Application of Hopfield Network to Human-Guided Contour Extraction.

T o develop a contour extraction tool for image simulations, the applicability of the Hopfield network is examined on the edge image around the roughly specified guide points. Our computational theory is that the edge map of the stretched belt-like images along the guide points should obey the following four constraints. (1) In thc longitudinal dircction, t l~c contour should consist of only one pixel. (2) Contour points are usually locatcd close to those in the neighboring columns. (3) Contour points arc usually located on the detected edge pixels in thc edge map. (4) Contour points are usually located near the horizontal center of the edge map. Furthermore, to obtain a size independent energy function, we developed a scaling relationship. Using the energy function developed according to these observations, the experimental results are shown in which contour extraction is succesful for the most part.