This is an implementation based on the paper
Scale-Sets Image Analysis
by LAURENT GUIGUES, JEAN PIERRE COCQUEREZ and HERVÉ LE MEN,
which talks about the segmentation of an image based on the scale-sets representation.
Usage: scale set [--help] [--version] image budget
Scale set segmentation algorithm
Positional arguments:
image path to input image
budget number of maximum merge allowed by user
Optional arguments:
-h, --help shows help message and exits
-v, --version prints version information and exits
./scale-sets example.jpg 500The scale-sets representation consists of segmenting the image into a set of regions,
which is a combination of smaller regions that depend on budget set by the user.
The number of merges is determined by this budget.
The combination of two regions is determined by comparing their energies.
A region will merge with another region if the resulting energy is lower than the one of any other possible
combination.
Here, the energy of a region is its variance, plus its perimeter multiplied by the lambda factor.
The algorithm starts by creating a region for each pixel in the image.
Then, it iterates over all the regions to merge them with the constraint describe above.
The possible merges are stored in a priority queue, which is sorted by the energy of the resulting region.
The merges are done in order of this priority queue, starting with the merge which has the lowest energy.
Regions are stored in a disjoint set, which allows to keep track of the parent of each region, while searching for them in a reasonable amount of time.
The complexity of the algorithm is complicated to find but annotations can be found in the code.
Here is a working example for our algorithm :
