Abstract

This paper presents an optical approach to the necessary preprocessing step of scene segmentation that conditions raw image information into a form useful for pattern recognition with optical correlators. Potential regions of interest (ROI) are identified and sorted from uninteresting background regions. By eliminating superfluous areas of an image, preprocessing maximizes the speed at which the correlator in the recognition system can operate. System output consists of the edge maps within these ROI. General architectures are based on correlators followed by a threshold. The ROI filter is not matched to any particular object but rather locates deviations from a homogeneous background that have the proper size (determined from available range information). These filters are based on extracting zero- and first-order moments within a window. For objects on a constant background, thresholding at zero intensity yields the object centroids. We present adaptations that allow this approach to be used on more general images where backgrounds can vary. Simulated results using FLIR (forward looking infrared) image data are shown.

© 1991 Optical Society of America