The watershed transform is a key building block for morphological segmentation. In particular, a gray-scale segmentation methodology results from applying the watershed to the morphological gradient of an image to be segmented. The watershed methodology has become highly developed to deal with numerous real-world contingencies, and a number of implementation algorithms have been developed.
A classical approach for producing edge images is to apply a gradient and then threshold the resulting gradient image to produce a binary edge image. A salient difficulty with this approach is the selection of an appropriate threshold value. Even if we ignore the issue of producing false edges by choosing too low a threshold, problems remain. If the threshold is too low, then the edges can be very wide and require extensive thinning, which might leave them inaccurate; if the threshold is too high, then many edges will not be detected and those that are may be severely broken. The conundrum is illustrated by attempting to apply the thresholded-gradient procedure to segmenting the lean meat of an image of a raw beef steak. Figure 7.1 shows (a) the original image, (b) its morphological gradient, (c) the result of a low threshold, and (d) the result of a high threshold. Note how difficult it would be to construct a closed contour by choosing and tracking the thresholded edges.
The literature has devoted many complex methods to deal with this segmentation problem. The watershed approach is a simple yet powerful tool to solve this type of segmentation. Figure 7.2 illustrates the beef segmentation based on the watershed transform (with details to be subsequently explained). For this example, we use the watershed from markers. This watershed requires a set of markers. Each marker must be placed on a sample region of the object to be segmented.
Online access to SPIE eBooks is limited to subscribing institutions.