From Image Measurements to Object Hypotheses
Allen R. Hanson, Edward M. Riseman · 1987
A BASIC STEP IN THE CONSTRUCTION OF A SYMBOLIC INTERPRETATION OF AN IMAGE IS THE INITIAL ICONIC TO SYMBOLIC MAPPING, WHICH ASSOCIATES PORTIONS OF THE IMAGE WITH HYPOTHESIZED OBJECT IDENTITIES. OUR APPROACH INVOLVES BUILDING AN INTERMEDIATE SYMBOLIC REPRESENTATION OF THE IMAGE DATA USING KNOWLEDGE-FREE SEGMENTATION PROCESSES. FROM THE INTERMEDIATE LEVEL DATA, A PARTIAL INTERPRETATION IS CONSTRUCTED BY ASSOCIATING AN OBJECT LABEL WITH SELECTED GROUPS OF THE ABSTRACTED IMAGE EVENTS THAT ARE REPRESENTED AS SYM- BOLIC TOKENS. THIS BOTTOM-UP STEP IS NECESSARY IN ORDER TO ACTIVATE POTEN- TIALLY RELEVANT KNOWLEDGE STRUCTURES, AND TO PROVIDE SPATIAL CONSTRAINTS ON THE APPLICATION OF SUCH KNOWLEDGE. THIS PAPER DESCRIBES A SIMPLE MECHANISM FOR GENERATING OBJECT HYPOTHESES THAT RELIES ON COVERGENT EVIDENCE FROM A VARIETY OF ATTRIBUTES OF REGION TOKENS PRODUCED BY A SEGMENTATION PROCESS. WE INTRODUCE THE IDEA OF A CONSTRAINT FUNCTION, WHICH IS AN EXTENDED REAL-VALUED FUNCTION DEFINED OVER A SINGLE REGION ATTRIBUTE. A SIMPLE CONSTRAINT MAPS A SINGLE ATTRIBUTE VALUE OF A REGION TOKEN INTO A GRADED RESPONSE WHICH CAN BE VIEWED AS A `VOTE'' FOR THE ASSOCIATED OBJECT. SIMPLE CONSTRAINT FUNCTIONS ARE HIERAR- CHICALLY ORGANIZED INTO COMPOUND CONSTRAINTS THAT ARE APPLIED TO A SET OF TOKEN ATTRIBUTES IN ORDER TO GENERATE INITIAL OBJECT HYPOTHESES. OUR USE OF THESE CONSTRAINT-BASED TECHNIQUES AS A FOCUS-OF-ATTENTION MECHANISM IS COMP