Edgeless active contouring, for vector-valued natural image segmentation

We propose here an efficient geometric active contouring method based on the level set approach for extracting objects from natural images described in vector-valued form. Natural images are characterized by absence of global minima for mean squared error, an energy minimization formulation based on the the principles of the calculus of variations, that helps in effective segmentation based on boundary information. The approach adopted is to treat this segmentation as a minimum partition approximation problem, using additional regularization terms. The constraints for stopping the evolving curve are derived by coupling information from each of the vectors of the vector described image. The coupling effect from each vector increases the segmentation accuracy. The results are qualitatively compared with an existing Chan et al. (1999) model and are found to be much superior.