The Image Processing Corner

Algorithm evaluation criteria

As explained in the previous section, the quality of a segmentation is hard to evaluate. Sensible criteria are rather:

• Execution time.

• Result stability, i.e. two close images give two close results.

• The limited parameter quantity and their sensitivity. Indeed, an algorithm which requires several parameters tuning for each image in not acceptable.

Some region-based segmentation methods

There are many ways to segment an image (not exhaustive):

• The "level set" or the "snakes" methods consists in minimizing an energy function. Many variants of this energy function exist and give different kind of results.

• The histogram-based color segmentation in an adapted color space is a very fast method, often used in video processing.

• The watershed method comes from the morphomathematic theory. It requires a closed gradient image and a manual or not selection of "seeds" (initial points) which will grow until filling the gradient sheds.

• The region split and merge methods. The split method starts from the whole image and splits it if a criterion is not fulfilled. The process is recursively applied on each splitted part of the image until all criterion are fulfilled. The merge method starts from each pixel and merges it with its neighbors if a criterion is not fulfilled. The process is recursively applied on each merged part of the image. Some other methods simultaneously split and merge regions.

• Region growing methods. "Seeds" (initial points) grow while competiting with others. The choice of how to solve conflicts gives many variants.

The next section presents an fast hierarchical region-based segmentation algorithm based on a pure fusion method.