A new method for image segmentation

A method is presented for finding a threshold surface which involves the ideas used in other methods but attempts to overcome some of their disadvantages. The method uses the gradient map of the image to point at well-defined portions of object boundaries in it. Both the location and gray levels at these boundary points make them good choices for local thresholds. These point values are then interpolated, yielding the threshold surface. A method for fitting a surface to this set of points, which are scattered in a manner unknown in advance, becomes necessary. Several possible approaches are discussed, and the implementation of one of them is described in detail. Two versions of the C.K. Chow and T. Kaneko algorithm (1972) and the present algorithm are applied to a few images, and the latter is shown to give better results, matching human performance quite well