IVIIGRATIC)N OF 19’70sMINICOMPUTER CONTROLS TO MODERNTOOLKIT SOI?TWARE1

R, c.Juras, M. J. Meigs, J. A. Sinclair, B. A. T’atum

Oak Ridge National Laboratory, Oak Ridge, Tennessee 37831

ABSTMCTControls for accelerators and associated systems at theHolifield Radioactive Ion Beam Facility (HRIBF) atOak Ridge National Laboratory have been migratedfrom 1970s-vintage minicomputers to a modern systembased on Vista and EPICS toolkit software. Stabilityand capabilities of EPICS software have motivatedincreasing use of EPICS for accelerator controls. Inaddition, very inexpensive subsystems based on EPICSand the EPICS portable CA server running on LinuxPCs have been implemented to control an ion sourcetest facility and to control a building-access badgereader system. A new object-oriented, extensibledisplay manager has been developed for EPICS tofacilitate the transition to EPICS and will be used inplace of MEDM. EPICS device support has beendeveloped for CAMAC serial highway controls.

1 PROJECTIn 1992 Oak Ridge National Laboratory beganconstruction of the first radioactive-ion-beam facility inthe United States, an innovative fast-paced, low-costproject utilizing existing accelerators of the shut-downHolifield Heavy Ion Facility[l], The facility housedtwo accelerators: ORIC, a K= 100 cyclotroncommissioned in 1964 as a high-current, light-ionaccelerator, modified in the early 1970s to accelerateheavy ions and modified again in the late 1970s toserve as a booster, and a 25 MV tandem electrostaticaccelerator built in the late- 1970s, which could operatestand-alone or as an injector into the ORIC booster.

Work consisted of modifying ORIC to once again serveas a high-current hydrogen and helium accelerator,stripping out injection components and restoring theORIC internal ion source, constructing a 300 kVplatform to house RIB ion sources, constructing a high-resolution mass separator for isobaric contaminant ionbeams, building a new beam line from ORIC to theRIB platform and building an injection beam line fromthe RIB platform to the tandem electrostatic accelerator.At the same time, beam lines were constructed to twonew experimental devices.

2 CONTROLS DILEMAAll of the new equipment required controls butprovision of controls by extension of either of theexisting accelerator control systems proved to be notfeasible. In fact, both accelerator control systems wereoverdue for upgrade or replacement. ORIC’s controlsystem was a mixture of a subset of the original hard-wired controls and a MODCOMP-minicomputer/CAMAC system installed as part ofORIC ‘s conversion to a booster in the 1970s.Because most of ORIC’S controls were associated withnow-obsolete booster-mode components and becausewe intended to extensively modernize ORIC, ORIC’SCAMAC hardware was stripped out and new controlswere implemented with Allen-Bradley PLC hardware.The tandem accelerator’s control system consisted of aPerkin-Elmer-minicomputer/CAMAC system designedin 1976. The tandem accelerator CAMAC controlhardware would remain but the Perkin-Elmerminicomputer desperately needed replacement. Bothaccelerator systems were programmed largely inassembly language with some FORTRAN and, as aresult, control system extensions were extremely labor-intensive and upgrade of control computers to moremodern computers would have required a largeprogramming effort. Our staff did not include a fhll-time programmer and engineering personnel wereextremely busy with facility modifications.

3 SEARCH FOR A SOFTWARETOOLKIT

We looked outside for a control system softwaresolution and found three possibilities: TACL beingdeveloped for CEBAF, EPICS developed at LosAlamos National Laboratory (LANL) and in use atseveral accelerator facilities and Vsystem, a commercialoffshoot of LANL also in use a several acceleratorfacilities. Vsystem is a product of Vista ControlSystems, Inc. The choice quickly narrowed to EPICSor Vsystem. Both systems are “toolkits” consisting ofa dynamic, distributed database component (ChannelAccess for EPICS and Vaccess for Vista), an operatorinterfhce component (DM or MEDM for EPICS andVdraw for Vista) and other took such as sequencers,

1Research supported by the U.S. Department of Energy under contract DE-AC05-960R22464 with Lockheed MartinEnergy Research Corp.

“The submitted manuscript hss beenauthored by a contractor of the U.S.Government under contract No, DE-

AC05-960R22464. Accordingly, the U.S.Government retains a nonexclusive,royalty-free license to publish or reproducethe published form of this contribution, orallow others to do so, for U.S. Governmentpurposes.”

. .

strip charts, alarm managers, etc. EPICS at the timeoperated in a UNIX development environment, withVMEIVxWorks being required to house the dynamicdatabase. Vista at the time only operated in aVAXiVMS environment with VMEIELN optionallyused for the remote portion of the distributed database.EPICS had little support for CAMAC but extensivesupport for Allen-Bradley which was our preference ibrbeam line and ORIC controls. Vista sothvare includedCAMAC support, important for tandem acceleratorcontrols, and Vista was willing to add Allen-BradleyPLC support.

4 OR.ICAND RIB PLATFORM SYSTEMOPERATIONAL

A major factor in the decision to use Vista sotlware wassuperior documentation and ease of installation and use,important because we did not have a full-timeprogrammer, were extremely busy and had shortdeadlines. Licensing and other costs were not

isignificantly different. Indeed, we had Vsystemrunning, “out of the box” in a few days andsuccessfully controlled Allen-Bradley PLC componentsof the RIB platform and beam lines in short order.Ultimately the system controlled 12,057 Allen-Bradleychannels and 1,244 serial RS-232, RS-422 and RS-i

44 485 channels through VMIC VMIVME-6015 VME

serial controllers,!{iY The Vdraw component of Vsystem worked very wellj and is flexible and easy to use, but we have continued

jto experience stability problems with Vaccess/ELNrelated to memory usage in ELN nodes as the numberof connections to ELN nodes is increased. The problemis managed by administratively limiting the number of

j open connections.

jAllen-Bradley 177 l-series PLC hardware has proved to!be cost-ef%ctive and reliable at the chassis level. Themost troublesome system problems have been in the1

< area of communication to PLC hardware fiorn the! Allen-Bradley VME-based AB6008SV remote I/O

scanner module which Allen-Bradley has thus fm beenunable to completely resolve. Controlnet, ethernet andother communication links to Allen-Bradley PLCS willeventually supplant the AB6008SV remote I/Omodule.

As the time came to upgrade tandem acceleratorcontrols, we were able to add a fill-time softwaredeveloperlsystem manager to our stat% At about thesame time, D?3C discontinued development of ELN,which was unfortunate both because of our investmentand because in some ways ELN with its protected-memory capabilities was superior to VxWorks.Fortunately, Vsystem version 3.1 had, by this time,been ported to Unix/VxWorks. Therefore, tandemcontrols were implemented with a DEC UNIX host anda Heurikon VME/VxWorks processor talking with

$j

existing CAMAC by means of a Kinetic System2917/3922 VME-CAMAC interface and we planned to

\

migrate the VMS/ELN controls for the RIB platformand ORIC to UNKWVxWorks with Vsystem version3.1. Although we achieved our goal of retiring theaging 1970s minicomputers and the tandem acceleratoroperated for over a year with this configuration, therewere several problems. Throughput between DECUNIX and our VxWorks node was extremely poor,necessitating development of special “gateway”soilware to but%r ethernet transactions. In addition,Vista software in this configuration would fi-eeze severaltimes per week and Vista was never able to locate theproblem(s). Resulting tandem accelerator downtime(and nighttime trips by members of the engineering staffto restart Vsystem) forced a change.

5 MIGRATION TO EPICSWe decided to move to EPICS after an analysis of thetwo software architectures[2], and because EPICS willbe used increasingly at ORNL following the decision touse EPICS to control the new Spallation NeutronSource (SNS)[3]. However, HRIBF is now in hlloperation making it a high priority to petiorm thetransition with minimal impact on operations. Inaddition, operations personnel have grown accustomedto control screens based on Vdraw and desired not tohave to find their way around a radically new set ofscreens. Engineers desired some of the nicerdevelopment features of Vdraw, such as symbol supportand flexible widget tools. Additionally, MEDM didnot work well on our DEC UNIX nodes. For thesereasons, a new extensible display manager wasdeveloped which is used in place of EPICS DM orMEDM and which can be used to communicate withboth EPICS and Vsystem channels, allowingincremental transition to EPICS. The new displaymanager is also an experiment in object-orienteddisplay manager design; new widgets can be designedas derived-class objects quickly and easily satisfyingoperator desires for new widgets or special widgetbehavior. New display widgets are implemented in amanner similar to creation of custom EPICS recordsupport. Anyone familiar with the power of EPICSextensible record support would appreciate the sameflexibility in display manager extension. Using thisobject model for display manager design permitsdistribution of updates to the object base class withoutimpact on custom widgets developed by users of thedisplay manager. Users of the display manager are ableto develop custom widgets which may be shared withother sites.

EPICS driver software was developed to communicatewith CAMAC serial highways. Tandem acceleratorcontrols are CAMAC-based with 18 CAMAC crateslocated on six CAMAC serial highways. The serialhighways originate fkom two CAMAC matescontaining six Kinetic Systems 3992 serial highwaydrivers and the VME-based Heurikon VxWorksprocessor communicates with these crates through aKinetic Systems 29 17-VME/3922-CAMAC link.

The EPICS portable channel access (CA) server hasbeen used to implement control and monitoring ofserial devices using inexpensive PC hardware runningthe Linux operating system. As a test of the stabilityof this configuration, a building-access badge readersystem was implemented and has been running hmany months without the need for a single reboot.Running the portable CA server on Linux was valuablefor off-line development of control screens withsimulated control channels and the portable CA serveron Linux was used to implement assignable knobs andassignable analog meters for accelerator control and PChardware running VxWorks The EPICS database willsimilarly be used throughout the facility to avoid thehigh cost of VME hardware. The primary use ofPC/VxWorks will be to interface Allen-Bradley PLChardware by means of ISA-bus-based Data HighwayPlus interfaces, and to control serial devices.

6 REFERENCES[1] Dowling, D.T. et. al. ,“Status of the RadioactiveIon Beam Injector at the Holifield Radioactive IonBeam Facility,”; Particle Accelerator Conf./ Int. Conf,on High-Energy Accelerators, Dallas, May 1-5,

[2] John W. Sinclair, “Software ArchitectureConsiderations for Ion source Control Systems”, 7thht. Conf. on Ion Sources, Taormina, Italy, Sept.7-l 3,1997

[3] W.R. DeVan, et. Al. ,“Distributed ImplementationPlan for a Large, Distributed Accelerator ControlSystem”, Proceedings of the 1997 InternationalConference on Accelerator and Large ExperimentalPhysics Control Systems, Beijing, China, November3-7, 1997

metallic foils running across it. Uneven and insufficient lighting at the paper mill itself as well as the nature of the stroboscopiclight and the way it is applied are of concern. Very dark images with an uneven bright spot on the side are obtained. The second

artifact was caused by water beads flying over the web and generated by the shaking action of the table, therefore introducingbright spots covering large portions of the image. Figure 2 illustrates an image of the web prior to processing.

Figure 2. A raw image of the web

2.1. Image Enhancement

The raw images obtained from the web as described in section 2. merely distinguish some structure or differences in theappearance of the various regions of the slurry surface. A background subtraction algorithm is applied to enhance the appear-ance of the images and make them more amenable to processing and analysis. The background subtraction algorithm will bedescribed in the following.

This process was inspired from the early work of Stanley Sternberg [22]. The essence of this routine is to remove smoothcontinuous backgrounds from the image. The preprocessing consists of a background removal step in which the slowly vary-ing portion of the image is separated and then subtracted from the original image. The resulting image contains the moresharply defined features of interest. A further enhancement step then applies a histogram stretch. The background subtractionis implemented using the rolling ball method [22]. This method, based on gray scale morphology, is the same as erosion anddilation with a spherical structural element. Conceptually, consider the image as a 3-D graph where intensity is the thirddimension. Then, place a ball underneath the surface of this graph and roll it under the entire image. In the case of a uniformimage, the top point of the ball would be tangent to the image everywhere. Taking the locus of points defined by the topmostpoint of the ball as the background, the result is just the image itself. For an arbitrary image, define the topmost point of theball at aparticular(.x, y) point as the background value for that point. The complete set of points forms the background image.

An important parameter influencing feature sizes in the image is the ball radius, providing a filter on the size of the imagearea affecting the background value. After the background image is subtracted, the remaining image generally has lessdynamic range than the original. A histogram stretch is applied in which the tails are truncated to expand the dynamic rangeand enhance the features of interest that remain. Results of the background subtraction are shown on Figure 3.

1

Figure 3. The background subtraction algorithm applied to the image of Figure 2.

2.2. Facet ModeI

The facet model is a powerful tool in image processing. Its uses range from edge detection [23,24,25], background normaliza-

tion [26], shape [27] and surface topography [28], to image segmentation procedures involving detection of corners, curves,valleys, and ridges [29,30]. The facet model principle is based on the minimization of the error between the image thought ofas a piecewise continuous gray level intensity surface and the observed data from the physical scene [31]. The image is consid-ered as a noisy discretized sampling of the surface. The general forms of the facet model include piecewise constant, piecewiselinear, piecewise quadratic, and piecewise cubic. In the constant model, each region in the image has a constant gray level. In

the sloped model, each region has a gray level surface that is a sloped plane[29]. The model used in this work is the cubic poly-nomial defined by Equation (1),

flX,y) = kl + k,x + k,y + kdx’ + k5xy + k6y2 +(1)

k7x3 + k8x2y i- k9xy2 i- k10y3

where f(x, y) is the gray level value at pixel location (x,y) whose neighborhood is to be fitted. A local vector of the 10 coeffi-cients, computed as weighted sums of the values in the local neighborhood, is found for each pixel (x,y). A discrete orthogonalpolynomial basis permits independent estimation of each coefficient as a linear combination of the data values in the neighbor-hood of (x,y). Those polynomials are given by Equation (2) for the 1-D case. The 2-D polynomials are obtained by taking thetensor product of the 2 sets of I-D polynomials.

Let the discrete integer index set R be symmetric in the sense that r E

nomial. The discrete polynomials are iteratively constructed as follows:

Define PO(r) = 1. Suppose PO(r), . . . . P._,(r) have been defined.

P.(r) = rn+an.lrn-] +... +a, r+ao.

R implies -r G R. Let F’n(r) be the nth order poly-

(2)

P.(r) must be orthogonal to each polynomial Po( r) , . . . . P. _I (r) . We then have the set of n linear equations

~ Pk(r)(rn+an_ lrn-l+... +alr+CJo)= O,k = 0,..., I-I (3)r~R

Solving for the set of equations yields the set of discrete orthogonal polynomials

Pi+,(r) = rPi(r)– fliPi_)(r) (4)

where,

pi=xrPi(r)Pi_l(r)

~Pi_,(i-’)reR

Po(r) = 1, Pi(r) = r

The first five polynomials are given as

Po(r) = 1

P~(r) = r

P2(r) = r’–~P’o

()~3=r3_ &.r

W2

(5)

(6)

The facet model consists of solving an equal weighted least square fitting problem by minimizing the error

e’ = xp(’)- i %Pn(r)]2 (7)reR n=o

in terms of the a. coefficients. d(r) is the data value observed (grey level values). The coefficients of the bivariate cubic of

Equation (l), kl, kz, . . . . k. can then be determined. An error image describing the quality of fit is also generated. Given the

10 coefficients ki defining the polynomial at pixel location (x,y), a number of topographic measurements can be determined.

Image intensity surface patches are labeled and grouped according to the categories defined by monotonic, gray level, andinvariant functions of directional derivatives, namely the gradient and the Hessian of the facets given by Equation (8).

1afz andaf~

r .

Ia’f a’f.—ax2axay

a’f a’f——ayaxay2

(8)

— —

The signs of those quantities are used to identify the region’s label, This results in the following categories: (1) Peak, (2)Ridge, (3) Saddle, (4) Flat, (5) Ravine, (6) Pit, (7) concave Hillside, (8) Saddle Hillside, (9) slope Hillside, and (10) convexHillside. The image can then be represented in a rich and hierarchical structure using these topographic units. The topographicstructures properties are defined in [16]. The facet model coefficients were computed for images of the wet end of the paper

web using a window size of 13 x 13. Smaller size windows were also tested and were found not to yield good results. Com-

parative images with window sizes of 5 x 5 and 13 x 13 are shown in Figure 4.

6

..

(a) (b)

Figure 4. Topographic images using (a) 5x5 window and (b) 13x13 window.

It is noticeable from the results of the facet model that a number of elongated structures with distinguishable topographic

characteristics are present. From our observations on more than 500 processed images, the nonunifonnities seem to have cer-tain common topographic characteristics. Values that are of the types: hillside convex, hillside concave, and hillside saddle allyield a good characterization of the structures in question. Based on those observations, a multilevel thresholding was appliedto the topographic images where only those values corresponding to hillsides (concave, convex, saddle) were retained.’

2.3. Mathematical Morphology for Binary Image segmentation

Using the image in Figure 4(b), a binary version is computed by leaving only the hillsides (concave, convex, and saddle). Thatimage still contains some noise and small size components and needs some further cleaning. Morphological processing is usedtowards the filtering of small size features and the segmentation of the image. A closing with a structuring element in the hori-zontal direction followed by an opening with the same structuring element are applied. The results of this operation are shownin Figure 5.

Figure 5. Filtering of elongated structures using morphological operators.

A closing followed by an erosion is subtracted from the dilated version of the same image to yield a set of boundaries of

the various structures in the image, see Figure 6. The objects resulting from the morphological processing were then filteredaccording to their size, orientation, and elongation. Figure 7 shows the result of that operation and represents the characteriza-tion of the surface of the slurry using minimum bounding rectangles (MBR) to identify the nonuniformities.

7

..

Figure 6. Binary and boundary images of the slurry.

(a) (b)Figure 7. (a) Original image of the slurry, (b) results of the surface characterization algorithm(MBR),

3. REAL-TIME IMPLEMENTATION OF THE FACET MODEL

A real-time facet model implementation is being developed for computing the sufiace feature parameters at 30 frames per sec-ond. This implementation is on a Detachable Maxpci containing dedicated pipeline processing hardware for image processing

applications. An initial design has been completed in which the necessary computations have been mapped to the hardware.Figures 8 and 9 show a block diagram of the design. The cubic polynomial approximation consists of ten terms with coeffi-cients kl through klo Each coefficient is calculated by convolving the input image with a mask. The mask is predetermined to

compute the least square fit of the Chebyshev polynomials to the input image. On the Datacube board, dedicated convolutionhardware is used for this computation. These computations are performed at 40 Mhz., so that a 512 x 512 image will require 7ms. for each coefficient calculation. With a 100-point convolver, two convolutions can be performed at the same time with upto a 7 x 7 convolution mask. The total time for the coefficient calculations is approximately 35 ms. After convolution, the coef-ficients are stored in memory and are used in the next step of the parameter calculations. Using the arithmetic blocks and look-

up tables, the gradient, the gradient magnitude, the eigenvalues and eigenvectors of the Hessian are calculated. From thesequantities the various feature values can be determined.

cam

I Convolution Masks

Figure 8. Block diagram of the real-time implementation of the facet model

4. CONCLUSION AND FUTURE WORK

In this paper, we propose a new application for the facet model algorithm. We show that the surface of a paper web at the wetend can be characterized using image enhancement algorithms (the background subtraction) followed by topographic descrip-tions via the facet model. Mathematical morphology is then used for the final segmentation of the results. Geometric filteringyields well-segmented images with clearly defined nonuniformities. Measurements of location, size, and orientation of thestructures are also computed. The initial work on the real-time implementation of the facet model is addressed. The algorithmimplemented yields a very good rate of detection of the nonuniformities on the surface of the web. Future work will involve the

use of a laser-based structured light profiler in conjunction [32] with the CCD camera in order to study the third dimension ofthe web not represented by the 2-D images.

ReferenCeS

[1] K. Humphrey, Image Analysis, Pira International, 1995.

[2] S. I. Shapiro and R. H. Shearin, “On-Line Sensors: Exploring the Trends;’ Tappi .hmrml, 78, No. 9, pp. 83-84, September1995.[3] J.-P. Bernie and W. J. M. Douglas, “Local Grammage Distribution and Formation of Paper by Light Transmission Image

Analysis;’ Tappi Journal, 79, No. 1, pp. 193-202, January 1996.[4] A. J. Niemi, and C. J. Backstrom, “Automatic Observation of Dry Line on Wire for Wet End Control of the PaperMachine;’ Pulp and Paper Canada, 95, no. 2, pp. 27-30, February 1994.

[5] M. Whitaker, “Optimizing Formation Through Consistency Measurement - a Wet End Revolution;’ Paper Technology, 36,

no. 2, pp. 22-26, March 1995.

[6] A. Kiviranta, “The Role of Table Activity in Fourdrinier Forming,” Canadian Pulp and Paper Association 79th AnnualMeeting Technical Section, Montreal, Canada, January 1993.[7] C. K. Aidun, “Hydrodynamics of Streaks on the Forming Table;’ Tappi Journal, August 1997.[8] T. Nomura, K. Wada, and T. Shimizu, “High Consistency Sheet Forming;’ Tappi Journal, pp. 115-122, January 1989.

9

,

,

J,

. .

1

I

J

,

Figure 9. Block diagram of the real-time implementation of the facet model (continued)

[9] H. Dahl, H. Holik, and E. Weisshuhn, “The Influence of Headbox Flow Conditions on Paper Properties and Their Con-stancy,” Tappi Journal, pp. 93-98, February 1988.[10] C. K. Aidun and A. E. Kovacs, “Hydrodynamics of the Forming Section: The Origin of Nonuniform Fiber Orientation;’Tappi Journal, 78, No. 11, November 1995.[11] S. I’Anson, “3-D Laser Surface Profiling of Paper and Board;’ Paper Technology, 38, no. 2, pp. 43-49, March 1997.[12] T. A. Vuori and C. L. Smith, “Three-Dimensional Imaging System with Structured Lighting and Practical Constraints:’Journal of Electronic Imaging, 6, No. 1, pp. 140-144, January 1997.[13] H. Kaplan, “Structured Light Finds a Home in Machine Visionj’ Photonics Spectra, January 1994.

[14] D. J. Purll, “Solid State Image Sensors in Automated Visual Inspection”, B. G. Batchelor, D. A. Hill and D. C. Hodgson,

Eds., Elsevier Science Publishing Company, Inc., New York, NY, pp. 255-293, 1985.[15] D. Zhu and A. A. Louis Beex, “Robust Spatial Autoregressive Modeling for Hardwood Texture Analysis”, Journal of

Visual Communication and Image Representation, 36, no. 9,, September 1993, Academic Press.[16] P. Araman, D. Schmoldt, T. H. Cho, D. Zhu, R. Conners, and D. Earl Kline;’ Machine Vision Systems for ProcessingHardwood Lumber and Logs”, AZ Applications, 6, no. 2, April 1992, pp. 13-26.[17] C. T. Ng, R. W. Conners, and P. A. Araman, “An Experimental Machine Vision Prototype for the Inspection of PlanarWebs”, Proc. of the twenty fourth Southeastern Symposium on System Theory, March 1-31992, pp. 483-487.

[18] R. W. Conners, C. T. Ng. T. H. Cho, and C. W. McMillin, “Computer Vision System for Locating and Identifying Defectsin Hardwood Lumber”, Proc. of the SPIE, vol. 1095, 1989, pp. 48-63.

[19] C. W. Klm and A. J. Koivo, “Hierarchical Classification of Surfaced Defects on Dusty Wooden Boards”, Proc. of 10th.Int’1. Conf. on Pattern Recognition, 1990, pp. 775-779.

[20] J. E. Graf, S. T. Enright, and S. I. Shapiro;’ Automated Web Inspection Ensures Highest Quality Nonwovens”, Tappi

Journal, 78, no. 9, September 1995, pp. 135-138.[21] J. P. Bernie and W. J. Murray Douglas, “Local Grammage Distribution and Formation of Paper by Light Transmission

Image Analysis”, Tappi Journal, 79, no. 1, January 1996, pp. 193-202.[22] S. R. Sternberg, “Biological Image Processing”, Computer, January 1983, pp. 22-34.

[23] A. Rangarajan, R. Chellappa, and Y. T. Zhou, A model-based approach for filtering and edge detection in noisy images,IEEE Trans. on Circuits and Systems, 37, pp. 140-144, 1990.

10

,Ci.!

[24] E. P. Lyvers, O. R. Michell, M. L. Akey, and A. P. Reeves, “Subpixel Measurement Using Moment Based Edge Opera-

tor”, IEEE Trans. PAMI, 12, pp. 1293-1309, 1989.[25] H. Jiang and Y. Cai, “Edge Detection from Surface Fitting by Orthogonal Polynomial”, Acts Autornatica Sinica, 16, pp.202-209, 1990.[26] M. Jankowski, “An Iterated Facet Model Approach to Background Normalization”, Proc. of the SPIE, 2238, pp. 198-206,

1994.[27] C. P. Ting, R. M. Haralick, and L. G. Shapiro, “Shape from Shading Using Facet Model”, Pattern Recognition, 22, pp.

683-695, 1989.[28] Q. Zheng and R. Chellappa, “Estimation of Surface Topography from Stereo SAR Images”, IEEE ICASSP-89, Glasgow,

UK, 3, pp. 1614-1617, 1989.[29] R. M. Haralick and L. G. Shapiro:’ Computer and Robot Vision”, Addison Wesley Publishing Company, Inc., 1st cd.,

1992.

[30] P. I. Radeva, “A Rule-based Approach to Hand X-ray Image Segmentation”, Computer Analysis of Images and Patterns,

5th Int’1. Conf., CIAP 93 Proc., Budapest, Hungary, pp. 641-648, 1993.[31] S. D. Pathak, Y. Kim, and J. Kim, “Efficient Implementation of Facet Models on a Multimedia System”, Optical Engi-

neering, 35, pp. 1739-1745, 1996.[32] H. Sari-Samaf, J. S. Goddard, Jr., J. C. Turner, M. A. Hunt, and B. R. Abidi, “On-Line Characterization of Slurry for Mon-itoring Headbox Performance”, Tappi Proc., Milwaukee, WI, 1998, pp. 97-115.