Willamette Valley Images

Here are some example segmentations of the Willamette Valley image using various parameter settings to control the segmenter. The process I used is to create an initial watershed segmentation and then apply the morphological reconstruction algorithm to that segmentation. The reconstruction algorithm basically just allows you to set a parameter to erase all boundaries that have less than a specified amount of contrast.

The variables n and h mentioned in the image labels below are the parameters that control the process:

• n
  The number of levels in the multiscale edge detector.
  
  This controls how large a window is used in looking for edge pixels and largely dictates the character of the initial watershed segmentation. The smaller n is, the finer the edge that will be recognized.
  • n=1 means look in a 1 pixel radius of the pixel being considered, i.e., use a 3x3 window.
  • n=2 means look in a 2 pixel radius, i.e., use a 5x5 window
  • n=3 is the largest value I've tried. I generally use n=2.
  
  
• h
  The depth of local minima in the edge image that will be erased.
  This controls the amount of contrast between adjacent regions that will be recognized as a boundary.
  
  It affects how large the resulting regions tend to be. Large values of h mean that a boundary must have very high contrast to survive. This in turn means that regions tend to get very large since most boundaries are erased.

Each image that you click on below will open a new browser window on a much larger version of that image so that you can have several of them open at once and resize the windows yourself to make comparing the segmentations easier. Note one thing: when the new browser comes up in Netscape, it's directly over the old browser so it looks like it came up in the original window, but that's not the case. You can just grab the top bar of the browser frame and drag the new browser off to the side. The same thing happens for each new image that comes up.

In each of the images, the program's segmentation is laid over the original image using blue lines. The segmentation that Pat did using small polygons appears as red lines and boundary points where the two match are in green.

Initial Watershed Segmentation

segmentation for n=2, h = 0

Reconstruction Segmentation for n=2, h=4

segmentation for n=2, h = 4

Reconstruction Segmentation for n=2, h=10

segmentation for n=2, h = 10

Reconstruction Segmentation for n=2, h=15

segmentation for n=2, h = 15

Reconstruction Segmentation for n=2, h=20

segmentation for n=2, h = 20

Reconstruction Segmentation for n=2, h=50

segmentation for n=2, h = 50

Other Examples from Puget Sound

To give you an idea of how the algorithm behaves on other types of images, I have included a link here to a page that I put together for someone else last year with some segmentations from another data set covering parts of Puget Sound. The original images were thermal infrared aerial photos, but they were converted to grey scale to do the segmentations. There is no reference segmentation for these, so the program's boundary outputs are shown in red and there are no blue or green lines. To see these examples, click on the image below.

Puget Sound color infrared segmentations