Robust and accurate image segmentation using deformable templates in scale space

This paper presents a new image segmentation algorithm using deformable templates in scale space. The deformable templates are greylevel patterns with clearly defined image features to represent ideal segmentation results of some generic percepts. To segment a specific target in an image, the algorithm deforms the corresponding generic template to match the actual state of the target. To reduce the probability of being stuck at local minima and to speed up the process of convergence, the algorithm deforms the templates in scale space from coarse to fine and uses the normalized cross-correlation to provide initial states for the deformation process. To achieve the best accuracy for localizing object boundaries, the algorithm also employs the 2D optimal edge detection functional developed by Qian and Huang at the finest scale. Experimental results on real images are given in the paper.