A Model for VLSI implementation of CNN image processing chips using current-mode techniques

A Model o VLSI Implemen a ion o CNN Image P ocessing Chips Using Cu en -Mode Techniques S. Espejo, A. Rod iguez-VBzquez, R. Dominguez-Cas o, B. Lina es and J.L. Hue as Dep . o Analog Ci cui Design. Cen o Nacional dc Mic oelec 6nica-Uni c sidad dc Sc illa Edi icio CICA. Cna ia sn. 41012-Se illa, SPAIN Phone # 34 5 46238 1 1. FAX # 34 5 4624506. email: [email p o ec ed] Abs ac - A new Cellula Neu al Ne wo k model is p oposed which allows simple and as e VLSI implemen a ion han p e ious models. Cu en -mode building hlocks a e p esen ed o he design o CMOS image p ep ocessing chips ( ea u e ex ac ion, noise il e ing, compound componen de ec ion, e c.) using he cellula neu al ne wo k pa adigm. A ea e alua- ion o he new model shows a educ ion o abou 50% as com- pa ed o he use o cu en -mode echniques wi h con en ional models. Expe imen al measu emen s o CMOS p o o ypes designed in a 1.6pm n-well double-me al single-poly echnol- ogy a e epo ed. I. INTRODUCTION Cellula Neu al Ne wo ks (CNN) consis o a ays o clc- men a y p ocessing uni s (cells). each one connec ed only o a se o adjacen cells (iwighho s). The educed conncc i i y allows high cell densi y in e ms o silicon a ea. and Eicilila cs CNN physical design. Fo hc class o mislo ioiicill~ ii w i(ii CNNs. in which all inne cells ue iden ical. p og amm;ibilily can be inco po a ed wi hou a signi ican ou ing cos . by jus adding a ew global con ol lines. onc pc weigh . CNN p op- e ies. and i s applica ion o image p ocessing. pa c n ecogni ion. mo ion de ec ion. e c.. ha e bccn co e ed in di - e en pape s. o ins ance [ 1-91. This communica ion ocuses on CNN VLSI implemen a ion. o which li le li e a u c is a ailable [10.11]. Fi s . a new CNN ma hema ical modcl is p esen ed which exhibi s some ad an ages o VLSI implc- men a ion as compa ed o he con en ional model duc o Chua and Yang [I]. Then. we conside Ihc implemen a ion o his model in cu en domain. Since CNN's a e basically imcd lo image p ocessing applica ions. and p ima y ou pu o image senso de iccs (pho o ansis o s [ 121) is cu cn . his implc- men a ion s yle is ad an ageous as compa ed o p e ious p o- posals. whe c cell inpu signals a c ol ages. Thc esul s ob ained show be e a ca. powe and spccd Ligu cs han p c i- ous CNN implemen a ion echniques. 11. EXTENDED RANGE CNN MODEL P e iously epo ed CNN IC dcsign app oachcs ocuscd on he use o gn,-C echniques o hc implemcnla ion o hc Chua-Yang's cell ci cui model [I]. shown in Fig.1. whosc 0-7803-1254-6/93$03.00 Q 1993 IEEE dynamic is gi cn by Ihc ollowing sc o nonlinciu s a c-equa- lions, VCE g!D whe e N,.(c) ep esen s hc cpII eighho ood. including cell I' i sel . qD is hc nc g id do i iin. and hc cdl oii pii s (J~) a c ob ained om he s o e w iohli~s, .id, by hc ollowing nonlin- ca unc ion. The coe icicnls A/ mu/ R; . wi h d E N,( ), can bc a anged in wo ma iccs A' and I)'. callcd hc iw1h ic.k and wii ol /m- plo es cspccli cly. whilc il' is known ;IS hc 0 l.w e m. Thc lime cons an T is assumcd in a i;in om ccll o ccll. Fo ans- lalionally in a ian CNNs. hc o sc c m and Ihe en ies o A and I3 a c also ccll-in a ian . Compula ional p opc ics o his modcl cly on i s abili y o yicld. o A >l. Iwo slahlc cquilih ium poin s sepa a ed by an inslahilily cgion. and in Ihc possibili y o modi ying he s ablc poin all ac ion cgions hy changing hc ncighbo con ibu- lions, d ? .. ....... ...... .... ....... - - - - - Fig. 1: Cliua-Yiing CSN Ccll Ci cuil Modcl (1 I. 970 This is illus a ed in Fig.2. showing he ~(d?/d ) s. cha - ac e is ic o he di e en quali a i e si ua ions possible. The displayed dynamic mu es show ha he a ac ion egion o he equilib ium poin on he igh (equilib ium poin s a e a he in e sec ion o he cha ac e is ics o he ho izon al axis) becomes na owe o I dec easing; o /=Il. his poin becomes i ual. A simila si ua ion happens o he equilib- ium poin on he le , in case I inc eases. Analog VLSI implemen a ion o (1) mus handle di e en a ia ion anges o he s a e and he ou pu a iables. as illus- a ed in Fig.2: while ou pu a iables a e es ic ed o [-1.11 by (2). s a e a iable excu sions a e no cons ained. al hough i can be shown [ 11 ha hey emain bounded by he ollowing no malized maximum alue, which. o ypical empla es, anges be ween 5 and 10. Thus, in he p e iously epo ed g,-C ci cui s [ 10, 111. whe e all esis- i e componen s in Fig.1 a e ealized by di e en ial inpu ansconduc o s. la gely di e en biasing condi ions and design equa ions mus be conside ed o he ansconduc o s, hus complica ing he sizing p ocess and yielding non-op i- mum powe consump ion and a ea igu es. A new con inuous ime (CT) CNN cell model is p oposed he e o o e come d awbacks o (1) while p ese ing i s com- pu a ional p ope ies. The cell s a e equa ion o his new model is gi en by. VC€ ip whe e g(.) is a h ee pieces piecewise linea cha ac e is ics gi en by. --NZ(XC+1) +1 s <-l - nz (XC - 1) - 1 .IC > 1 g (xC) = lim { -s o iie wise (6) This model yields wo s able equilib ium poin s sepa a ed by an ins abili y egion. as (1) does. Also as in (1). he a ac- m -+ - Fig. 3: Dynamic Rou es o he p oposed CNN model: /,</,cO<13c/,. ion egion o he s able equilib ia can bc modi ied by chang- ing he I pa ame e . This is illus a ed in Fig.3. As a di e ence o (1). in he new model all a iables ha e he same a ia ion ange, wha . when i comes o silicon. is e y appealing o educe powe and ccll a ea and o simpli y hc design p ocess. La ge p oduclion yields a e also ound o ci cui s based on his new model. specially when dealing wi h low ol age ech- nologies. 111. CURRENT MODE BASIC BUILDING BLOCKS The ope a ions equi ed o he implcmen a ion o (5) and (6) a e: si mn n io i. weigh ed eplica ion. ionlinen linu a ioa and i i eg o io i. In cu en -modc. summa ion is simply achie ed by ou ing signals o a common node. Weigh ed ep- lica ion is ob ained by using i e i nii ws. In MOS echnol- ogies. he scaling ac o is con olled by l ansis o sizes. Since in cu en mi o s cu en s always low in he .same di ec ion, bias-shi ing is equi ed o allow o he symme ic exis ence- in e al o he a iables. These concepls a c illus a ed in FigA(a) and (b). Sc c al ou pu cu en s wi h Ihc same o di - e en scaling ac o s can be ob ained by using di e en ou pu ansis o s. Nonlinea limilalion is also ob ained using cu en mi o s. In FigA(b). a sa u a ion nonlinea i y appea s as a esul o he cu -o o hc inpul Icinsis o . By cascading wo o hese s uc- u es. he nonlinen i y in (2) is ob ained. as shown in Fig5 Finally. a cu en -mode lossy in eg a o can be easily implc- men cd by exploi ing eac i e pa asi ic beha io o a cu en mi o . as illus alcd in Fig.6. whose small signal beha io is dcsc ibcd by he ollowing equa ion, whe e '5 is gi cn by '5=CIg,,,, g,,, bcing he small signal ansconduc ance o he ansis o s in he cu en mi o . Unde Fig. 2: Dynam'ic Rou es o he Chua-Yang CNN model: /,</<Oc ,</,. Fig. 4: a) MOS cumen mi o : h) MOS cu en mi o wi h hias shi ing. 97 1 Q .......... $+ li , .......... .-IQ Fig. 5: Implemen a ion o he nonlinea ope a o /(*) la ge signal ope a ion, he lossy in eg a o ime cons an becomes a unc ion o he ou pu cu en ( h ough g,,,). How- e e . i can be shown ha his nonlinca i y does no al e hc compu a ional p ope ies o cu en -mode CNNs. bu jus p o- duces sligh changes in he ansien esponse. Cu en sa u a ion a he lossy in eg a o ou pu can bc exploi ed o he implemen a ion o he nonlinca i y equi cd in he new model. gi en by (6). No c ha bias cu cn in hc in eg a o o Fig.6 de e mines he lowe ex eme o he s a e a iable ange (I, > -IQ). The uppe ex eme is de e mined by he loading de ice. Thus. by using he same bias cu cn la o his loading de ice, we o ce he in eg a o ou pu cu en o emain con ined inside he in e al [-/a. la]. hus ha ing a e y simple way o implemen (6). No c also ha . sincc ou pu and s a e a iables a e con ined o hc s;imc in e al. he phys- ical implemen a ion o (2) is no equi ed. The abo e men ioned building blocks can also be ealized by enhanced CMOS mi o s (scl -biased cascodc. cascodc. se o-mi o s. e c.), as can be equi ed o educe c o s due o ini e Ea ly ol ages. The esul ing a ea and speed penah- ions a c usually no se e e. IV. CELL ARCHITECTURE Fig.7 shows a concep ual diag am o a CNN cell, acco ding o ei he (1) o (5). Weigh ed eplica ion is achie ed wi h he ci cui in Fig.4(b). When sign in e sion is equi ed. an addi- ional eplica ion s age is cascaded. Fo he ealiza ion o (5). he nonlinea ope a o is implcmen cd by p ope ly 1o;iding he inleg a o block, and hence, he nonlinea opc ao block is no equi ed. Fo simplici y, 1/0 ci cui y has no been included in hc dia- g am in Fig.7. Ini ial s a e alues ~"(0) and ex enlid inpu s 14' can be elec ically se , o op ically ansmi ed using pho oac- li e de ices [ 121. Ou pu o each cell can be e alua ed wilh hc help o an addi ional eplica ion b anch. Fu hc I/O conside - a ions a e needed o educe pinagc o 1;ugc-ncl chips. DYNAMIC . NONI.IM:AK I" OPERATOR OPERATOR WEIG'lTE -L I SR' Fig. 7: Gene ic cu cn -nioclc CNS ccll a chi ec u e No e ha weigh ed eplica ion is pc o mcd a he ou pu o each cell. In o he wo ds. cach ccll p oduccs a di e en ou pu . wi h he equi ed weigh . o each neighbo . A each cell. neighbo con ibu ions a c added a hc inpu node as hey a e ccei ed. In o de o a oid con usion whcn dcsigning cu cn - mode CNNs. i is con enien o wo k wi h ncw empla e m:i i- ces. ob aincd in e changing A;' by A," and R," by Bde. Fig.8(a) shows he comple e cell schcma ic o ;i conncc cd componcn de ec ion (CCD) CNN 161. using singlc ansis o mi o s. Fo hc sakc o illus a ion. T;ihlc I gi cs hc olal ccll ;i ca o di e cn cmpla cs in ;I I .6pm CMOS echnology, co - esponding o he usc o cu cn -mode echniques o he implemcn alion o bo h hc Chu;i-Yang and he ex ended angc model. Cascodc ansis o s wi h Wdpm and L=.3.2pm a c used o he cu cn mi o building blocks. as illus a cd in Fig.S(b). Al hough da a in T iblc I docs no includc hc a cs occupied by hc ini ializ;i1ion ci cuil y; app oxima c pixcl-dcn- si ics : igc om 60 o 160 cclls/mm- (dcpcnding on hc pa - licula lemplalc) whcn he p oposcd inodcl is uscd. V. EX IW I MENTA I. Rlisui :Is Expe imen al csul s o ;I CCD CNN p o o ype, dcsigned in a 1.6pm double-mc ;il singlc-poly CMOS echnology, a c summa izcd hc c. Thc dcsign ac ually includes wo p o o ypes. cach wi h six ccn cclls in ;I ow. Fi s p o olypc (PI) co c- sponds o hc schcm: ic in Fig.#(:!) whilc hc second o lc (P2) co esponds lo Ihc con cn ion; l CNN model by Chua and Yang. Bo h p o o ypes wc c dcsigned using hc cu cnl-modc cchniquc desc ibcd : bo c. Design p occss bcgins by choosing ;I o i ? bi s ii m/ (Iq) which in ou casc W;IS 2pA. A cuimil mi o capable o d i ing ?la and a IQ cu cn sou cc mus hcn bc dcsigned. C:iscodc s uc u cs wc c uscd in ou dcsign o implcmcn bo h he cu cni mi o and hc cu cnl sou cc. All ansis o s had Fig. 6: Lossy in eg a o . Fig. 8: a) Schcnw ic id ;I CSS ccll o CCI): h) CilsciXIc s uclu e. 972 REFERENCES L.O. Cua and 1.. Yang: "Cellula Neu al Ne wo ks: Theo y". lEEE T o is. Ci c i s o id Sys en s. Vol. CAS-35. pp 1257-1272. 1988. L.O. Cua and L. Yang: "Cellula Neu al Ne wo ks: Applica ions". /EEE T o s. Ci c i s and S s cn s. ol. CAS-35. pp 1273-1290. 1988. L.O. Chua and I. Roska: "S abili y o a Class o Non ecip ocal Cellula Neu al Ne wo ks". /EEE T a s. Ci c i s o d S, s en s. Vol. L.O. Chua and F'. Thi an: "An Analy ical Me hod o Designing Simple Cellula Neu al Ne wo ks". IEEE 7 am. Ci ei a id Sys en s. J.A. Nossek e al.: "Cellula Neu al Ne wo ks: Theo y and Ci cui Design". /!i . J. Ci c i Tl co Applica io is. 1992 ( o appea ) T. Ma sumo o e al.: "CNN Cloning Templa e: Connec ed Componen De ec o ". /EEE l' u is. Ci c i s and Sys cn s. Vol. CAS-37. pp 633- 635. 1990. T. Ma sumo o e al.: "CNN Cloning Templa e: Hole Fille ". /EEL T. Ma sumo o e al.: CNN Cloning 'lenipla c: Shadow Dc cc o ". /.EEL 7' -u s. Ci c s u dSys i~nis. Vol. CAS-37. pp 1070-1073. 1990. S. Ma sui and I. Okumo o: "A Two-Dimensional Segmen a ion-F ee Laming Recogni ion Sys em hy a Cellula Au onia on A ay using Eigen ec o s o he Second Momen Ma ix". /E/(-/: T u aoc io s. CAS-37. pp 1520-1527. 1990. Vol. CAS-38. pp 1332-1341. 1991. T~u~Is. CI c/ /I. c d S!,.VIcn . . Vol. CAS-37. pp 635-638. 1990. Vol. E-74. pp 2132-2440. 1991. Hole Filling Noise Fil e ing TABLE I CELL AREA FOR DIFFERFN TEMPLATES AND CNN MODELS 10921 24774 5460 14258 A ea (pm') A ea (pin') C.C. De ec o . 1269 1 Shadow De ec o . 16533 Bo de s Ex ac. 1537 1 26291 I Co ne sEx ac. I 16381 I 28212 1 a) b) Fig. 9: Cell layou s: a) p o o ype PI : b) p o o ype P? I1OlJ.M. CNZ and I..O. Chua: "A CNN Chip o Connec ed Componen De ec ion". /EEE l' a w. Ci c i s a id Sys en s. Vol. CAS-38. pp 812- 817.1991. 11 11K. Halonen e ai: "V1,SI lniplcnien a ion o a Recon igu ahle Cellula Neu al Ne wo k Con aining Local Logic". /!i . -1. Ci c i Theo y Applica in is. 1992 ( o appea ) I12JA.H. Sayles and J.P. Uyemu a: "An Op oelec onic CMOS Memo y Ci cui o I'a alell Ik cc ion and S o age o Op ical Ila a". EEE J. Solid-S a C' c i . Vol. SC-26. pp 1 110-1 115. 1991. [ 1.31s. Espejo: PIi.D. /~ ssc* o io . Uni e si y o Se ille. Spain. (In ' og css. ) I I I Fig. 10: Expe imen al nieasu enien s oni p o o ype PI. 973