Optimizing boundary maps for image segmentation

We propose an algorithm for image segmentation using the output of any boundary detector. Our algorithm constructs closed regions from the brittle boundaries of boundary maps which are usually obtained when boundary detectors are used for natural images. The proposed algorithm works on a bottom-up process which analyzes neighbourhoods of boundary points to detect discontinuities and connects them. We call this process optimization of boundary maps which is followed by dilation of boundaries and the watershed transform to obtain image segments. The segmentation results can be controlled with user-tuned parameters.