Watershed Segmentation

Watershed segmentation [7] finds pixels at the midpoints between features by treating the pixel values as topographic elevations. The areas between features become catchment basins which depict the influence zones in the image (see figure below). Images should be low-pass filtered first to avoid over-segmentation.

The output is a new image in which each catchment basin is given a different gray scale value. The actual watershed points are those points on the new image with a value of 0. These points can be selected by inverting and increasing the contrast, or by applying the formula

$$if(i>0) i=255;$$

to the image. These edges form a tessellation of the image while the constant color regions indicate the catchment basins.

Watershed segmentation of an emulsion autoradiogram.
A Original image
B Watershed segmentation superimposed on the image. The image was low-pass filtered twice using a 5$\times$5 kernel before segmentation. The segmentation edges (i=0 points) were pasted over the original using the ``Mask..Add'' function.
C Watershed segmentation of grayscale-inverted image, created as in A except the image colors were inverted before segmentation. The watersheds now track the closest distance between the grains instead of the greatest distance between them.
D Actual watershed segmentation obtained in B, with a spectrum colormap added.
E Quick segmentation result superimposed on the same image in A for comparison. Image was filtered once with a 15$\times$15 low frequency removal filter, using a kernel multiplier of 2 before quick segmentation. Line around outside is an edge artifact caused by the filtering.

If the gray scale of the image is inverted first, the result is that lines are drawn connecting the highest elevation routes through the features instead of around them (see figure).

The segmentation pattern can be superimposed on the original image by the following procedure:

1. Increase the contrast of the watershed image by 1000x so that it appears as a network of black lines on a white background.
2. Select "Process...Mask" and set "Image to change" to the image number of the original image and "Image for mask" to the image number of the watershed image.
3. Select "Mask (1&=2" and click "Accept". The black watershed lines should now be superimposed on the original image. If the original image is color, the watershed lines can be set to any color and added to the image.