Depth Recogition of Stereo Images Using Hierarchiral Depth Perception System.

A neural network called a “Hierarchical Depth Perception System” is proposed in this paper for extracting the depth information from a stereo pair of images. Since isparity, i. e. difference between left image and right one, is assumed to be continuous everywhere in conventional methods which use competition and cooperation between neural units, they do not work well for images containing discontinuous disparities. We use a hierarchical system with the following features ; cooperation by the use of Perceptron, removal of isolated active nuits, cooperation and competition between units. In this system, some functions are activated at the positions with continuous disparities, and other ones are activated in other cases. The structure and functions of the system are described, and an application of the system to the depth perception of random dot stereograms is shown.