A Combinatorial Approach forClassification of Patterns withMissingInfornation and Random Orientation
Thomas E. Flickmember, Lee K. Jones · 1986
A maximumlikelihood approach isdeveloped forapat- ternrecognition problem wherethepatterns aredescribed byconfig- urations ofsimple easily recognized parts called primitives. Theap- proachiscapable ofdealing withthree types ofnoise: measurement noise inthelocation andshapeofobserved primitives, undetected or missing primitives (leakage), andtheunexpected appearance ofextra primitives (false alarms). Theapproach iscalled combinatorial be- causethelikelihood function dictates thatobserved primitives mustbe assigned toknownprimitives inallpossible combinations. Duetothecomplexity ofthelikelihood function, practical classifiers mustbebased onlikelihood function approximations. Several arepro- posed, andmostofthese aresimple enough tobeusedinagradient search strategy forrecognizing distorted patterns withrandomorien- tations. Examples areincluded toshowthecharacteristics ofcombinatorial classifier performance. IndexTerms-Combinatorial classifier, distortion, maximumlikeli- hood,missing information, phantoms, statistical pattern recognition. I.INTRODUCTION IN asyntactic approach topattern recognition asin(1), anobserved object maybeidentified fromtherelative locations ofsmall easily detected shapes called primi- tives. Therecognition process involves matching thelist ofprimitives (andrelations between them) withamodel called alanguage comprising allpossible lists andrela- tions whichmaybeproduced byobserving aparticular class ofobjects. A language maybestored asa setof production rules called agrammar. Theremaybemany different grammars stored inalibrary andthetrick of identification istofind theonewhichproduced thegiven observation. Thisisapopular methodofobject identification be- cause itallows verycomplex objects tobedescribed bya fewsimple parts andrules. Forexample, aship observed intheinfrared maybedescribed asaconfiguration ofhot spots, bright patches caused byheatescaping fromthe engine room,stacks, windows, etc.Ships ofdifferent classes maybeseparated bythenumber, location, and Manuscript received December 2,1985.Recommended foracceptance byJ.Kittler. Thisworkwasproduced withthefacilities andsupport ofthe NavalResearch Laboratory.