A hierarchical, automated target recognition algorithm for a parallel analog processor

A IIicrarcllical al)proacll is dcscrit.d for all autolt~atd target rccogtlitioll (A’J’1{) systcItI, VIG1l,AN’1’11}, that uses a Inassivc]y l)arallel, analog ]jrocessor (31) ANN). ‘J’lic 31) ANTNT l)rocessor is caI)ablc of Ixmforlning 64 com curwrlt, illncr I)roclucts of size IX409G every 250 llatloseco]lds. A colltIjlctc 64x64 raster scan of a 256x256 ittlage call be evaluated by tflc 31)ANNT wit}l its 64 ltlodifial~lc tetllplatcs in al)out 16 Iuilliscconcls. ‘1’0 fully utili7,c tl[c a[lalog I)rocessor a]id accon]~nodate its IIigll l)altdwidtlt rcquireltlcnts, tllc v e c t o r s (tern I)latcs) loadd oil to the 31)ANAT ]jrovidc climcllsionality rcduc(, ion for a hacli-cnd set of classifiers. ‘1’hc tc]ur)latcs used in the reported algoritllltl arc llicrarcllically gelwratcd sets of cigcllvcctors take]) froln a })artitiolid set of library object images. As itlforltlatiou is accuInulatd al)out tltc target (e.g. oljjcct class), a rnorc rdind set of ei.gmlvectors reflecting this kIIowldge call bc loadccl and IIlorc sl)ccializd classifiers utilized. ‘J’hc classifiers ~Jrovidc inforlnation related to tllc Al’l{ task: location, class, sul)-class, aI]d oricntatioIl of target(s). \\7c rel)ort some prcli Illinary results that cxaltlinc tllc ~mrfor]tlancc of orientation classifiers. \Vitfl no ktlowldgc aboltt object class 01 oriclltatio]l, a ]Ieut-al ]Ictlvork aclticvm 94 .2(X in detertniuirlg to wllicll OIIC of tflrcc classes froln wrt ical (30°, 45°, or 60° ) an ol~jcct ifllagc is cJriellted (+300). Using al I eigcllvector tcttll)latc set .gcv]eratd fro]]) a clistributio~l }vllerc bc~tll object class and oric]ltatic]ll are kHown, a neural llctwork classifier acllicvcs 96% in orict]ting the object to within +22,.5°. ‘1’ltis illfor~tlation catl bc used to load cvel) IItorc slwcific cigcllvcctor sets Jvllicll sl]oulcl lead to I[lorc accurate ot)ject locatiotl duri[lg tracking and an cnllatlccrnetlt itl ol)jcct rwogt]itiotl tasks.