Date of Award
Doctor of Philosophy (PhD)
Department of Electrical and Computer Engineering
Steven K. Rogers, PhD
This dissertation details the development of a new computational vision model motivated by physiological and behavioral aspects of the human visual system. Using this model, intensity features within an artificial visual field of view are extracted and transformed into a simulated cortical representation, and a saccadic guidance system scans this field of view over an object within an image to memorize that object. The object representation is thus stored as a sequence of feature matrices describing sub-regions of the object. A new image can then be searched for the object (possibly scaled and rotated), where evidence of its presence is accumulated by finding sub-regions in the new image similar to those stored and with the same relative spatial configuration. A set of over 450 experimental trials demonstrates the model is capable of memorizing and then recognizing arbitrary objects within arbitrary images, as well as correctly rejecting images that do not contain the memorized object. A new context based recognition paradigm is introduced that solves the problem of a priori assignation of recognition thresholds, and also can be generalized to solve thresholding problems commonly found in pattern recognition environments. A demonstration is provided of the model's applicability to real world problems by memorizing a face and text string, and then successfully searching a video sequence for their presence.
DTIC Accession Number
Keller, John G., "SCAN-IT: A Computer Vision Model Motivated by Human Physiology and Behavior" (1999). Theses and Dissertations. 5120.