Table Of Content

AFOSR Research Programs in Image Fusion yengn a.TJ secnei ceSf idL nyartsime h fCeotarotceriD hcraes ecRifitnei c fSeocif feOcr orFiA .A.S .,U A,VnotgnilrA [email protected] Abstract - The U.S. Air Force Office of Scientific Sensors Directorate (SN) and the Information Directorate Research supports basic research in image fusion within )FI( lliw eb .detartsulli roF ,elpmaxe eseht owt several programs. These programs are presented in setarotcerid fo LRFA krow rehtegot no a noisuf“ context of Air Force technology needs for targeting, network” for technology demonstrations that integrate image exploitation, and autonomous systems. Programs srosnes dna .gnissecorp RSOFA setubirtnoc ot NS include research involving human perception and neural ygolonhcet secnavda ni levon ,srosnes gniledom foeht processing in other biological systems, algorithms for environment, data storage and processing. AFOSR also fusion from multiple sources and platforms, and novel setubirtnoc ot FI ygolonhcet secnavda ni atad dna sensors. Available mechanisms for support of information fusion and image exploitation. collaborative research will also be presented. hguorhT eht LRFA ygolonhcet ,setarotceridRSOFA dednuf ecneics si dednuorg ni snoitacilppa fo tseretniot Keywords: Sensors, Displays, Shape Encoding, the Air Force. In addition to SN and IF, other Algorithms, Learning, Control. setarotcerid fo eht LRFA etubirtnoc ot noisuf .seigolonhcet emoS noitamrofni no eht krow fo hcaefo eht ygolonhcet setarotcerid si elbaliava ta 1 Introduction http://www.afrl.af.mil nekaT sa a ,elohw ,revewoh eht LRFAygolonhcet ehT riA ecroF eciffO fo cifitneic ShcraeseR )RSOFA(si troffe detaler ot noisuf sedulcni lla eht stnemelederiuqer : the single manager of basic research for the U.S. Air gnizihpromoporhtna ,tahwemos evitceffe seye dnasdnah ecroF .)FASU( cisaB hcraeser si denifed ni eht.S.U connected with a brain. The sensing goal is development esnefeD tnemtrapeD sa eht tsom latnemadnuf fokrow fo a nommoc gnitarepo erutcip tsubor htiw tcepserot along a research and development spectrum from rosnes ,seigolonhcet dna eht gnitceffe laog si rofdeeps egdelwonk yrevocsid ot ygolonhcet .noitartsnomed ehT dna ycarucca fo taerht.esnopser RSOFA stroppus ynam saera fo ,ecneics tub sesucofsti troppus no scipot detcepxe ylgnorts ot etubirtnoc oteht ygolonhcet sdeen fo eht .FASU esehT smargorpera 1.2 AFRL Partnering in Fusion Technologies executed largely through grants and contracts to industrial and academic researchers. Description of ehT LRFA osla skrow ylesolc htiw rehto yratilim tnerruc scipot fo RSOFA troppus si elbaliava ta eht departments, with industry and other government .)lim.fa.rsof aR.Sw OweFwmA/ o/eh:gpatpth( agencies. The AFRL maintains a virtual distributed ,noisuF eht cipot fo siht ,gniteem si ton naRSOFA ,yrotarobal fo hcihw eht reilrae denoitnem noisufkrowten sucof aera esuaceb RSOFA senifed scipot nicifitneics si a ,trap ot etanidrooc ygolonhcet stnempolevedssorca ,smret ton lacigolonhcet .seno ,sselehtreveN lareves the institutions involved. RSOFA smargorp etubirtnoc ot noisuf .seigolonhcet I ehT lluf egnar fo noisuf seigolonhcet si fo tseretniot will describe program contributions each in turn, the Air Force, and the AFRL is working in a leadership emphasizing those that take advantage of biological role on many, if not most, of them. This role requires langis ,gnissecorp fo laiceps tseretni ot.em gnitanidrooc stroffe fo ynam hcraeser dnatnempoleved .snoitazinagro ehT ecneics detroppus yb RSOFA si conducted in many of these organizations and, when not, 1.1 AFRL Fusion Science and Technology is connected tightly to them through the AFRL. RSOFA si a trap fo eht riA ecroF hcraeseRyrotarobaL )LRFA( dna sesu sti cifitneics smargorp ot troppus 2 Fusion Science ygolonhcet stnempoleved nihtiw s’LRFA lareves ygolonhcet .setarotcerid esehT ygolonhcet setarotcerid The AFOSR supports research in areas related to fusion tcudnoc emos cisab ,hcraeser tub evah yramirp of information and data, including multi-spectral image seitilibisnopser rof ygolonhcet tnempoleved dna sensing, fusion and processing for recognition and .noitacilppa emoS fo eht detaler-noisuf krow ni eht identification of targets, as well as integrated distributed Report Documentation Page Form Approved OMB No. 0704-0188 Public reporting burden for the collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering and maintaining the data needed, and completing and reviewing the collection of information. Send comments regarding this burden estimate or any other aspect of this collection of information, including suggestions for reducing this burden, to Washington Headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite 1204, Arlington VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to a penalty for failing to comply with a collection of information if it does not display a currently valid OMB control number. 1. REPORT DATE 2. REPORT TYPE 3. DATES COVERED 2000 N/A - 4. TITLE AND SUBTITLE 5a. CONTRACT NUMBER AFOSR Research Programs in Image Fusion 5b. GRANT NUMBER 5c. PROGRAM ELEMENT NUMBER 6. AUTHOR(S) 5d. PROJECT NUMBER 5e. TASK NUMBER 5f. WORK UNIT NUMBER 7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING ORGANIZATION Directorate of Chemistry and Life Sciences Air Force Office of Scientific REPORT NUMBER Research Arlington, VA, U.S.A. 9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR’S ACRONYM(S) 11. SPONSOR/MONITOR’S REPORT NUMBER(S) 12. DISTRIBUTION/AVAILABILITY STATEMENT Approved for public release, distribution unlimited 13. SUPPLEMENTARY NOTES The original document contains color images. 14. ABSTRACT 15. SUBJECT TERMS 16. SECURITY CLASSIFICATION OF: 17. LIMITATION OF 18. NUMBER 19a. NAME OF ABSTRACT OF PAGES RESPONSIBLE PERSON a. REPORT b. ABSTRACT c. THIS PAGE UU 7 unclassified unclassified unclassified Standard Form 298 (Rev. 8-98) Prescribed by ANSI Std Z39-18 retemarap lortnoc fo eht selcihev devlovni ni eseht.sksat yam eb lufesu ni gnitarapes edam-nam stcejbo morf Main topics of research are considered in turn. natural backgrounds -- camouflage breaking. In the approach described above, the infrared sensing 2.1 Sensors material is obtained from novel 3-5 materials, spectral tuning is obtained using interference filters constructed The AFOSR supports research on sensing materials and of Germanium and Silicon Oxide, and polarimetric devices for multi-spectral imaging. Some research sensitivity is obtained using a standard grating technique. supported in Biological programs takes inspiration from RSOFA osla stroppus hcraeser ot enibmocdnab-itlum yduts fo larutan .smetsys ehT ,nohtyp rof ,elpmaxenac optical and radar sensing elements into compact, low fuse the visible and infrared spectra to detect and capture ,rewop wol thgiew .seludom enO tnerruc hcaorppa its prey. Research supported in this program has recently sevlovni hcraeser ta eht latnemadnuf ecived ngisedlevel determined [1] that the heat sensitive organ of the python ot poleved nommoc erutrepa secived rof evitca dna yam ekat egatnavda fo dezilaiceps enarbmemserutcurts evissap gnisnes fo OE dna .FR gnidroccA ot enonoiton that provide a degree of spectral filtering for underlying rof evitca ,sedom rof ,elpmaxe RADAL noitanimulli heat transduction mechanisms. The microstructure of the dluow eb detaludom ta evaw-mm seicneuqerf os taht overlaid membrane, specifically, the spatial distribution return light might be detected simultaneously and fo norcimbus stip nihtiw eht ,enarbmem si detcepsusot demodulated with the radar returns for precise pixel play a role in spectral filtering and is the object of current noitartsiger fo eht owt.segami .yduts ehT llarevo laog fo siht hcraeser si otenimreted It has been long known that interaction can occur whether infrared or heat sensitive detectors might be neewteb citengamortcele sdleif dna suoirav lairetam operated at ambient temperatures with higher sensitivity excitations. Brillouin scattering results from the and resolution than those available using current gnilpuoc fo citengamortcele sdleif dna citsuoca sevawni .seigolonhcet matter. For the first time, as far as we know, researchers gnikrow htiw eht riA ecroF eciffO fo cifitneicShcraeseR are exploiting this electromagnetic-acoustic coupling to attempt to extend Air Force imaging capabilities. In a real material medium such as f ,egailo,doow soil, concrete, or plastic, the material polarizability and material conductivity are dependent on the local pressure. lairetaM ytilibaziralop srefer ot eht eerged ot hcihwdexif charge in the material can be distorted by an electric field. A measure of material polarizability is the cirtceleid .tnatsnoc lairetaM ytivitcudnoc srefer oteht ycnednet fo eerf segrahc nihtiw eht lairetam ot evomni Figure 1 : srelttaR dna snohtyP esneS elbisiVdna eht ecneserp fo na cirtcele .dleif elihW ytilibaziralopdna Infrared ytivitcudnoc era snoitcnuf fo lacol ,erusserp eseht lairetam seitreporp osla ecneulfni citengamortceleevaw In the optoelectronic materials program, AFOSR speed in the medium. Thus, there is a connection supports research on sensor materials for focal plane neewteb erusserp dna eht tnemevom focitengamortcele syarra dna rehto ,secived ot revoc eht eritnemurtceps energy, and, it is this interaction that is being exploited with fast, sensitive, high resolution images. Current to enhance imaging. approaches include quantum well infrared photodetection rosseforP samohT sknaB ]2[ dna seugaelloc taeht and quantum dot arrays. These approaches provide fast Center for Scientific Computation at NC State ,)sdnocesocip( evitisnes snet( fo ,)nivleK-illim high ,ytisrevinU rednu RSOFA ,pihsrosnops evahdemrofrep sryyo ta.srirsnosrnefn aoseyi dttaea icfrioalvppa computational research that indicates that this ehT RSOFA scinortceleotpo slairetam margorp osla citsuoca-citengamortcele )erusserp( noitcaretni nac eb supports research on integrated focal plan imaging of exploited to help identify materials. Banks and multiple spectral bands. The interspersed spectral seugaelloc evah dekrow ta evaworcim seicneuqerfdna images that result offer advantages in image registration evah nwohs taht na citsuoca evaw tnorf nac tca saa and sensor fusion because images in multiple bands are ylkaew evitcelfer gnivom rorrim nihtiw a muidemhcus obtained at the same time through the same optics. sa .lios gnisU trohs evaworcim seslup dna gnitiolpxeeht One approach to integrated multi-spectral sensing tsav ecnereffid neewteb citengamortcele sdeeps ni sevlovni noitcurtsnoc fo eerht yb eerht lexip syarrabusni slairetam dna citsuoca ,sdeeps citsuoca evaw stnorfnac hcihw hcae fo eht enin gnisnes slexip-bus seviecerna be interrogated as they spread in a structure. image filtered into one of three infrared bands and one of Interrogating the acoustic or pressure fronts appears to be three polarimetric orientations. Polarimetric images are able to aid in the characterization of the material between of interest because man-made objects, with their planar eht ecafrus fo eht tcejbo dna eht noisilloc“ ”tniop foeht ,secafrus yllacipyt od ton tcelfer thgil fo lla.snoitaziralop . eesvl au t wpcend oivonertarahcswftiumoca desuF segami nac osla eb detcurtsnoc ot edivorpmetsys This research is aiming at exciting military operators and image analysts with new information that .snoitacilppa ,yllacificepS ti si detcepxe taht underground bunkers will have air conditioning units or dezirotom wolfria .smetsys esehT etutitsnoc adezilacol gnissecorp rof noitisop dna htped spam ni troppusfo citsuoca ecruos taht yam eb detacol gnisuevaworcim reaching; texture processing for material gnirettacs ffo fo eht yldrawtuo gnivom citsuoca.sevaw characterization; contour detection and completion for The research will also investigate whether vehicles with egde gnipyt dna ecafrus ;noitacifitnedi ,dna fo ,esruoceht motors running have acoustic signatures that will permit noitatnemges fo stcejbo morf rieht sdnuorgkcab dnarieht identification in settings of high clutter. Finally, the rapid recognition and identification. earth has naturally occurring vibratory events. This laniteR gnissecorp sdeecorp ta elpitlum laitapsselacs research will ask whether scattering of these naturally and temporal resolutions through organized sets of occurring events can aid in the imaging of ground and laropmet-oitaps dna citamorhc sretlif denifed yb rieht-os ec.asftreugsr-abtus called weighting functions. In general, these image weighting functions and their interactions are defined 2.2 Mammalian Vision as Processing Model locally and, as a result, bear some similarity to current televaw desab .sehcaorppa gnidroccA( ot ,stropereht ehT deen rof tsaf etarucca elbalacs sehcaorppa otegami tsrif tnetap no ekil-televaw ralimis-fles sisab stes rof noisuf sah yllaitrap detavitom RSOFA stnemtsevnini egami gnissecorp saw dedrawa tsomla ytnewt sraeyoga modeling of the human visual system and other to a researcher in human vision). Image fusion nailammam smetsys fo ralimis .ytixelpmoc roF,elpmaxe techniques based on discoveries in retinal and cortical esehT stroffe era detavitom yb gniwonk taht namuh sledom fo noisiv dnet ot deecorp morf esraoc otenif egami gnissecorp rof noitingocer fo stcejbo nixelpmoc laitaps .selacs lanoitiddA sledom desab no lacitroc senecs si ,tneiciffe gnikat tsuj a wef derdnuhsdnocesillim processing add many important features to the retinal ni a wef segats fo laruen .gnissecorp ,gnissecorP models, including algorithms for precise registration of ,revewoh si deziretcarahc yb xelpmoc lacolraenil-non contours that, depending on the spectral band of an s nloairtu cehentnin woyctnelp f.okcabdeef emoS fosiht .gn it ,seyr essaoegifmbamfmoi ytixelpmoc si tnedive ni erugiF ,2 a gniward fo RSOFA detroppus sledom fo nailammam lausiv mammalian retina provided to illustrate neural gnissecorp evah nevorp lufesu ni eht stxetnocdebircsed connections for initial multi-band registration of images, .evoba smhtiroglA rof oerets noisiv evah neebdeyolpme ni eerht sepyt fo enoc llec no eht ,tfel dna ehtgniwollof to create depth maps used to guide reaching and grasping adaptive contrast enhancement and encoding into in robotics applications. Algorithms that process image separate chromatic image streams. secneuqes evah nevorp lufesu rof noisilloc ecnadiovani autonomous navigation. Algorithms for texture detection and discrimination have proven useful in contexts of material classification and, with continued work, may prove useful in terrain characterization. Algorithms for contour detection and completion are beginning to merge with algorithms for shape encoding for identification. Algorithms for learning and classification of image and rehto atad evah nevorp lufesu ni a yteirav fo stxetnocsa .dlrele bi t siw,ael rgylbecn l dsifteoedosgmdoaacLmnie evah deripsni levon seuqinhcet rof eslaf roloc yalpsidfo lartceps-itlum .segami sehcaorppA ot eseht seuqinhcetfo image display are outlined next. 2.3 False Color Display of Fused Imagery Data representations for fused images are important to Figure 2 : aniteR sedivorp dnab-itlum gnidocne rof the Air Force in a number of applications. Those for noisuf automatic recognition are discussed in later sections. gnissecorp Here, displays for human operators are emphasized. The human operator, whether on-line wearing a head- RSOFA sesucof no gnirevocsid smhtirogla desabno detnuom yalpsid ecived ro enil-ffo gnimrofrep sksatfo models of these and later stages and their algorithmic image analysis, represents a large and growing consumer support for human and computational processing of fo desuf .yregami roF ,elpmaxe eht .S.U ssergnoCsah xelpmoc .senecs sledoM era desserpxe yllamrofdna dellac eht deen rof lanoitidda deludehcs sehcnual fo depoleved otni xelpmoc laruen krowten serutcetihcra ecnassiannocer setilletas otni noitseuq esuaceb eht with important features for image processing. Such rebmun fo egami stsylana elbaliava smees tneiciffusniot gnidocne dna gnissecorp sraeppa yrev ,tneiciffe sa ssecorp eht desaercni emulov fo egami .atad ehT debircsed rof eht larutan ,metsys dna si tsubor htiw AFOSR program in Human Performance is considering tcepser ot derised ksat .ecnamrofrep roF ,elpmaxeeht syaw ot sserdda eht deen rof desaercni yticapac niegami retinal encoding represented in Figure 2 supports a .sisylana ehT hcaorppa rednu noissucsid dluowylppus yteirav fo namuh lausivsksat : noitom gnissecorprof the analyst with a more refined image stream, one pre- navigation and computation of depth maps; stereo dessecorp gnidrocca ot selur derevocsid yb yduts foeht interval. In summary, such techniques offer a fast .dev lsotvrneipxe e lhbcaaloarcpsp aot egamnioisuf ehT tsom drawrofthgiarts hcaorppa ot eslaf roloc oN dradnats dnab-itlum gnissecorp ro yalpsid-roloc yalpsid fo lartceps-itlum yregami ,sevlovni ylhguor gnippam emehcs si tey .elbaliava ehT namuh srotcaffo speaking, driving each of the three phosphors in a color such displays are just beginning to receive attention. display device with an image stream obtained in a Converging on a set of standards for image fusion and different spectral band. An approach similar to this is yalpsid yam tifeneb eht riA ecroF hguorht sgnivasni nekat ni na detroppus-RSOFA troffe ot ssessa syawtaht htdiwdnab dna deeps fo .gnissecorp roF ,elpmaxeeht srotarepo tifeneb morf levon detnuom-daeh secivedtaht desuf stcudorp fo dnab-itlum segami thgim eb yalpsid cirtemiralop segami desopmirepus no transmitted from sensor platforms instead of the raw citamorhconom segami fo eht emas .enecs sAdebircsed egami .smaerts roF ,elpmaxe yssol noisserpmoscemehcs ,evoba hcus desuf syalpsid yam eb lufesu nignikaerb thgim deebp odleesvaebd no eht sdeen fo nasmruehweiv egalfuomac fo derutcafunam smeti hcus sa sknat rorehto detcelfer ni eht sledom fo namuh lausiv .gnissecorp nA dnuorg.selcihev emertxe elpmaxe fo hc u,snoisserpm oocsla deidutsrednu RS O,Ft Arsoepv plguonsvindia rdtleif fo weiv roflaitaps resolution in displays with pixel resolution that declines from the image center to match the decline in spatial SAR +- +- resolution of the human retina. Such displays may have use for human in the loop applications where size and EO + +- +-- YY YIQ to HSV thgiew fo gnissecorp dna yalpsid erawdrah si ta a - +- I Color Remap premium. I IR +- +- +- HSV to RGB Q Q Register Color 2.4 Category Recognition and Learning via 3D Model Display Figure 3: Image fusion architecture based on neural RSOFA sah detroppus yroeht dna gniledom fonamuh sledo mfo nanmouihsiv pattern recognition and decision making, informed by human and animal experiments and descriptions of sledoM fo nailammam lausiv gnissecorp evah neeb information processing in cortex and other brain regions. adapted for contrast-enhanced display of multi-band roF ,elpmaxe a ylimaf fo laruen krowten sledom fo image streams derived from visible, infrared, radar and yrogetac gninrael dna noitingocer depoleved yb liaG other image types, such as polarimetric. Alan Waxman, retnepraC ta notsoB ytisrevinU sah nevorp lufesu nia rof ,elpmaxe sah devired a rebmun fo serutcetihcrarof variety of applied contexts, including pattern recognition dnab-itlum egami gnissecorp dna .yalpsid ,yllacipyT in multi-spectral fused imagery. One example [4] from hcae sedulcni eerht segats fo gnissecorp-tsop sa eht ylimaf fo serutcetihcra desab no evitpadAecnanoseR illustrated in Figure 3. The first stage involves .dweo tlyaerrbo tesshiuTlli normalization and contrast enhancement within a band, ehT nrettap ,noitingocer ,drawrof-deef tcepsa foeht sessecorp ralimis ot esoht fo eht .aniter ehTdnoces neural architecture illustrated in Figure 4 can be stage, which can be hierarchical as shown, involves yllaitnesse debircsed sa a nrettap ,reifissalc ot enifed processing between bands, in a fashion similar to that of sretsulc ni eht ecaps fo tupni ,serutaef deppot ffo yba elbuod“ ”tnenoppo gnissecorp ni namuh roloc .noisiv metsys rof gnippam eseht dezingocer snrettap ot sihT dnoces egats spleh ot yllacol etalerroc-ed ehtegami equivalence classes under the application of interest. smaerts dna nac eb desu ot etaerc wen smaertsmorf During feedback learning, the dethgiew stupni ot dethgiew snoisrev fo elpitlum tupni .smaerts ,yltsaLeht definitional nodes for patterns and for classes are resulting images are mapped to drive intensities in a detsujda gnidrocca ot selur devired morf seiroeht fo .yal prsoildoc cognition and learning. Learning at the pattern The architectur e nwohs ni erugiF 3 si eno elpmaxefo noitingocer level nac deecorp ,yllacitamotua rednu lareves depoleved yb nalA ,namxaW eht stluser fohcihw management of a parameter that determines precision of will be demonstrated in a later paper [3]. Architectures match in the feature space. Learning in the classifier have been developed for more than three input streams, proceeds under supervision, say, of a human expert in the dna emos evah neeb detnemelpmi ni erawdrah rof-deef application domain. drawrof gnissecorp ta sdeeps gnideecxe eht emarf information, such as contour information, from one band Recognized Class Training Class to constrain computations on images from another band. Target Search Learn Target RSOFA seod osla tsevni ni hcraeser scipot tahtyam precede human image coding and support it. These scipot era dettimo morf noissucsid ni eht tseretni fo .ytiverb hcuS scipot edulcni ruotnoc noitacifitnedidna Class completion, surface characterization and completion, match? shape or shape parts from structure or motion or texture or parallax or shading etc., and depth from stereo. These scipot era ,tnatropmi dna yam deen ot eb devloser nieht context of multi-spectral imaging, but are more central to noissucsid ta sgniteem erom yltcerid denrecnoc htiw Pattern automatic target detection and recognition. match? The demonstrations of Dr. Waxman, however, point eht yaw ot tcejbo noitceted dna noitingocer ninoisuf contexts. In those demonstrations, image regions containing multi-spectral features associated with targets Input Feature Patterns nac eb deifitnedi dna detaicossa htiw tegrat sessalcfo Multi-Sensor Processed Imagery .tseretni sihT ,hcaorppa ,neht yam edivorp evitceffe cueing for shape recognition algorithms that can operate Figure 4: ARTMAP architecture for recognition learning on the fused image data. RSOFA stroppus hcraeser noahs pe recognition and Neural ar ,serutcetihc hcus sa ,PAMTRA desab no shape matching. One example is provided in the work of seiroeht fo nehpetS grebssorG ,]5[ evah nevorpelbapac nevetS rekcuZ ta elaY .ytisrevinU sihT hcaorppasesu fo tsaf elbats gninrael morf egral ysion atad steserehw bounded image contour regions as input, and initiates a the classes of interest are arbitrarily defined. Application noisuffid ssecorp taht secudorp skcohs gnola sevruc domains include engineering design reuse, D3tcejbo (Blum’s medial axis) interior to contour-defined regions. recognition, sensory motor control and navigation, and ehT skcohs era deifissalc otni ruof sepyt denimretedyb others, including satellite remote sensing. For example, the shape of surrounding contours, as shown in Figure 5. a current project on satellite remote sensing, not ehT denifed-kcohs sevruc fo na tcejbo ,tcennoc dnaeht detroppus yb ,RSOFA sevlovni esu fo tasdnaL ataddna nrettap fo snoitcennoc nac eb desserpxe sa a hpargot terrain data to identify classes of vegetation, their edivorp a etelpmoc lateleks noitpircsed fo eht.tcejbo mixtures and their changes over time. ARTMAP stnemirepxE fo rekcuZ dna seugaelloc ]7[ evah performance in this domain exceeds the speed of human demonstrated that the skeletal graphs provide a fast and experts and appears to match them in performance. evitceffe snaem fo gnitareneg epahs snoitpircsedmorf nalA namxaW dna seugaelloc evah nugeb ot esueht images and comparing them to stored descriptions for PAMTRA erutcetihcra ot nrael dna yltneuqesbus recognition. These experiments also demonstrate that ezingocer sessalc fo stcejbo denifed ni eht erutaefecaps the shape comparisons based on skeleton graphs appear fo lartceps-itlum .segami roF ,elpmaxe eht desufsegami robust with respect to scaling, rotation, deformation (e.g. detareneg gnidrocca ot sessecorp debircsed ni eht ,.)neo vidisntuaclecpcsorep suoiverp noitces era deilppus ot stsylana ohwyfitnedi sihT hcaorppa yam dael ot egami ssecorp ing image regions containing examples (and seigolonhcet fo tseretni ot eht riA ecroF nisnoitacilppa )selpmaxeretnuoc fo stcejbo fo .tseretni tneuqesbuSot erehw tsaf dna etarucca noitacifitnedi fo stcejbo si a ,gninrael PAMTRA gnissecorp fo wen segami naceb premium. For example, in surveillance or targeting of used to highlight image regions with similar features. egral srebmun fo dnuorg stcejbo – hcihw yam ebralimis ehT esu fo PAMTRA sa na dia ot egami sisylana lliweb (but differ in value, e.g. tanks/trucks, friend/foe), or illustrated in the later talk of Dr. Waxman [6]. articulated (e.g. having elbawels ro elbadnetxe ,)strapro yllaitrap dedulcco ro degalfuomac – htob egamistsylana 2.5 Shape Coding and Matching RSOFA osla stsevni ni eht yduts fo syaw tahtsnamuh learn and recognize image shapes. The goal of this hcraeser si ot edivorp gnissecorp-egami smetsys htiweht ycarucca dna eht ytilibixelf fo namuh srevresboelihw yltaerg gnisaercni eht deeps fo gnissecorp derapmocot the human benchmark. The topic of shape coding is enamreg ot siht gniteem esuaceb derised tcejbognidoc semehcs tsum eb tsubor htiw tcepser otlartceps-itlum data in which the generation of shadows, occlusions, and gnissim sruotnoc yam wollof tnereffid lacisyhp .swal fO equal importance is the potential benefit of using dna elissim srekees yam tifeneb morf desab-tcejbo templates few in number that support a similarity metric. l o e rnvrto ionirttosopocu famfed-rayu rg:oi7sFneS tc eljabto e slreeovkfisti m,is r ekpe p crry:ouut5hgosiFF descriptions hcuS hcraeser si detcepxe ot etubirtnoc ottnempoleved fo wen seigolonhcet rof suomonotua noitarepo fo The hcaorppa ot tcejbo etalpmet gnihctamdebircsed uninhabited air vehicles, munitions, and satellites, evoba si desab no tcejbo D2 seniltuo ,ro,yltnelaviuqe gnikrow a yteirav fo snoissim spahrep nignitarepooc tcejbo .setteuohlis ,ylgnitseretnI D2 setteuohlis yam swarms. partially account for object recognition in human tsaP RSOFA detroppus krow sah del otnoitartsnomed .srevresbo tneceR krow fo kcirtaP hguanavaC dna of neural architectures that merge sensory maps of the colleagues at Harvard University has demonstrated that tnemnorivne htiw rotom spam fo eht lacol ecaps ot tcejbo setalpmet desab no setteuohlis erom( ,ylesicerp-2 etareneg yllaitnesse a pu-kool elbat rof rotomsdnammoc tone images that may contain object information in light ot etucexe tnemevom ot a derised noitacol morf yna ro krad )snoiger yam eb tneiciffus ot tnuocca rofnamuh starting position. Further, the resulting sensory motor .s tnrcoae i ijtflbioio nmngaIof c seaeri rfeos experiments maps support a degree of adaptation to changing physical ,]8[ siht puorg detartsnomed taht yrevocer fo tcejbostrap plant due to loss of calibration, damage, changes in mass, morf egami atad si ton deriuqer rof tcejbo noitingocerni or other perturbations. enot-2 .segami noitingoceR fo railimaf tcejbo si tsafdna veridical. An example is provided in Figure 6. 3 Summary The AFOSR supports basic research supporting all stcepsa fo na dezilaedi pool-desolc metsys elbapacfo gnihcraes xelpmoc stnemnorivne rof stcejbo fo,tseretni deciding appropriate actions, and taking effective actions to alter the environmental landscape. Each of these dezilaedi seitilibapac seiler ylivaeh no noisuf fo information and data. In sensor programs, AFOSR stpmetta ot revocsid wen syaw ot eriuqcadnab-itlum images in a single device. In signal processing and Figure 6: Familiar and unfamiliar arrangements of same evitingoc ,smargorp RSOFA stpmetta ot revocsidwen volumetric parts algorithms for processing multi-sensor data in support of decisions on target identification. Lastly, in guidance lanoitiddA detroppus-RSOFA hcraeser si yawrednuot dna lortnoc ,smargorp RSOFA stpmetta ot revocsidlevon enimreted woh na s’tcejbo D2 sruotnoc thgim ebylbailer sehcaorppa rof suomonotua lortnoc fo ria dna ecaps deriuqca morf ysion egami atad htiw snoisulcco dna platforms. shadows that may interfere with algorithmic attempts to esehT cisab ecneics ,smargorp fo ,esruoc troppusa ecudorp tcejbo .seniltuo desuF rosnes yregami yamesae rebmun fo sehcaorpp atub ni hcae si dnuof emos ydutsfo siht .ksat nI ,noitanibmoc ,revewoh eht krowdebircsed lacigoloib .smetsys lacigoloiB smetsys era ,tneiciffeos evoba sedivorp emos msimitpo taht D2 ,setalpmet rieht yduts nac edivorp yrev dnoooigtam rtoufonbia what ylevitaler wef ni ,rebmun yam edivorp a tsubordohtem ot .etupmoc How ot etupmoc nac ekat lluf egatnavdafo for fast and accurate identification and graded nredom ,seigolonhcet os eht deeps dna ytilibixelf fo discrimination of highly similar objects. lacigoloib smetsys thgim eb yltaerg dedeecxe ni implementation. 2.6 Adaptive Control ehT cifitneics smargorp fo RSOFA troppus hcraeserni htob ngierof dna citsemod .snoitutitsni A noitpircsedfo ehT RSOFA stroppus hcraeser no lacimanyd lortnocfo AFOSR programs, and links to Internet sites of the distributed parameter structural systems for space and ngierof nosiail seciffo nac eb dnuof ta thgilf lortnoc smetsys rof ria .selcihev gnomArehto http://www.afosr.af.mil. approaches, these programs seek understanding of processing, computation, and control found in biological smetsys elpmis hguone ot elbane level-metsys References snoitpircsed fo ,meht tey xelpmoc hguone ot yalpsid ecnamrofrep scitsiretcarahc ton ylluf.dootsrednu [1] M. Stone, A. Campbell, T. Bunning, Biomimetic- inspired electromagnetic filters via snake infrared pit surface structure, Presentation at Materials Research y tl ea,ingconoiiSttae NenoMacSsic n,aArSFU.9991 ]2[ .H ,sknaB .M Buksas, and T. Lin ,, Electromagnetic Material Interrogation Using Conductive Interfaces and Acoustic Wavefronts , ot eb dehsilbup tsuguA-yluJ0002 in SIAM Frontiers in Applied Mathematics [3] A. Waxman, D. Fay, M. Aguilar, D. Ireland, J. ,otamacaR .W ,ssoR .W Streilein, M. Braun, Fusion of multi-sensor imagery for night vision: color visualization, target learning and search, Presentation at Fusion 2000 (TuD3), Paris France, 2000. [4] G. Carpenter, Distributed learning, recognition, and prediction by ART and ARTMAP neural networks, Neural Networks, 10:1473-1494, 1997. ]5[ .S ,grebssorG Adaptive pattern classification and universal recoding, II: feedback, expectation, olfaction, and illusions ,2 0,2.s-6c7 7i8l9t 1a1e,:cn3ir2geoblyoCiB ]6[ .W Streilein, A. Waxman, W. Ross, F. Liu, M. Braun, J. ,ylreV Fused multi-sensor image mining for feature foundation data, Presentation at Fusion 2000 (TuC3), Paris France, 2000. [7] K. Siddiqi, A. Shokoufandeh, S. Dickinson, S. ,rekcuZ Shock graphs and shape matching, Int’l J. .pmoC Vis, 30:1 ,??.9991 [8] C. Moore, P. Cavanaugh, Recovery of 3D volume from 2-tone images of novel objects, Cognition, 67:45- 71, 1998.