Calculate bounding box of arbitrary pixel-based drawing

Given a contiguous drawing of arbitrary pixels (e.g. on an HTML5 Canvas) is there any algorithm for finding the axis-aligned bounding box that is more efficient than simply looking at every pixel and recording the min/max x/y values?

Just scanline from top left to right and down to get y top,and similar algorithm with different directions for the rest.

---

Edit by Phrogz:

Here's a pseudo-code implementation. An included optimization ensures that each scan line does not look at pixels covered by an earlier pass:

function boundingBox()
  w = getWidth()            # Assuming graphics address goes from [0,w)
  h = getHeight()           # Assuming graphics address goes from [0,h)
  for y=h-1 to 0 by -1      # Iterate from last row upwards
    for x=w-1 to 0 by -1    # Iterate across the entire row
      if pxAt(x,y) then
        maxY=y
        break               # Break out of both loops

  if maxY===undefined then  # No pixels, no bounding box
    return               

  for x=w-1 to 0 by -1      # Iterate from last column to first
    for y=0 to maxY         # Iterate down the column, up to maxY
      if pxAt(x,y) then
        maxX=x
        break               # Break out of both loops

  for x=0 to maxX           # Iterate from first column to maxX
    for y=0 to maxY         # Iterate down the column, up to maxY
      if pxAt(x,y) then
        minX=x
        break               # Break out of both loops

  for y=0 to maxY           # Iterate down the rows, up to maxY
    for x=0 to maxX         # Iterate across the row, up to maxX
      if pxAt(x,y) then
        minY=y
        break               # Break out of both loops

  return minX, minY, maxX, maxY

The result (in practice) performs about the same as the brute-force algorithm for a single pixel, and significantly better as the object gets larger.

Demo: http://phrogz.net/tmp/canvas_bounding_box2.html

For fun, here's a visual representation of how this algorithm works:

       
       
       
       
       

It doesn't matter in what order you choose to do the sides, you just have to make sure that you take the previous results into account so that you are not double-scanning the corners.