Efficient Boundary Tracking Through Sampling

The proposed algorithm for image segmentation is inspired by an algorithm for autonomous environmental boundary tracking. The algorithm relies on a tracker that traverses a boundary between regions in a sinusoidal-like path. Boundary tracking is done by efficiently sampling points, resulting in a significant savings in computation time over many other segmentation methods. For noisy images, the traversed path is modeled as a change-point detection problem between two states. Change-point detection algorithms such as Page’s cumulative sum procedure are adapted in conjunction with other methods to handle a high level of noise. A modification for the multiple-region case is also presented as a hybrid of a topology-detecting segmentation algorithm and boundary tracking. Applications to high resolution images and large data sets such as hyperspectral are of particular interest. Irregularly shaped boundaries such as fractals are also treated at very fine detail along with accompanying fractal dimension calculations, which follow naturally from the boundary tracking data.