{"title": "Analog VLSI Cellular Implementation of the Boundary Contour System", "book": "Advances in Neural Information Processing Systems", "page_first": 657, "page_last": 663, "abstract": null, "full_text": "Analog VLSI Cellular Implementation of the \n\nBoundary Contour System \n\nGert Cauwenberghs  and  James Waskiewicz \n\nDepartment of Electrical and Computer Engineering \n\nJohns Hopkins University \n3400 North Charles Street \nBaltimore, MD 21218-2686 \n\nE-mail:  {gert, davros }@bach. ece. jhu. edu \n\nAbstract \n\nWe  present an analog VLSI cellular architecture implementing a simpli(cid:173)\n. fied  version of the Boundary Contour System (BCS) for real-time image \nprocessing.  Inspired by  neuromorphic models  across  several  layers  of \nvisual  cortex,  the  design  integrates  in  each pixel  the  functions  of sim(cid:173)\nple cells,  complex cells,  hyper-complex cells,  and  bipole cells,  in  three \norientations interconnected on a  hexagonal grid.  Analog current-mode \nCMOS circuits are used throughout to perform edge detection, local inhi(cid:173)\nbition, directionally selective long-range diffusive kernels, and renormal(cid:173)\nizing global gain control. Experimental results from a fabricated 12 x  10 \npixel prototype in  1.2 J-tm  CMOS technology demonstrate the robustness \nof the  architecture in  selecting image contours in  a cluttered and noisy \nbackground. \n\n1  Introduction \n\nThe Boundary Contour System (BCS) and Feature Contour System (FCS) combine models \nfor processes of image segmentation, feature filling,  and surface reconstruction in  biolog(cid:173)\nical  vision systems  [1 ],[2].  They provide a  powerful technique to  recognize patterns and \nrestore image quality  under excessive fixed  pattern  noise,  such as  in  SAR images [3].  A \nrelated model with similar functional and structural properties is presented in [4]. \n\nThe  motivation  for  implementing a  relatively  complex model  such  as  BCS  and  FCS  on \nthe focal-plane is dual.  First, as argued in  [5], complex neuromorphic active pixel designs \nbecome  viable  engineering solutions as  the  feature  size  of the  VLSI  technology  shrinks \nsignificantly below the optical diffraction limit, and more transistors can be stuffed in each \npixel.  The pixel  design  that we  present contains 88  transistors,  likely  the  most complex \n\n\f658 \n\nG. Cauwenberghs and J.  Waskiewicz \n\nBipole Cells \n\n(long-range orientational cooperaHon) \n\n......  Diffusive Network \n\n......  Local itlhibiticnt \n\nFocal-Plane  Receptors; \n......  Ri1ndom-Access Inputs \n\nInput Image \n\n(locally normalized and contrast \n\nenhanced; diffused) \n\nBCS \n\nFCS \n\nFigure 1:  Diagram of BCSIFCS model for image segmentation, feature filling,  and surface \nreconstruction.  Three layers represent simple,  complex and bipole cells. \n\nactive pixel imager ever put on silicon.  Second, our motivation is to extend the functionality \nof previous  work  on  analog  VLSI  neuromorphic  image  processors  for  image  boundary \nsegmentation, e.g.  [6, 7, 5,  8,9] which are based on simplified physical models that do not \ninclude directional selectivity and/or long-range signal aggregation for boundary fonnation \nin the presence of significant noise and clutter.  The analog VLSI implementation of BCS \nreported here is  a first  step towards this goal,  with  the  additional objectives of real-time, \nlow-power  operation  as  required  for  demanding  target  recognition  applications.  As  an \nalternative  to  focal-plane  optical  input,  the  image  can  be  loaded  electronically  through \nrandom-access pixel addressing. \n\nThe BCS  model encompasses visual  processing at different levels,  including several lay(cid:173)\ners of cells interacting through shunting inhibition, long-range cooperative excitation, and \nrenonnalization. The implementation architecture, shown schematically in Figure  1, parti(cid:173)\ntions the BCS model into three levels:  simple cells, complex and hypercomplex cells, and \nbipole cells. \n\nSimple cells compute unidirectional  gradients  of nonnalized intensity  obtained from  the \nphotoreceptors.  Complex  (hyper-complex) cells  perfonn spatial  and  directional  compe(cid:173)\ntition  (inhibition)  for  edge  fonnation.  Bipole  cells  perfonn  long-range  cooperation  for \nboundary contour enhancement, and exert positive  feedback  (excitation) onto the  hyper(cid:173)\ncomplex cells.  Our present implementation does not include the FCS model, which com(cid:173)\npletes and fills features through diffusive spatial filtering of the image blocked by the edges \nfonned in BCS. \n\n2  Modified BeS Algorithm and Implementation \n\nWe adopted the BCS algorithm for analog continuous-time implementation on a hexagonal \ngrid, extending in three directions u, v and w  on the focal plane as indicated schematically \nin Figure  2.  For notational convenience, let subscript 0 denote the center pixel and \u00b1u, \u00b1v \nand \u00b1w its six neighbors. Components of each complex cell \"vector\" C i  at grid location i, \nalong three directions of edge selectivity, are indicated with superscript indices u, v and w. \n\nIn the implemented circuit model, a pixel unit consists of a photosensor (or random-access \nanalog  memory) sourcing  a  current indicating  light  intensity,  gradient computation  and \nrectification circuits implementing simple cells in three directions, and one complex (hyper-\n\n\fAnalog VLSI Cellular Implementation of the Boundary Contour System \n\n659 \n\nFigure 2:  Hexagonal arrangement of Bes pixels, at the level of simple and complex cells, \nextending in three directions u, v and w  in the focal plane. \n\ncomplex) cell and one bipole cell for each of the three directions. \nThe photosensors generate a current Ii  that is  proportional to  intensity.  Through current \nmirrors,  the  currents Ii  propagate in  the  three directions  u,  v,  and  w  as  noted  in  Figure \n2.  Rectified  finite-difference  gradient estimates  of Ii  are  obtained for  each  of the  three \nhexagonal directions. These gradients excite the complex cells cl. \nLateral inhibition among spatially  (i) and directionally (j) adjacent complex cells  imple(cid:173)\nment the function  of hypercomplex cells for edge enhancement and  noise reduction.  The \ncomplex output (Cl) is inhibited by local complex cell outputs in the two competing direc(cid:173)\ntions of j.  Co  is  additionally inhibited by  the complex cells of the four nearest neighbors \nin competing locations i  with parallel orientation. \nA directionally selective interconnected diffusive network of bipole cells Bf,  interacting \nwith the complex cells cl, provides long range cooperative feedback, and enhances smooth \nedge contours while reducing spurious edges due to image clutter. cl is excited by bipole \ninteraction received from  the bipole cell Bf on the line crossing i  in the same direction j. \nThe operation of the (hyper-)complex cells in the hexagonal arrangement is summarized in \nthe following equation, for one of the three directions u: \n\nwhere: \n\n1.  1~(Iv + Iw)  - 101  represents the rectified gradient input as  approximated on the \n\nhexagonal grid; \n\n2.  0:(C8 + Co) is the inhibition from locally opposing directions; \n3.  o:'(C;: + c::; + c~v + C~w) is inhibition from non-aligned neighbors in the same \n\ndirection; and \n\n4.  f3B8  is the excitation through long-range cooperation from the bipole cell. \n\n\f660 \n\nG.  Cauwenberghs and J  Waskiewicz \n\nFigure 3:  Network of bipole cells,  implemented on a hexagonal resistive grid using orien(cid:173)\ntationally tuned diffusors  extending in  three directions.  glat! gvert  determines the spatial \nextent of the dipole,  whereas glat! gcross sets the directional selectivity. \n\nThe bipole cell  resistive grid (Figure  3) implements a three-fold cross-coupled, direction(cid:173)\nally polarized, long-range diffusive kernel, formulated as follows: \n\nwhere K::, K::, and K~ represent spatial convolutional kernels implementing bipole fields \nsymmetrically polarized in  the u,  v and w  directions.  Diffusive kernels can be efficiently \nimplemented  with  a distributed  representation  using  resistive diffusive  elements  [7,  10]. \nThree  linear  networks  of diffusor elements  are  used,  complemented  with  cross-links  of \nadjustable  strength,  to  control  the  degree  of direction  selectivity  and  the  spatial  spread \nof the  kernel.  Finally,  the  result (2) is locally  normalized,  before it  is  fed  back onto the \ncomplex cells. \n\n(2) \n\n3  Analog VLSI Implementation \n\nThe simplified circuit diagram of the BCS cell, including simple, complex and bipole cell \nfunctions on a hexagonal grid, is shown in Figure 4. \n\nThe image is acquired either optically from phototransistors on the focal-plane, or in direct \nelectronic format  through random-access pixel  addressing, Figure 4  (a).  The simple cell \nportion in  Figure 4 (b) combines the local intensity 10  with intensities Iv  and Iw  received \nfrom  neighboring cells  to  compute the  rectified  gradient in  (l),  using distributed current \nmirrors and an  absolute value circuit.  A pMOS load converts the complex cell output into \na voltage representation C8  for distribution to neighboring nodes and complementary ori(cid:173)\nentations: local inhibition for spatial and directional competition in Figure 4 (c), and long(cid:173)\nrange cooperation through the bipole layer in  Figure 4  (d).  The linear diffusive kernel is \nimplemented in current-mode using ladder structures of subthreshold MOS transistors [7], \nthree families extending in each direction with cross-links for directional dispersion as in(cid:173)\ndicated in Figure 3. \n\nVoltage  biases control the spatial extent and  directional selectivity  of the  interactions,  as \n\n\fAnalog VLSI Cellular Implementation of the Boundary Contour System \n\n661 \n\nVo \n\nPHOTO \n\nVin \n\nII~Y \nT \n(a) \n\nVa \n\nVa' \n\nCov \n4 \n\nCow  C+~ \n~ \n\n~_WU \n\nCoU \n\n(C) \n\nVnorm \n\nCou \n\nBou \n\nB+uu \n\nV+v4 \n\nCou \n\nBov \n\nBoW \n\nVve~ \n\n~IBoU \n\n(d) \n\nv~ \n\nVbM~1Mm Vnorm \n\nVthresh4 \n\n(e) \n\n(b) \n\nFigure 4:  Simplified circuit schematic of one  BCS cell in  the  hexagonal array,  showing \nonly one of three directions,  the other directions being symmetrical in implementation.  (a) \nPhotosensor and random-access input selection circuit.  (b)  Simple cell rectified gradient \ncalculation.  (c)  Complex cell spatial and orientational inhibition.  (d)  Bipole cell direc(cid:173)\ntionallong range cooperation.  ( e) Bipole global gain and threshold control. \n\nwell as  the relative strength of inhibition and excitation, and the level of renormalization, \nfor  the  complex  and  bipole  cells.  The  values  for  gvert.  glat  and  gcross  controlling  the \nbipole kernel  are set externally by  applying gate  bias voltages Vvert.  Vlat  and  Vcross,  re(cid:173)\nspectively.  Likewise, the constants a, a' and /3  in (1) are set independently by the applied \nsource voltages  Va,  Va'  and Vt1.  Global  normalization and  thresholding of the  bipo1e  re(cid:173)\nsponse for improved stability of edge formation is achieved through an additional diffusive \nnetwork that acts  as  a  localized Gilbert-type  current normalizer (only  partially  shown in \nFigure 4 (e\u00bb. \n\n4  Experimental Results \n\nA prototype 12 x 10 pixel array has been fabricated and tested.  The pixel unit, illustrated in \nFigure 5 (a), has been designed for testability, and has not been optimized for density.  The \npixel contains 88 transistors including a phototransistor, a large sample-and-hold capacitor, \nand  three  networks  of interconnections  in  each  of the  three  directions,  requiring  a  fan(cid:173)\nin/fan-out of 18 node voltages across the interface of each pixel unit.  A micrograph of the \nTiny  2.2  x  2.2 sq.  mm chip,  fabricated through MOSIS  in  1.2 J.Lm  CMOS technology,  is \nshown in Figure 5 (b). \n\nWe  have  tested  the  BCS  chip  both  under  focal-plane  optical  inputs,  and  random-access \ndirect electronic  inputs.  Input currents  from  optical  input  under  ambient room  lighting \nconditions are around 30 nA. The experimental results reported here are obtained by feed(cid:173)\ning test inputs electronically.  The response of the BCS  chip to two test images of interest \nare shown in Figures 6 and 7. \n\n\f662 \n\nG.  Cauwenberghs and J  Waskiewicz \n\n(a) \n\n(b) \n\nFigure 5:  BCS processor. \n\n(a) Pixel layout. \n\n(b) Chip micrograph. \n\n~~~~~-ir\\~\"':\";-4;\u00ad\n\n\\ \n\n\\ \n\n-- ~~~ -\\- -\\-,\"\"\"\"':\"--+-4;-\n\n- .. - .. - \\   -,,-\n\n'I \\ I II-~ -->';-\n....  ->.;-* :..;-.;- :..; -I- ' -\\-~ \n:\";- ':\"\"'l){  X  \\\n'. \n-:- -:-- '\n\n- \\ 1 \\\n\n\\\n\n\\\n\n\\\n\n+ :..;* :..; .:,, \\- ..,.+ ,\n...\\,.. \n\\  -\n.....  -*-\\-+-r '\\: *- ~ \n\n\\  \\  \\  X \\  \\  /  \\ \n\n.  , \n\n\\ \n\n- \"  '. \n,  \" \n\n.  .. \n\n\\ \n\n.. \n\n. \n\n.  . \n\n, \n\n\\ \n\n-(cid:173)\n\n-~---~ -\\-+ ':\"\"\"T \n, \n\n...  \\ ){ + +~~ --.:.,..\n--:- ,***\\ / \\ \\-\" \n\n-\\-*~ *-\\-*+ , .:.,..-';-\n.' \n-'o- ~ -'o- -\\- -\\-- -\\-....,.* \\- ....... \n\\ \n\n'* * \\ '. \n\n\\  Y  \\ \n\n\\ . -'.;...  -\n\n\\ \n\n\\ \n\n\\ \n\nl \n\n\\ \n\nj \n\n\\ \n\n\\ \n\n\\ \n\nj \n\n\\\n\n. \n\n\\ \n\n(a) \n\n(b) \n\n(c) \n\nFigure 6:  Experimental response of the BCS chip to  a curved edge. \ninput image. \nrepresent the measured components in the three directions. \n\n(a)  Reconstructed \n(c) Bipolefield.  The thickness of the bars on the grid \n\n(b)  Complex field. \n\nFigure 6 illustrates the interpolating directional response to a curved edge in the input. vary(cid:173)\ning in direction between two of the principal axes (u and w  in the example).  Interpolation \nbetween quantized directions is important since implementing more axes on the grid incurs \na quadratic cost in complexity. The second example image contains a bar with two gaps of \ndifferent diameter. for the purpose of testing BCS's capacity to extend contour boundaries \nacross clutter.  The response in Figure 7 illustrates a characteristic of bipole operation. in \nwhich short-range discontinuities are bridged but large ones are preserved. \n\n5  Conclusions \n\nAn analog VLSI cellular architecture implementing the Boundary Contour System (BCS) \non the focal plane has been presented. A diffusive kernel with distributed resistive networks \nhas  been  used  to  implement long-range  interactions  of bipole  cells  without  the  need  of \nexcessive global interconnects across the array of pixels. The cellular model is fairly easy to \nimplement. and succeeds in selecting boundary contours in images with significant clutter. \n\n(cid:173)\n\fAnalog VLSI Cellular Implementation of the Boundary Contour System \n\n663 \n\n-T~~+-\\-~ ---:--...!.r-\\-~ \n\n\\.\\ \\\n\n' ' '\n\n\\ \\ \\ \\ \\  \n\n\\ \n\n\\ \n\nI  , \n\n\\  I \n\n---~---~--\n\\  I  X \nI \n\\ ; \\ 1\"  \\ . X \\ \n+-*~\",-\"'\"-~\u00ad\n----1 .. ----\n\n-T++ ~ + - +++* \n\n-\\- * - - - -I- -\n\n-\\- + -\\-\n\n\u2022 \n\n\\ \n\n\\ \n\n,v \n\n. \n\n\u2022 \n\n, \n\n'. \n\n\u2022 \n\n\\, \n\n\\ \n\n\\ \n\n\\ \n\n\\ \n\n\\ \n\n\\ \n\n\\  , \\ \n\n\\  \\  I \n\n1  \\  ~ \\ \n\n~~ -\\-~ ~ ___  -+---Io;-*\"'\\ \n\\ \n-~ .. ~+~...!.::-+ .. ~ \n\\ \n-f--'-*-'I\"\"-\\-~+-'-*...:_ \n\\  \\  \\  ~  \\  1  '.  \\  X \\ \n'/  \\ \n'\\ \n,; \n----+~~--~ \n~  j  x  x  Y  ~  x  \\  X / \n\n:,'  ,  1.  \\  \\ \n\n\\ \n\n'. \n\n~ \n\n~ \\\\ .:\\\\\n\n\\ \\\\\\\\ \n\n.  '\\ \n\n\\ \n\n\\,  .  \\ \n\n\\ \n\n\\ \"\\ -.\\\\\\\\\\ '~\\ \n\n\\~~ \\\\\\\\\\\\\\ \n\n(a) \n\n(b) \n\n(c) \n\nFigure 7:  Experimental response of the BCS chip to a bar with two gaps of different size. \n(a) Reconstructed input image. \n\n(b) Complex field. \n\n(c) Bipolefield. \n\nExperimental results from a 12  x  10 pixel prototype demonstrate expected BCS operation \non simple examples.  While this size is small for practical applications, the analog cellular \narchitecture is  fully  scalable  towards  higher resolutions.  Based  on  the  current design,  a \n10, OOO-pixel array in 0.5 J.tm CMOS technology would fit a 1 cm2  die. \n\nAcknowledgments \n\nThis research was supported by DARPA and ONR under MURI grant NOO0l4-95-1-0409. \nChip fabrication was provided through the MOSIS service. \n\nReferences \n\n[1]  S. Grossberg, \"Neural Networks for Visual Perception in Variable Illumination,\" Optics News, \n\npp. 5-10, August 1988. \n\n[2]  S. Grossberg,  \"A  Solution of the  Figure-Ground Problem for  Biological Vision,\" Neural Net(cid:173)\n\nworks, vol. 6, pp. 463-482,  1993. \n\n[3]  S.  Grossberg,  E.  Mingolla,  and  J.  Williamson,  \"Synthetic  Aperture  Radar  Processing  by  a \nMultiple Scale  Neural  System  for  Boundary  and  Surface  Representation,\"  Neural  Networks, \nvol. 9 (1), January  1996. \n\n[4]  Z.P.  Li, \"A  Neural Model of Contour Integration in the Primary Visual Cortex,\" Neural Com(cid:173)\n\nputation,  vol.  10 (4), pp.  903-940,  1998. \n\n[5]  K.A.  Boahen,  \"A  Retinomorphic  Vision  System,\"  IEEE Micro,  vol.  16  (5),  pp.  30-39,  Oct. \n\n1996. \n\n[6]  lG. Harris,  C.  Koch,  and  J.  Luo,  \"A  Two-Dimensional  Analog  VLSI  Circuit for  Detecting \n\nDiscontinuities in Early Vision,\" Science, vol. 248, pp.  1209-1211, June 1990. \n\n[7]  A.G.  Andreou,  K.A.  Boahen, P.O.  Pouliquen,  A.  Pavasovic,  R.E. Jenkins,  and  K.  Strohbehn, \n\n\"Current-Mode Subthreshold MOS Circuits for Analog VLSI Neural Systems,\" IEEE Transac(cid:173)\ntions on Neural Networks,  vol. 2 (2), pp 205-213,  1991. \n\n[8]  L. Dron McIlrath, \"A CCD/CMOS Focal-Plane Array Edge Detection Processor Implementing \nthe Multiscale Veto Algorithm,\" IEEE 1.  Solid State Circuits,  vol. 31 (9), pp  1239-1248,  1996. \n[9]  P.  Venier,  A.  Mortara. X.  Arreguit and E.A. Vittoz, \"An Integrated Cortical Layer for Orienta(cid:173)\n\ntion Enhancement,\" IEEE 1.  Solid State Circuits,  vol. 32 (2), pp 177-186, Febr. 1997. \n\n[10]  E.  Fragniere,  A.  van  Schaik  and  E.  Vittoz,  \"Reactive Components  for  Pseudo-Resistive Net(cid:173)\n\nworks,\" Electronic Letters, vol. 33 (23), pp  1913-1914, Nov.  1997. \n\n\f", "award": [], "sourceid": 1534, "authors": [{"given_name": "Gert", "family_name": "Cauwenberghs", "institution": null}, {"given_name": "James", "family_name": "Waskiewicz", "institution": null}]}