Kasvand(ID:7345/)Organically derived picture description language References: in Watanabe, S. (ed.) "Frontiers of Pattern Recognition" Academic Press 1972 view details in Watanabe, S. (ed.) "Frontiers of Pattern Recognition" Academic Press 1972 view details Yet, in the literature on psychology of vision, on the effects of brain damage, on how to draw pictures, etc., there is a wealth of hints on how biological systems appear to s the scene analysis problem. To biological systems scene analysis "comes naturally", one is not even aware that the problem exists. Some observational results will be compared with computational procedures. Extract: A simplified picture language Appendix D: A simplified picture language A rather simple picture language for two dimensional, monochromatic objects was formulated, programmed and tested experimentally. It illustrates how the first 7 variability-dimensions (Table 2) were handled. The language was based on the follow premises: a) Select local features (point features) which require no knowledge about picture content, in terms of which the objects can be described and for which a normalization procedure exists. These are, of course, spatial gray level gradients, contours, curvatures of contours, etc. b) Fragment or segment the unknown objects (if complicated) such that reasonably coherent descriptions of the segments (atoms) can be obtained prior to any knowledge about the objects. The algorithm used the point features to pinpoint the atoms. (Translation problem eliminated.) c) Form an atom description which can be normalized before knowing what it represents. A polar coordinate representation of the atom allowed normalization for size and rotation Since the as yet unknown atom is now described, its gray level is known. d) Recognition of an atom is a comparison operation where the description of the unknown atom is compared with a list of known atoms and the best match is selected. The numerical value of this match is a measure of distortion. e) A (complicated) object is a spatially organized collection of atoms. It may suffice to identify one atom only, if no ambiguity results (partial view). The atoms were selected only by the computer (algorithm). The online operator informed the machine which atoms belonged together to form a complex object. The spatial interrelationships were computed automatically since the necessary data was available. Besides numerous shortcomings the language behaved fairly well. In retrospect, however, the greatest conceptual difficulty occurred in assigning atom priorities or in trying to define which atoms were important in a given complex object. The atom areas (segments) into which a complex object was fragmented were essentially mutually exclusive. The program had no ability to disregard (i.e., not "see") elements in the picture during the segmentation procedure. The problem is best illustrated with a very simple example: If the object is a circle with a bar in it (like the letter 0) one could describe it as upper semicircle * lower semicircle * bar, or circle = upper semicircle* lower semicircle, object = circle * bar. An unrelated line through this object (0) ruins the description since the semicircles are split into sectors. Basic difficulties arose due to irrelevant details in a complex object. in Klinger, A.; Fu, K. S.; Kunii, T. L. "Data Structures, Computer Graphics, and Pattern Recognition" (Largely based on IEEE Computer Society conference held in Los Angeles, May 1975) Academic Press, NY 1977 view details |