Sometimes image regions are incompletely specified in
the sense that only a few representative points in each
region are known. In order to determine a segmentation of
such an image it is necessary to construct the region
boundaries. A fast algorithm called a diffusion algorithm
is given that determines region boundaries by diffusing
region labels outward from the known points.
In this paper the problem of texture boundary formation is
approached by a region growing technique that is based upon a structural
model of texture.
Region growing is based upon similarities among
texture elements and upon the spatial proximities of these elements. The region grower itself is a global algorithm that makes use of a minimal spanning tree of what is called the associated texture graph of
a texture. A primary advantage of the method is that the texture
regions are self-organizing in the sense that no artificial windows need
be superimposed over the source textures. While the region described in
this paper uses extrema-based texture elements, the application to other
types of elements is straightforward.
A description of a coherent optical processor which utilizes matched spatial filters in order to do pattern recognition is presented. The processor has been interfaced to a PDP-11-40 computer which controls the input film drive, the filter stage stepper motors, digitizes, and stores data and is used for data analysis. On-line to the computer are various peripherals including: a DEC-writer, a storage scope, a plotter, and a display terminal. As an example of a pattern recognition problem, we discuss the application of the system to the identification of biological specimens.