Abstract

A small bank of optical holographic filters can be designed to recognize a 3-D object from a large solid cone of views. Invariance to image intensity and location is also exhibited. The method of design is first to decompose the set of 3-D views into a set of characteristic image components (eigenimages). These orthogonal image components contain complete information about all the object views to be recognized. An iterative method is applied to combine these eigenimages with a different relative phasing so that the information contained in them can easily be extracted when implemented optically. The iterative method produces a solution for the entire bank of holographic filters. We have found that a bank of about twenty filters is sufficient to recognize a thousand or more object views, while still retaining specificity of recognition. We discuss factors that allow recognition of the discrete training set of images as well as the continuum of intermediate views. The most important factor for continuum recognition is choosing the correct sampling of the different object views.

© 1987 Optical Society of America