (redSum, greenSum, blueSum) = 
  pixelRGBs.foldLeft((0, 0, 0)) {case ((rSum, gSum, bSum), Array(r, g, b)) => (rSum + r, gSum + g, bSum + b)}

new Color(redSum / numPixels, greenSum / numPixels, blueSum / numPixels)

UPDATE: To deal with both 3 and 4 channels, I would write

pixels.grouped(numChannels).foldLeft((0, 0, 0)) {case ((rSum, gSum, bSum), Array(r, g, b, _*)) => (rSum + r, gSum + g, bSum + b)}

_* here basically means "0 or more elements". See "Matching on Sequences" in http://programming-scala.labs.oreilly.com/ch03.html

Answer 2

This is insane overkill for this problem, but I do a lot of partitioned reductions over datasets, and have built some utility functions for it. The most general of them is reduceBy, which takes a collection (actually a Traversable), a partition function, a mapping function, and a reduction function, and produces a map from partitions to reduced/mapped values.

  def reduceBy[A, B, C](t: Traversable[A], f: A => B, g: A => C, reducer: (C, C) => C): Map[B, C] = {
    def reduceInto(map: Map[B, C], key: B, value: C): Map[B, C] =
      if (map.contains(key)) {
        map + (key -> reducer(map(key), value))
      }
      else {
        map + (key -> value)
      }
    t.foldLeft(Map.empty[B, C])((m, x) => reduceInto(m, f(x), g(x)))
  }

Given that heavy machinery, your problem becomes

val sumByColor:Map[Int, Int] = reduceBy(1 until numPixels, (i => i%numChannels), (i=>pixel(i)), (_+_))
return Color(sumByColor(0)/numPixels, sumByColor(1)/numPixels, sumByColor(2)/numPixels)

Stand mute before the awesome power of higher order programming.

Answer 3

This is a great question, since I think the solution you have provided is the idiomatic solution! The imperative model really fits this problem. I tried to find a simple functional solution that reads well, but I could not do it.

I think the one with pixels.grouped(3) is pretty good, but I am not sure it is better than the one you have.

My own "non imperative" solution involves defining a case class with the + operator/mehtod:

import java.awt.image.Raster
import java.awt.Color

def calculateColorFromRaster(raster:Raster): Color = {
  val minX = raster.getMinX()
  val minY = raster.getMinY()

  val height = raster.getHeight()
  val width = raster.getWidth()
  val numPixels = height * width

  val numChannels = raster.getNumBands()

  val pixelBuffer = new Array[Int](width*height*numChannels)
  val pixels = raster.getPixels(minX,minY,width,height,pixelBuffer)

  // pixelBuffer now filled with r1,g1,b1,r2,g2,b2,...
  // If there's an alpha channel, it will be r1,g1,b1,a1,r2,... but we skip the alpha

  // This case class is only used to sum the pixels, a real waste of CPU!
  case class MyPixelSum(r: Int, g: Int, b: Int){
    def +(sum: MyPixelSum) = MyPixelSum(sum.r +r, sum.g + g, sum.b + b)
  }

  val pixSumSeq= 0 until numPixels map((i: Int) => {
    val redOffset = numChannels * i
    MyPixelSum(pixels(redOffset), pixels(redOffset+1),pixels(redOffset+2))
  })
  val s = pixSumSeq.reduceLeft(_ + _)

  new Color(s.r / numPixels, s.g / numPixels, s.b / numPixels)
}