membership/ei/pdfs/ei15-2.pdf june 2005 vol. 15, …gsharma/papers/imagingarithei...to download the...

1SPIE’s International Technical Group Newsletter

ELECTRONIC IMAGING 15.2 JUNE 2005

ELECTRONICIMAGING

JUNE 2005VOL. 15, NO. 2

Newsletter of the SPIE ElectronicImaging Technical Group

Imaging arithmetic: Physics ∪ Math >Physics + MathGaurav Sharma, Electrical and Computer Engineering and Biostatistics & ComputationalBiology Depts., University of Rochester

For several real-world problems, the mathematical meth-ods of signal and image processing are most successfulwhen they also incorporate the insight offered by thephysics of the problem. Imaging systems are a particu-larly fertile ground for problems in this class becausethey deal specifically with the capture of physical scenesand with the reproduction of images on physical de-vices. Solutions for these problems that combine physi-cal understanding and modeling with appropriate math-ematical tools of signal/image processing, offer advan-tages significantly greater than the sum of the individualparts. In this article, we illustrate how the happy mar-riage between physical insight and modeling and ap-

To download the electronic version of this newsletter, please go to:http://spie.org/membership/ei/pdfs/ei15-2.pdf

Gabriel Marcu,Gabriel Marcu,Gabriel Marcu,Gabriel Marcu,Gabriel Marcu, Apple Computer, Inc.

Editor/Technical Group Chair

propriate mathematical methods leads to novelsolutions. Additional examples and referencesmay be found in the identically-titled paperpresented at the January 2005 IS&T/SPIE Elec-tronic Imaging meeting.1

Show-through in document imagingDuplex or double-sided printing is commonlyused for hardcopy documents, with most maga-zine and book pages being prime examples.When a duplex-printed page is scanned, infor-

Figure 1. Scans of two sides of a duplex-printed page.3

Continued on page 10.

Figure 2. Scanned data after show-through correction.3

ContentsA simple optical system for real-time size measurements of nuclearfuel pellets, Thomas P. Karnowski,Andrew K. Kercher, John D. Hunn, andL. Curt Maxey, Oak Ridge NationalLaboratory ........................................ 2

Machine vision for defect detec-tion on silicon wafers, FabriceMeriaudeau, Patrick Gorria, Le2i Labo-ratory, Univ. of Burgundy; PierrickBourgeat, BioMedIA Laboratory,CSIRO; Kenneth Tobin, Oak Ridge Na-tional Laboratory ............................. 3

Three-dimensional inspection ofhighly-reflective metallic objectsby polarization imaging, OlivierMorel, Christophe Stolz, FabriceMeriaudeau, and Patrick Gorria,Laboratory Le2i ................................ 4

Making 3D binary digital imageswell composed, Marcelo Siqueira,Jean Gallier, Department of Computerand Information Science, Univ. of Penn-sylvania; Longin Jan Latecki, Depart-ment of Computer and Information Sci-ences, Temple Univ. .......................... 5

Information processing of motionin facial expression and the geom-etry of dynamical systems, Amir H.Assadi, Brenton W. McMenamin, De-partment of Mathematics, Univ. of Wis-consin; Hamid R. Eghbalnia, Depart-ment of Mathematics, Department ofBiochemistry, Univ. of Wisconsin .... 6

Digital bullet scratches, Jan Lukas,Jessica Fridrich, and Miroslav Goljan,Department of Electrical and ComputerEngineering, SUNY Binghamton ..... 7

Motion-based particle filtering forhead tracking applications, NidhalBouaynaya and Dan Schonfeld, Multi-media Communications Laboratory,Univ. of Illinois ................................. 8

Managing levels of detail for real-time rendering in immersive vir-tual environments, D. Paillot, F.Mérienne, M. Neveu, and S. Thivent,Le2i laboratory, Institut Image ....... 9

Deflectometric inspection of dif-fuse surfaces in the far-infraredspectrum, Jan W. Horbach and SoerenKammel, Institute for Measure-ment and Control Systems, Univ. ofKarlsruhe ........................................ 12

SPIE’s International Technical Group Newsletter2


A simple optical system for real-time sizemeasurements of nuclear fuel pelletsThomas P. Karnowski, Andrew K. Kercher, John D. Hunn, and L. Curt Maxey, Oak Ridge National Laboratory

Advanced nuclear reactor designs usefuel forms that are built up from tensof thousands of tiny nuclear fuel pel-lets. Called TRISO (tri-isotropic) par-ticles, these employ a dense layer ofsilicon carbide to trap radioactive fis-sion products. The coated TRISO fuelacts as a containment system to pre-vent the release of fission products tothe environment during accidents.Furthermore, these advanced nuclearreactor designs are regarded as an im-portant part of the future hydrogeneconomy because they can be used toproduce hydrogen more efficientlythan electrolysis.1 Rapid counting andmeasurement of TRISO spheres is anecessary technology for the devel-opment of these materials. A surveyof commercial equipment foundavailable devices lacking in speed,resolution, and accuracy. We there-fore elected to develop this capabil-ity in-house.

Based on the methods used byWallisch and Koss2 we employed alight obstruction concept where a slitof light (either an aperture or a focused beam)is blocked by a particle. The system projectslight through a target transport cell and collectsthe light onto a photo receiver (see Figure 1).The signal from this receiver is digitized by ahigh-speed analog-to-digital conversion unit(ADC), and is shown in Figure 2. When no lightis blocked, the signal is high, but it then shrinksas the sphere passes through the slit and reachesa minimum when the sphere blocks the maxi-mum amount of light. The maximum lightblockage is proportional to the diameter of thesphere.

The error in our measurement can be lim-ited by the ADC resolution since we are tryingto estimate the radius with the intensity of theobstruction. We attempt to better estimate thepeak location by fitting the data to a parabola.In our software this is accomplished by find-ing the sampled peak, then locating 80 valueson either side of it. These 161 values are usedto find the coefficients of a least-squares errorfit to a parabola, and to estimate the ‘true’ mini-mum value of the curve. This improves the ac-curacy of the ADC resolution by a factor of100.

Another problem was linked to the size ofthe particle-handling system, which was de-

Figure 3. An abnormal event where multiple particles passthrough the laser line. This shape deviates considerably from aparabola and can be detected as an abnormal event for latermanual inspection.

Figure 2. A typical event where a particle passes through thelaser line and is detected. The shape is similar to a parabola.

Figure 1. Schematic of the optical system.Particles are carried by vacuum pump through atarget cell where they are illuminated by a laserline. The resultant blocked signal is detected andprocessed by the computer.

signed for 1000µm spheres. Whensmaller spheres (800µm) passthrough the cell, some overlappingcan occur, causing a signal that de-viates significantly from the idealevent (see Figure 3). We detect thesein real-time by determining the errorbetween the event and a three-pointparabolic fit. Large errors indicateseveral particles may have been de-tected together. This provides aneasy-to-compute method for findingmulti-particle events that can bemanually screened with softwaretools.

Experiments have estimated the countingaccuracy as having an error of less than 0.075%with a 95% confidence. The size-measurementaccuracy was on the order of 11µm standarddeviation for spheres 1000µm in diameter. Al-though the current particle transport systemdoes not support the maximum detection rate,electronically-generated data showed rates of200 particles per second, implying a through-put of 720,000 particles per hour. These ratesand accuracies will improve the research anddevelopment cycle for the manufacturing ofthese pellets: which will ultimately lead to safer,

more-efficient nuclear reactors producing en-ergy to benefit us all.

Thomas P. Karnowski,Andrew K. Kercher, John D. Hunn,and L. Curt MaxeyOak Ridge National Laboratory, TN

References1. ORNL Review 35 (2), 2002.2. K. Wallisch and P. Koss, Automatic size analysis of

coated fuel particles, Nucl. Tech. 35, pp. 279-283,1977.



Machine vision for defect detection on silicon wafersFabrice Meriaudeau, Patrick Gorria, Le2i Laboratory, University of Burgundy; Pierrick Bourgeat, BioMedIA Laboratory,CSIRO; Kenneth Tobin, Oak Ridge National Laboratory

Silicon wafers are extensivelyused in the semiconductor andmicroelectronics industries.With this material, it is of ex-treme importance to obtain adefect-free surface to improveyield and microchip perfor-mance. Current practice in theindustry is to inspect the wa-fers for surface defects onlyat the end of the final polish-ing stage. Techniques such asatomic force microscopy,scanning tunneling force mi-croscopy, scanning electronmicroscopy, x-rays andacoustic electron microscopyhave all been used to performthe surface-defect character-ization. However, all thesemethods require cumbersomeequipment that is expensive touse on a daily basis in a pro-duction factory.

We have developed an al-ternative technique that relieson the use on the direct-to-digital holography(DDH).1 This allows us to capture the compleximage of a scene, and then reconstruct the phaseand amplitude information. This gives us acomplete 3D representation of the scene withinthe wavelength depth of the laser used to pro-duce the hologram: thus it provides a new di-mension to texture analysis.2,3

In die-to-die wafer inspection (see Figure1), defect detection is based on the comparisonof the same area on two neighboring dies. Im-ages being thoroughly aligned, the dissimilari-ties between the images are a result of defectsin the area of one of the dies. The two imagesare subtracted, and a threshold level is selectedto locate any anomaly.

To optimize the signal-to-noise ratio, thethreshold value is established based upon thenoise level in the difference image. However,since multiple structures coexist in the samefield of view, the noise level can vary over asingle image preventing the use of a globalthreshold. To overcome this problem, a seg-mentation is done to create a mask of the dif-ferent regions (each region being a class forour classifier) in order to produce a measure ofnoise for each structure in the difference im-age. This leads to an individual threshold foreach region (see Figure 1).

The mask is created through a ‘classifica-tion-segmentation’ procedure. The complex

Figure 1. Principle of die-to-die wafer inspection and defect detection with the creation ofa mask for local noise and threshold estimation. Respectively, µ and σ are the averagegrey and the standard deviation in the different regions.

Figure 2. Bank of Gabor filters with three scalesand two orientations. Figure 3. Example of a segmented image.


image provided by the DDH is first processedto generate four new images: the phase, nor-malized-amplitude, and complex images; andalso the complex image with normalized am-plitude.5 Each image is then processed using abank of Gabor filters (see Figure 2), and thenused as input in a vote SVM (support vectormachine) classifier.6

A pixel is classified as pertaining to a givenclass (dynamic RAM, logic area, or blank area)if it is identified as such by at least two of thethree best classifiers. A so-far unclassified pixel—a pixel identified as being part of a differentclass by each of these three classifiers—is iden-

tified according to the resultsof the fourth. Figure 3 showsan example of a segmentedimage. The obtained resultsshow robust segmentation.4,6

The technique we devel-oped enables us to train theclassifier with a small set ofexamples and to obtain resultssimilar to those produced us-ing a full training set. The tech-nique can be applied to anytype of optical tool for waferinspection, but in the particu-lar case of the DDH, the extrainformation provided by thecomplex nature of the image,makes it even more powerful.The classification has been ap-plied to defect detection andhas led to a tremendous im-provement in the final results:4

the ratio of false detection tono detection.

Fabrice Meriaudeau, Pierrick Bourgeat*,Patrick Gorria, and Kenneth Tobin†

Le2i LaboratoryUniversity of Burgundy, France*BioMedIA LaboratoryCSIRO, Australia†Oak Ridge National Laboratory, TNE-mail: [email protected]



Three-dimensional inspection of highly-reflectivemetallic objects by polarization imagingOlivier Morel, Christophe Stolz, Fabrice Meriaudeau, and Patrick Gorria, Laboratory Le2i

In the field of industrial vi-sion, 3D (three-dimen-sional) inspection ofhighly-reflective metallicobjects is still a difficulttask. Indeed, highlights pre-vent the use of a 3D laserscanner, and phase-shift-ing-based 3D systems arebetter adapted to the inspec-tion of surfaces with lowcurvature. The ‘shape frompolarization’ technique1,2

can be extended to provide3D information on highly-reflective metallic surfaces.After being reflected from an object,unpolarized light becomes partially-linearly polarized: to what extent andin what direction depends on the sur-face normal and the refractive indexof the surface medium.3 Once the sur-face normals are computed, the sur-face profile can be obtained by inte-gration.

Partially-linearly-polarized lighthas three parameters: the intensity,and the angle and degree of polariza-tion. The latter gives the proportionof the linearly-polarized component.For characterization, we can use a ro-tating polarizing filter in front of thecamera. The relationship between theintensity and the rotation angle of the filter isgiven by a sinusoid. The phase shift gives theangle of polarization, and the degree of polar-ization is the relative magnitude of the sinu-soid. To speed up the sensing of polarizationcomponents, the filter is replaced with an elec-trically-controlled liquid-crystal polarizationrotator.

The polarization parameters of the reflectedlight are linked to the normals of the surface.On the one hand, the linearly-polarized com-ponent is orthogonal to the incidence plane,meaning that the angle of polarization gives theazimuth angle. On the other hand, by using thecomplex refractive index of the surface, andthanks to the Fresnel coefficients, a new rela-tion between the degree of polarization and thezenith angle can be found.4 Therefore, these twoangles determine the surface normals (see Fig-ure 1). The surface is then reconstructed fromthe normals by integration.

The acquisition system consists of a CCDcamera, the liquid-crystal polarization rotator,

and a diffuse dome light (see Figure 2) to sup-ply unpolarized light across the whole surface.After being reflected, the light—which be-comes partially-linearly polarized—is ana-lyzed. Since the degree of polarization is lowerfor metallic surfaces, a sensitive camera with10bit depth is used. In total, 18 frames are ac-quired with different polarization rotations from0° to 180°, with a constant step of 10°. The

l e a s t - m e a n - s q u a r emethod is applied duringthe acquisition process tocompute the polarizationparameters. Then, byknowing the object’scomplex index of refrac-tion, the normals are com-puted and finally the sur-face is reconstructed byintegration.

ResultsBecause of the manufac-turing process, metallicobjects made throughstamping and polishing

do not all come out the same way: thus3D information has to be extracted inorder to characterize them.

To compare the 3D surface ob-tained with our polarization system toa reference 3D surface, an object wascoated with a thin opaque layer andscanned with a 3D commercial scan-ner (Replica 500). After registrationof the two surfaces, the mean devia-tion between the surfaces was about30µm. The reconstructed surface ispresented on Figure 3. The resolutionalong the X and Y axes depend onlyon the lenses used and on the spatialresolution of the sensor: so X/Y reso-

lutions of up to three times this accuracy shouldbe achievable.

Figure 4 shows the comparison of a recon-structed surface from a flawless object and adefective part which has a shape defect (bump)on the left-bottom .

ConclusionPolarization imaging presents a new way ofreconstructing the 3D surface of highly-reflec-tive metallic objects. The acquisition equip-ment—diffuse dome light, CCD camera, andliquid crystal polarization rotator—are quitesimple. The system is fully automated, fromthe acquisition of the polarization parametersto the 3D surface reconstruction process. Fur-ther, it can be integrated with production linesto detect shape defects on objects made viastamping and polishing.

Figure 1. Reflection of a light wave from a highly-reflective surface.

Figure 2. System to acquire the polarization of thereflected beam.

Figure 3. Polarization imaging applied to the 3D reconstruction of highly-reflective metallic objects: (a) object of reference photograph, (b) 3Dreconstructed surface.

Continued on page xx.

Figure 4. Mean deviation map between a flawlesssurface and a surface with a shape defect.



Making 3D binary digital images well composedMarcelo Siqueira, Jean Gallier, Department of Computer and Information Science, University of Pennsylvania; Longin JanLatecki, Department of Computer and Information Sciences, Temple University

A three-dimensional binary digitalimage is said to be well composed if,and only if, the set of points in thevoxel faces consisting of every voxelface that is shared by a foregroundand a background voxel of the imageis a 2D manifold.1 A well-composedimage enjoys very useful topologicaland geometric properties. These prop-erties make simpler several basic al-gorithms in computer vision, com-puter graphics, and image processing.For instance, thinning algorithms donot suffer from the irreducible thick-ness problem if the image is wellcomposed.2 Also, algorithms that relyon curvature computation to extractapproximating iso-surfaces directlyfrom binary images can be applied to well-com-posed images with no need to handle specialcases resulting from ‘non-manifold’ topol-ogy.3,4

On the other hand, if a 3D digital binaryimage is the result of the digitization of a ‘solid’object, such as a bone, and it lacks the prop-erty of being well composed. In this case, thedigitization process that gave rise to it is nottopology-preserving. As the results in Refer-ence 5 show, if the resolution of the digitiza-tion process is fine enough to ensure preserva-tion of topology, then the resulting image iswell composed. This fact has motivated us todevelop an iterative and randomized algorithmfor ‘repairing’ non well-composed 3D digitalbinary images.

Our algorithm restores the given image byconverting background voxels into foregroundones. Although this algorithm always producesa well-composed image, it cannot guaranteethat the result is the same as would be obtainedfrom a topology-preserving digitization pro-cess. This is because our algorithm does notassume any knowledge about the original digi-tization process. Even so, if the number of back-ground voxels converted into foreground onesis not too large, the input and output imageswill be similar, which is satisfactory for sev-eral applications that benefit from using well-composed images.

The conversion process relies on the factthat well composedness is a local property: thatis, the condition of being well composed isequivalent to the nonexistence of two types oflocal critical configurations of image voxels1

(see Figure 1). The first of these is where fourvoxels share an edge, two of them are back-ground voxels, two of them are foreground

voxels, and the background (or foreground)voxels share an edge but not a face. The sec-ond is where eight voxels share a vertex, twoof them are background (or foreground), six ofthem are foreground (or background) voxels,and the two background (or foreground) voxelsshare a vertex but not an edge. It can be shownthat a 3D binary digital image is well composedif, and only if, it does not contain any instancesof these critical configurations.

Note that we can decide if a given 3D bi-nary digital image is well composed by simplyverifying if any 2×2×1 neighborhood of voxelsof the image is an instance of the first criticalconfiguration, and if any 2×2×2 neighborhoodof voxels of the image is an instance of the sec-ond. This test can be performed in linear timewith the number of voxels of the image. Thefirst step of our algorithm is to verify if the in-put 3D binary digital image is well composed.If so, the algorithm finds a subset P of back-ground voxels such that, if they were convertedinto foreground voxels, then the resulting im-age would be well composed. Then this con-version is performed.

Ideally, P should be as small as possible, sothat the input and output images are similar.Although such a smallest set can be found us-ing an exponential-time search, this is com-pletely unfeasible in the context of practicalapplications. Our repairing algorithm is notguaranteed to find the smallest set P, but itstime complexity is linear in the number ofvoxels of the input image. Our algorithm buildsP iteratively, starting with an empty set. Eachiteration inserts a background voxel into P.Each such voxel is randomly chosen from thosein the background in the critical configurationsof the input image, and when such a voxel is

converted into a foreground one, atleast one instance of a critical configu-ration is eliminated. This conversioncan also give rise to a new critical con-figuration, which is further eliminatedby choosing another backgroundpoint. However, this process is guar-anteed to converge to a correct solu-tion after a finite number of iterations.

We tested our algorithm againstseveral magnetic resonance (MR) im-ages of parts of the human body, suchas the brain, torso, and lungs. In allcases, our algorithm generated a well-composed image by converting fewerthan 0.0034*N, where N is the totalnumber of voxels in the input 3D bi-nary digital image. We also derived

an upper bound for the average number ofvoxels that needed to be converted. This upperbound supports the results obtained from ourexperiments with real imaging data, and pro-vides a theoretical measure of the effectivenessof our algorithm for making 3D binary digitalimages well-composed. For future work, weintend to study the existence of a linear-timealgorithm for computing the smallest subset ofbackground voxels that generates a well com-posed image after being converted into fore-ground voxels.

Marcelo Siqueira, Longin Jan Latecki,*and Jean GallierDepartment of Computer and InformationScienceUniversity of Pennsylvania, Philadelphia, PA*Department of Computer and InformationSciencesTemple University, Philadelphia, PAE-mail: [email protected]

References1. L. Latecki, 3d well-composed pictures, Graphical

Models and Image Processing 59 (3), pp. 164-172, 1997.

2. L. Latecki, U. Eckhardt, and A. Rosenfeld, Well-composed sets, Computer Vision and ImageUnderstanding 61 (10), pp. 70-83, 1995.

3. H. Dellingete, Initialization of deformable modelsfrom 3d data, Proc. 6th Int’l Conf. in ComputerVision, pp. 311-316, Bombay, India, 4-7 January1998.

4. N. Krahnstoever and C. Lorenz, ComputingCurvature-Adaptive Surface Triangulations ofThree-Dimensional Image Data, The VisualComputer 20, pp. 17-36, 2004.

5. A. Gross and L. Latecki, Digitizations preservingtopological and differential geometric properties,Computer Vision and Image Understanding 62(3), pp. 370-381, 1995.

Figure 1. Instances of the two critical configurations. For the sake ofclarity, only the background voxels are shown. The shared edge (left)of the first critical configuration and the shared vertex (right) of thesecond are heavily drawn.



Information processing of motion in facial expressionand the geometry of dynamical systemsAmir H. Assadi, Brenton W. McMenamin, Department of Mathematics, University of Wisconsin; Hamid R. Eghbalnia, Depart-ment of Mathematics, Department of Biochemistry, University of Wisconsin

The advent of wireless communication hasopened up a new era of multidisciplinary hu-man and computer vision. Video conferencingnow boasts transmission of realistic facial mo-tion, albeit at a cost in terms of bandwidth andimperfect resolution. However, communicat-ing the expression of a speaker’s facial motionand the listeners’ emotional response is a farmore complex problem. The subtlety rests inthe easy-to-experience fact that a drop in time-resolution causes the movement of facial fea-tures to be perceived as jerky, and that the view-ers often cannot ignore a drop in spatial reso-lution. Time-variation of image frames storedin a video file can be regarded as a dynamicalsystem realized through vector representationof images in sequences of frames in an Euclid-ean space of dimension D. How can we refinethis naive representation to reduce the stagger-ingly high computational resources required toa practical range while retaining low error ratesand other practical advantages?

Latecki et al.1 have treated some of the moresubtle points of digital geometry. The issue ofwhen a discrete object obtained via the sam-pling of visual motion—as for the above-men-tioned example of video—could be treated asa bona-fide geometric analogue of a curve.Further, the type of invariants associated withit are within the realm of digital geometry. Analternative, and still general, point of view isto regard a curve of this type as the flow line(an integral curve) of a dynamical system in ahigh-dimensional space. This works providedwe resolve the delicate issues of reconcilingperceptual and differential geometry (e.g. asdiscussed in Assadi et al.,2 Eghbalnia et al.3

and Townsend et al..4)The exact encoding of realistic facial motions

that convey emotions for a particular person Pforms a parameter space X

P that represents the

objective geometry of the problem, the geomet-ric features, and a representation of the spaceindependent of evaluations by observers. Thesubjective geometry of X

P varies between ob-

servers, reflecting individual experiences andvariation in the perception of facial expressions.The encoding, compression, and communicationof the subjective geometry of X

P is the ultimate

goal of an effective algorithm for transmittingthis genre of multimedia.

The first problem is to find a way to encodethe finer details of X

P in real-time, and to se-

lect a sparse representation adaptively by theobserver’s feedback to the system. Psycho-

physical experiments are required to weigh thevarious features in X

P according to the subjec-

tive valence of emotion for each individual: thedesign of such experiments poses an interest-ing problem in its own right. We recognize,then, that certain geometric invariants of digi-tal geometry of the original motion data couldbe measured and singled out statistically. Eachvideo frame is first encoded by an orthogonaldecomposition (see Eghbalnia et al.5), thus en-coding the motion in terms of significant coef-ficients of one such multi-scale decomposition.

Consider a discrete time-window W oflength w sliding in the interval {a=t

0,b=t

q-1;

t=t0,t

1,…tq-1}, and the principal component

analysis (PCA) of the data set k(W) consist-

ing of the frames of Xk(t) indexed by w. The

window W provides two parameters τ (the start-ing point in time) and w, thus it also suppliesthe principal values and principal directions of

k(W) and other functions of the pair (τ , w).

As in our earlier results in dynamic PCA formulti-channel time series13, each scalar func-tion of the principal values provide a functionF(τ , w). Also, a vector whose components areF

1(τ , w), F

2(τ , w), …, F

N(τ , w) provides an

information surface in RN. Every orthonormalframe of three top principal vectors provides asurface in the Stieffel manifold and the corre-sponding Grassmannian whose statistics areprovided by a relatively small number of al-most-constant clusters.

When the multi-channel time-series associ-ated with

k(W

1),

k(W

2),…

k(W

r),… convey

the same information-theoretic PCA output, theirimages under the mapping above are either thesame or very close together. The images are ei-ther significantly far from each other, or couldbe assigned to distinct clusters in terms of thestatistical analysis (such as the principal framesabove.) Each cluster encodes features that carryessentially the same Shannon information relat-ing to the multi-channel time series and

k(W),

thus providing a class of algorithms for comput-ing the sparse codes derived from the geometryof motion.

The statistical geometry of invariants of theparameters X

P for a sample of population could

provide effective algorithms for extraction ofperceptually significant features in motion (e.g.in unsupervised feature detection from facialmotion.) We have discussed problems of find-ing a sparse representation of motion featuresin circumstances such as normal conversationswithout discernible emotional expression or

with perceptually-distinguishable emotionalexpression. We have also outlined a frameworkfor study of their information-theoretic invari-ants. The algorithm provides a general approachto encode motion in terms of a particular genreof dynamical systems and the geometry of theirflow (integral curves). In particular, a sparsecoding of the video sequences is proposed thatpotentially classifies the features based on in-formation-theoretic arguments. Examination ofthese sparse codes determines which ones re-veal perceptually-significant events in motion,in distinction to those that fall into the categoryof frequent and common facial motions in anordinary serious conversation.

Amir H. Assadi*, Hamid R. Eghbalnia*†,

and Brenton W. McMenamin**Department of Mathematics†Department of BiochemistryUniversity of Wisconsin, Madison, WIE-mail: [email protected]

References1. L. J. Latecki, et al, IEEE Advanced Video and Signal

Based Surveillance, 2003.2. A. Assadi, S. Palmer, and G. Eghbalnia, Proc. of IEEE

NNSP ’99, Ed. T. Adali et al, 1999.3. H. Eghbalnia, A. H. Assadi, and J. Townsend, Invited Chap-

ter in Handbook of Geometric Methods in Vision, Ed.E. Bayro, Springer-Verlag, 2005.

4. J. Townsend, A. Assadi, and J. Busemeyer, Invited Chap-ter in Measurement And Representation Of Sensations,Eds. H. Colonius et al, Erbaum, Berlin, 2005.

5. H. Eghbalnia and A. H. Assadi, J. Neurocomputing 40,pp. 1155-1163, 2001.

Three-dimensional inspectionContinued from page 4.

Olivier Morel, Christophe Stolz,Fabrice Meriaudeau, and Patrick GorriaLaboratory Le2iLe Creusot, FranceE-mail: [email protected]

References1. S. Rahmann, Reconstruction of quadrics from two polar-

ization views, Proc. IbPRIA03 Iberian Conf. on Pat-tern Recognition and Image Analysis, pp. 810-820, 2003.

2. D. Miyazaki, M. Kagesawa, and K. Ikeuchi, Transparentsurface modeling from a pair of polarization images, IEEETrans. Pattern Anal. Machine Intell. 26 (1), pp. 73-82,2004.

3. L.B. Wolff, and T.E. Boult, Constraining object featuresusing a polarization reflectance model, IEEE Trans. Pat-tern Anal. Machine Intell. 13 (7), pp. 635-657, 1991.

4. O. Morel, C. Stolz, and P. Gorria, Application of Polari-metric Imaging to 3D Inspection of Highly Reflective Me-tallic Surfaces, Proc. SPIE 5606, pp. 82-89, 2004.



Digital bullet scratchesJan Lukas, Jessica Fridrich, and Miroslav Goljan, Department of Electrical and Computer Engineering, SUNY Binghamton

Is it possible to use the fine structure of pixelcolors in an image to identify the digital cam-era that took it, in much the same way that wecan identify a gun by looking at the scratchesit made on the bullets it fired? If so, how reli-ably can we distinguish between images ob-tained using different sensors or cameras? Isaccurate identification possible even from pro-cessed images? Reliable digital-camera iden-tification would especially prove useful in courtfor establishing the origin of images presentedas evidence, or, in a child-pornography case,to prove that certain imagery was obtained us-ing a specific camera and is not computer-gen-erated.

In classical film photography, there aremethods for camera identification that are com-monly used in forensic science. Some of thesemethods use camera imperfections, such asscratches on the negative caused by the filmtransport mechanism. To find an appropriateequivalent for the digital realm, one needs toturn to the process of digital data acquisition.

In a typical consumer digital camera, thelight from the photographed scene passesthrough the camera lenses. However, beforereaching a photo responsive sensor, the lightgoes through an antialiasing (blurring) filter andthen through a color filter array (CFA). Thephoton counts are converted to voltages, whichare subsequently quantized by an analog-to-digital converter. The digital signal is then in-terpolated (demosaicked) using color interpo-lation algorithms, color corrected, and the whitebalance adjusted. Additional processing in-cludes high-pass filtering and gamma correc-tion to adjust for the linear response of the im-aging sensor. Finally, the raw image is writtento the camera memory device in a user-selectedimage format (e.g., TIFF or JPEG).

Basing camera identification on artifactsintroduced by CFA or processing would notpermit us to distinguish between cameras thatshare the same sensors or processing algo-rithms. Thus, to obtain a unique identificationfingerprint, we need to turn our focus to theimaging sensor, which is usually a CCD(charge-coupled device) or a CMOS (comple-mentary metal oxide semiconductor) array.1

Imperfections of different kinds are introducedinto each sensor during the manufacturing pro-cess. As these imperfections are of a stochasticnature and are unlikely to be identical for dif-ferent sensors, they can be used to identify thecamera in a manner similar to identifyingpeople by their birthmark. Among the mostcommonly present defects are hot and dead

pixels. The problem with these is that somecameras remove them from images by post-processing and some cameras may not have anyof these defects.

Instead of relying on defective pixels,2 wehave proposed a new approach that uses thepattern noise of CCD arrays.3 This is, in fact, acollection of patterns caused by different phe-nomena, such as pixel-to-pixel non-uniformity,dust specs on optics, interference in optical el-ements, and dark current. However, not all ofthem are suitable for our goal. We are inter-ested in those noise components that do notchange with time, at least in the short term, andthat allow reliable extraction from typical im-ages. These requirements are best satisfied forpixel-to-pixel non-uniformity, which is causedby slight differences in the efficiency withwhich individual pixels absorb photons. Thedifferences are stochastic in nature and are in-evitably caused during the manufacturing pro-cess. As a result, each image the camera takesis overlaid with a weak noise-like pattern.

We extract this pattern using a denoisingfilter. To minimize the impact of the scene andimprove the estimation process, we averagethese patterns extracted from many imagestaken with the same camera, thus obtaining thecamera ‘reference pattern’. To determinewhether a specific image was taken with thiscamera, we simply extract the noise patternfrom the image and calculate its correlationwith the camera reference pattern. Based on thecorrelation value, we can reach a decision aboutthe image origin.

This methodology was tested on nine dif-ferent devices: low-end digital cameras; twosemi-professional DSLRs (Nikon D100 and

Sigma SD9); and the camera containing theFoveon X3 CMOS sensor (Sigma SD9). Ourset also contained two cameras of exactly thesame make (two Olympus Camedia C765 UZ).

The task that one typically encounters inpractice is to determine from several camerasthe camera that most-likely took a given im-age. This can be achieved simply by assigningthe image to the camera whose reference pat-tern has the highest correlation with the noisefrom the image. In our experiments with ninecameras and a total of 3000 images in both rawand JPEG formats, all images were correctlyclassified. We were also able to correctly iden-tify the camera from images that were pro-cessed using lossy JPEG compression withquality factors as low as 72 and additionallyresized with an adjusted contrast. Our experi-ments also confirmed that it is possible to dis-tinguish between cameras of the exact samemodel.

Jan Lukas, Jessica Fridrich,and Miroslav GoljanDepartment of Electrical and ComputerEngineeringSUNY Binghamton, NYE-mail: {jan.lukas, fridrich,mgoljan}@binghamton.edu

References1. G. C. Holst, CCD Arrays, Cameras, and

Displays, 2nd edition, JCD Publishing and SPIEPress, USA, 1998.

2. Z. Geradts, J. Bijhold, M. Kieft, K. Kurosawa, K.Kuroki, and N. Saitoh, Methods for identificationof images acquired with digital cameras, Proc.SPIE 4232, pp. 505-512, 2001.

3. J. Lukas, J. Fridrich, and M. Goljan, Determiningdigital image origin using sensor imperfections,Proc. SPIE 5685, 2005.

References1. C. E. Thomas Jr., et al. Direct to digital hologra-

phy for high aspect ratio inspection of semiconduc-tor wafers, 2003 Int’l Conf. on Characterizationand Metrology for ULSI Technology, Proc. AIPVol. 683, pp. 254-270, Austin, March 2003.

2. P. Bourgeat, F. Meriaudeau, P. Gorria, and K. W.Tobin, Content-based segmentation of patternedwafer for automatic threshold determination, Proc.SPIE 5011, pp. 183-189, 2003.

3. P. Bourgeat, F. Meriaudeau, K. W. Tobin, and P.Gorria, Patterned wafer segmentation, QualityControl by Artificial Vision, Proc. SPIE 5132,pp. 36-44, 2003.

Machine vision for defect detectionContinued from page 3.

4. P. Bourgeat, Segmentation d’images de semi-conducteurs appliquée à la détection de défauts,Ph.D. thesis, Université de Bourgogne, 2004.

5. P. Bourgeat, F. Meriaudeau, K. W. Tobin, and P.Gorria, Features extraction on Complex images,Proc. OSAV’2004, Int’l. Topical Meeting onOptical Sensing and Artificial Vision, pp. 103-110, Saint Petersburg, Russia, 18-21 October 2004.

6. F. Meriaudeau, P. Bourgeat, P. Gorria, and K.Tobin, Classifier Combination on Featuresextracted From Complex images: Application toDefect Detection on Silicon Wafers, Proc.QCAV’05, Japan, May 2005.



Motion-based particle filteringfor head tracking applicationsNidhal Bouaynaya and Dan Schonfeld, Multimedia Communications Laboratory,University of Illinois

Recent advances in multimedia and communi-cation require techniques for accurately track-ing objects in video sequences. The goal intracking is to estimate the posterior density ofthe target given all the observations. The esti-mate of the object is then given by the meanstate estimate. Particle filters provide a generalframework for estimating the probability-den-sity function of general non-linear and non-Gaussian systems.

Let Xk represent the target characteristics at

discrete time k (position, velocity, shape, etc).Denote all past observations up to time k byZ

1:k = {z

0,…,z

k}. Given a Markovian state space

model for Xk, we want to estimate the posterior

density p(Xk | Z

1:k). Let {X

k(n) , n=1,…,N} be a

set of samples drawn from a distribution q attime k. Let the set of weights {π

k(n),n=1…N}

be given by:

Then, it can be shown that the posteriordensity is approximated, at time k, by theweighted sample { (X

k(n), π

k(n)) , n=l,…,N }.1 The

conditional densities p(Xk | X

k-1) and p(Z

k | X

k)

are specified by the state space model. Noticethat, in theory, there is infinite number ofchoices of the importance function q, as longas its support includes that of the posterior dis-tribution. But of course when q is close to thetrue posterior, the particles are more effective.

We propose a motion-based proposal. Weadopt adaptive block matching (ABM) as themotion-estimation technique.2 We use a four-dimensional parametric ellipse to represent thestate vector of the head. The motion vector ofinterest, ∆X

k, is given by the difference between

the position of the center of the newly-fittedellipse and the position of the center of the pre-vious mean-state ellipse.

Each sample is translated in the x-y direc-tion by the motion vector estimated by theABM and diffused in the four coordinates ac-cording to a zero-mean white Gaussian noise.The motion-based importance function is thena sum of Gaussians:

where the notation N(µ, Σ) denotes the normaldistribution with mean µ and covariance ma-trix Σ.

In addition to the importance-functionevaluation, we also need to calculate the par-

ticle likelihood p(Zk | X

k) and transition prob-

ability p(Xk | X

k-l).

We use both face-color characteristics andedge-detection cues for the likelihood term. Weuse the color model proposed in Reference 3and the gradient model described in Reference4. We adopt a first order auto-regressive (AR1)process for the system dynamics.

Refinement processApplications, such as interactive imaging orvirtual reality environments, need an accurateelastic contour around the person’s head. Ac-tive contours take an initial estimate of the ob-ject and refine this estimate using an optimiza-tion procedure.

We simplify the 2D contour into a 1D modelby considering only observations λφ along thenormal lines σ of the contour. The total energyis a weighted sum of different energy terms.The weights are fixed for the whole video. Thedifferent energy terms are defined as follows:

is the exter-nal force making the contour move towards theedges, where I denotes the intensity.

is the internal en-ergy term and imposes smoothness on the con-tour.

is the shape energy term,

where σ specifies the accuracy of the initialcontour.

The causal definition of all the energy termsallows a dynamic programming scheme to findthe optimal sequence. To improve the perfor-mance of the active contour, we perform theoptimization procedure using properly-sampledinitial conditions. Specifically, we draw nor-mally-distributed samples with the mean of the

initial output of the tracker and a given covari-ance. Dynamic programming is performed foreach sample. The optimal elastic contour is thenthe one that corresponds to the minimum totalenergy.

ExperimentsOur experiments applied this technique to headtracking. We use challenging video sequenceswith more than 400 frames in cluttered envi-ronments. We simulate various tracking con-ditions, including camera zoom in and out, ap-pearance changes, fast and erratic movements,out-of-plane head rotation. During the track-ing, 50 particles are propagated. The refinementstage uses 10 perturbed initial samples.

ConclusionsIn this paper, we have demonstrated that real-time information extraction can be incorporatedinto the importance density function of particlefilters to significantly improve the trackingperformance. Specifically, we showed thatonline motion estimation can be used to trackfast-moving objects that undergo erratic move-ments. Moreover, the use of a robust contourextraction scheme allows us to employ the pro-posed tracking system in many applications(e.g., background hiding, virtual video-conferencing, video animation, robust process-ing, etc.).

Nidhal Bouaynaya and Dan SchonfeldMultimedia Communications LaboratoryDepartment of Electrical and ComputerEngineeringUniv. of Illinois, Chicago, ILE-mail: {nbouay1,dans}@uic.edu.

References1. A. Doucet, On sequential simulation-based methods for

bayesian filtering, Technical report CUED/F-INFENG/TR.310, 1998.

2. K. Hariharakrishnan and D. Schonfeld, Fast object track-ing using adaptive block matching, IEEE Trans. on Mul-timedia, to appear.

3. K. Nummiaro, E. Koller-Meier, and L. V. Gool, Objecttracking with an adaptive color-based particle filter,DAGM-Symp. Pattern Recognition, pp. 353-360, 2002.

4. M. Isard and A. Blake, Condensation conditional den-sity propagation for visual tracking, Int. J. Com-puter Vision 29, pp. 528, 1998.

Figure 1. Motion-based particle filter handles fast motion and out-of-plane rotation of the jumpingperson simultaneously .

Figure 2. Randomly-perturbed active contour results. The elastic contour nicely fits the head, whichcan then be easily extracted for further processes.



Managing levels of detail for real-time renderingin immersive virtual environmentsD. Paillot, F. Mérienne, M. Neveu, and S. Thivent, Le2i laboratory, Institut Image

Design reviews in virtual immersive en-vironments for the motor industry needa very-high geometrical quality, andreal-time visualization and interactivity.To be useful, such applications have tobe linked to the relevant computer-aideddesign (CAD) database. At the sametime, to work with virtual-reality appli-cations, CAD models have to be pre-pared in advance as digital mock-ups(DMUs). Currently, it takes severalmonths to get a full car model into avirtual reality application. In this article,a complete chain for exporting CAD de-signs into virtual reality models is pro-posed and developed (see Figure 1).

PreprocessingSince CAD models cannot be used inreal time, data preparation is required.This step is composed of three stages:the tessellation that converts a modelmade by surfaces into a set of polygons;the repair of the polygons (deletion ofmultiple entities, welding vertices/edges, etc.); and the deletion of all in-visible parts of the assembly (rib, riv-ets, screw holes, etc.). We implementthis stage using the P-Buffer algorithm,which requires as parameters the direc-tions of visualization represented by dif-ferent points of view of the assembly.For example, in the proposed imple-mentation, the reduction ratio for adashboard is typically about 75%. Allthese stages optimize the models for thebest visual quality. The digital mock-up operations must be as precise as thephysical mock-up.

For the design review of the inte-rior of the vehicle, all objects that arerendered using the virtual models are viewedin a cave-like environment: everything is ap-proximately the same distance from the ob-server and is distributed 360° around. In prac-tice, due to the requirements of real-time pro-cessing, a special scheme is needed to reducethe elements needed for rendering. This is re-ferred to as level-of-detail (LOD) manage-ment.1 Previously-published techniques, knownas decimation,2 can be used. Our reductionscheme offers the best trade off between theperceived quality of rendering and the numberof polygons required for scene rendering, basedon the response of the human visual system inthis environment.

Figure 1. The steps for exporting CAD models into virtual-realitymodels.

Figure 2. Geometrical-error dependency on the eccentricity.

Figure 3. Rendering scheme for two viewing directions: (a) looking atthe wheel; (b) looking on the right side on the dashboard.

Loading modelsThe proposed method involves displaying amodel with a controlled definition dependingon the visual acuity of the observer. Visual acu-ity is a function of the eccentricity of the retinaand can be evaluated by showing two modelswith various LOD at different eccentricities.The viewer is asked to distinguish betweenthem. An equal perception of two differentmodels means that the human visual system hasreached its limit. A statistical analysis allowsus to define the geometrical error limit depend-ing on the eccentricity. Results are presentedon the Figure 2 (black curve). This result issimilar to the curve proposed by M. Reddy,3

which represents the visual acuity de-pending on the eccentricity.

This perception limit can be used forLOD management. The viewing direc-tion is indicated in real time by a head-tracking system, so the eccentricity foreach model into the scene can be ad-justed. Using the dark curve in the Fig-ure 2, a decimated model can replace thereference without decreasing the percep-tion quality. To limit the computationaltime, three LOD schemes are chosen,corresponding to the grey line in the Fig-ure 2.

ResultsExperiments were performed on a dash-board of the 807-C8 car from PSAPeugeot-Citroen Company: there were54 objects in total. Each was decimatedwith three geometrical errors (0.25mm,1mm, and 2.75mm). When all the mod-els were displayed without decimation,the frame rate was 7fps for 500,000 poly-gons. With the LOD management de-pending on the eccentricity, the framerate and the number of polygons in thescene change in real time. Figure 3 il-lustrates the rendering for two viewingdirections: (a) looking at the wheel; and(b) and looking on the right side of thedashboard. The limits of the LOD zoneare marked in white. The proposedscheme reduces the number of polygonsto 94,000, thus increasing the frame rateto about 23fps without a decrease in theobserved visual quality of rendering.

D. Paillot, F. Mérienne, M. Neveu,and S. ThiventLe2i laboratoryInstitut ImageChalon sur SaÙne, FranceE-mail: [email protected]

References1. J. H Clark, Geometric Models for Visible Surface

Algorithm, Communication of the ACM 19 (10),pp. 547-554, 1976.

2. P. Heckert and M. Garland, Multiresolutionmodelling for fast rendering, Proc. of GraphicsInterface ’94, pp. 43-50, 1994.

3. M. Reddy, Perceptually Modulated Level OfDetail for Virtual environment, Ph.D. Thesis(CST-134-97), University of Edinburgh, 1997.



mation from the back-side printing can oftenbe seen in the resulting images (of the frontside of the page). This show-through is nor-mally an undesirable artifact in the scan thatone would like to remove. An example of show-through can be seen in Figure 1, where thescanned images from two sides of a duplexprinted page are shown with visible show-through in both.

Traditionally, thresholding is used to mini-mize the effect of show-through. Thresholdingsets scanned reflectance values above a selectedthreshold to unity (white). This approach workswell for pure black and white regions (e.g. Fig-ure 1(a)), but fails irretrievably for show-through seen in a light gray background, forinstance, in the region corresponding to the stateof Texas in Figure 1(b).

When scans of both sides of the page areavailable, the problem of show-through removalbecomes analogous to the signal processing prob-lem of echo cancellation.2 The images of the twosides help distinguish light gray printing on thefront side from show-through due to the backside. This by itself is, however, insufficient forthe development of a method for correction be-cause the show-through is a non-linear and im-age-dependent effect in the image intensity do-main (note that dark regions on the front sidehave no impact of show-through). A physicalunderstanding of the show-through process is re-quired in order to proceed. An elementary physi-cal model can be developed by considering theoptical properties of the elements involved.Through suitable simplification using the fact thatthe paper substrate scatters (or reflects) back amuch larger fraction of the incident light than ittransmits, it is possible to obtain a linearizedmodel for show-through of the form:3

(1)

where (x,y) denotes spatial location, Dsf(x,y) is

the (optical) density of the front side scan (de-fined as the negative logarithm of the scan re-flectance normalized by the white paper reflec-tance), D

f(x,y) is the density that would have

been obtained in the absence of any printingon the front side, h(x,y) is the ‘show-throughpoint spread function’, ⊗ represents the con-volution operator, and A

b(x,y) is the absorptance

of the back side (defined as one minus the back-side scan reflectance divided by the paper re-flectance). The show-thorough point spreadfunction (PSF) represents physical model termsthat ensure its value is small in spatial extentsimilar to a blurring filter.

Note that the equation is rather non-intui-

tive, in that two of the terms are in density anda third in absorptance. The equation would berather hard to arrive at purely through trialand error with different mathematical trans-forms. It, however, comes about naturally fromthe physical model of the problem with the useof the appropriate mathematical simplification.

Once the linearized model of show-throughis available, the mathematical signal-process-ing technique of adaptive filtering can be ap-plied to the problem of show-through correc-tion. This technique has the useful capabilityofautomatically estimating and tracking theshow-through PSFs h(x,y) are caused both bychanges in paper substrate and variations in reg-istration between the front and back-side im-ages due to imperfections in the scanning pro-cess. We present only sample results here andrefer the interested reader to Reference 3 foradditional details.

Corrected images (see Figure 2) correspond-ing to the scans of the two sides shown in Fig-ure 1 were obtained using the show-throughcancellation algorithm. From the results, it isclear that the algorithm is successful. It is ef-fective in eliminating the show-through notonly in white regions of the page but also inlight gray regions, such as those correspond-ing to the state of Texas in these figures.

Discussion and summaryThe example presented here illustrates how theproblem of show-through in document imag-ing can be solved by using suitable physicalmodels and mathematical tools. It is worthpointing out that the synergy and combinationof these elements is what allows the solution:either one by itself is insufficient. The use ofgradient-descent-based adaptive filtering byitself does not enable the solution. Instead, thephysical analysis that allows linearization ofthe problem is also a key element. In particu-lar, the problem becomes linear when imagedata on the front side is converted to densityand on the back side to absorptance: a rathernon-intuitive result that could be inferred onlythrough the use of the physical model. Like-wise, the physical modeling by itself is rathersimplistic and is not sufficient to solve the prob-lem since it does not readily allow us to com-pute the unknown show-through point-spreadfunction or the alignment between the imageson the two sides.

The synergy between physics and math-ematics is a particularly common and power-ful motif in imaging systems research that is

Imaging arithmetic: Physics ∪ Math > Physics + MathContinued from cover.

supported by numerous other examples. Theinterdisciplinary nature of work in imagingsystems makes these approaches that combinephysics and mathematics harmoniously evenmore compelling than other areas of research.

Gaurav SharmaElectrical and Computer Engineering andBiostatistics & Computational Biology Depts.University of Rochester, NYE-mail: [email protected]

References1. G. Sharma, Imaging Arithmetic: Physics ∪ Math > Phys-

ics + Math, Proc. SPIE 5667, pp. 95-106, Invited Paper,2005.

2. S. Haykin, Adaptive Filter Theory, Prentice Hall, NJ,2002.

3. G. Sharma, Show-through cancellation in scans of duplexprinted documents, IEEE Trans. Image Proc., 10 (5),pp. 736-754, May 2001.

movement is assumed. For this scenario, thepattern is fixed. The necessary phase shift ofthe pattern is achieved by the movement of thework-piece on the assembly line. As a periodicfringe pattern is used, one can automaticallycalculate the actual phase shift of the pattern.This is done by calculating the pixel-wise con-trast between two particular images: the firstimage, with phase shift zero; and a second im-age, the relative phase shift of which has to bedetermined. The pixel-wise contrast betweenthese images is averaged over the whole im-age region. Maximum contrast is reached for a180° phase shift, minimum contrast for 360°.With this knowledge, four images with 90°phase shift can be selected from the image se-quence acquired at the assembly line.Deflectometry in general is very successful forthe inspection of specular surfaces. Using ournew far-infrared extension of the technique,plastics and beamless metal sheets can be in-spected. For example, a high-quality examina-tion of car body parts can now be carried outprior to the costly varnishing.

Jan W. Horbach and Soeren KammelInstitute for Measurement and ControlSystemsUniversity of Karlsruhe, GermanyE-mail: {horbach, kammel}@mrt.uka.de

Deflectometric inspectionof diffuse surfacesContinued from page 12.



Join the SPIE/IS&T Technical Group...and receive this newsletter

JEI 2005 subscription rates (4 issues): U.S. Non-U.S.Individual SPIE or IS&T member* $ 55 $ 55Individual nonmember and institutions $305 $330

Your subscription begins with the first issue of the year. Subscriptions are entered on a calendar-year basis.Orders received after 1 September 2005 will begin January 2006 unless a 2005 subscription is specified.

*One online journal included with SPIE/IS&T membership. Price is for printed journals.

Send this form (or photocopy) to:SPIE • P.O. Box 10Bellingham, WA 98227-0010 USATel: +1 360 676 3290Fax: +1 360 647 1445E-mail: [email protected]

Please send me:■■ Information about full SPIE membership■■ Information about full IS&T membership■■ Information about other SPIE technical groups■■ FREE technical publications catalog

Technical Group Membership fee is $30/year, or $15/year for full SPIE and IS&T Members.

Amount enclosed for Technical Group membership $ _________

❏ I also want to subscribe to IS&T/SPIE’s Journal of Electronic Imaging (JEI) $ _________

Total $ _________❏ Check enclosed. Payment in U.S. dollars (by draft on a U.S. bank, or international money order) isrequired. Do not send currency. Transfers from banks must include a copy of the transfer order.

❏ Charge to my: ❏ VISA ❏ MasterCard ❏ American Express ❏ Diners Club ❏ Discover

Account # _______________________________________ Expiration date ______________

Signature ____________________________________________________________________

This newsletter is produced twice yearly and is available onlyas a benefit of membership of the SPIE/IS&T Electronic Imaging Technical Group.

IS&T—The Society for Imaging Science and Technology has joined with SPIE to form a technical group structurethat provides a worldwide communication network and is advantageous to both memberships.

Join the Electronic Imaging Technical Group for US$30. Technical Group members receive these benefits:• Electronic Imaging Newsletter• SPIE’s monthly publication, oemagazine• annual list of Electronic Imaging Technical Group members

People who are already members of IS&T or SPIE are invited to join the Electronic Imaging Technical Group for thereduced member fee of US$15.

Please Print ■■ Prof. ■ ■ Dr. ■ ■ Mr. ■ ■ Miss ■ ■ Mrs. ■ ■ Ms.

First (Given) Name ______________________________________ Middle Initial __________________

Last (Family) Name ___________________________________________________________________

Position ____________________________________________________________________________

Business Affiliation ___________________________________________________________________

Dept./Bldg./Mail Stop/etc. ______________________________________________________________

Street Address or P.O. Box _____________________________________________________________

City _______________________________ State or Province ________________________________

Zip or Postal Code ___________________ Country ________________________________________

Phone ____________________________________ Fax ____________________________________

E-mail _____________________________________________________________________________

(required for credit card orders)

(see prices below)

Electronic ImagingThe Electronic Imaging newsletter is published bySPIE—The International Society for Optical Engi-neering, and IS&T—The Society for Imaging Sci-ence and Technology. The newsletter is the officialpublication of the International Technical Group onElectronic Imaging.

Editor/Technical Group Chair Gabriel MarcuTechnical Editor Sunny BainsEditorial Assistant Stuart BarrManaging Editor/Graphics Linda DeLano

Articles in this newsletter do not necessarily consti-tute endorsement or the opinions of the editors,SPIE, or IS&T. Advertising and copy are subject toacceptance by the editors.

SPIE is an international technical society dedicatedto advancing engineering, scientific, and commer-cial applications of optical, photonic, imaging, elec-tronic, and optoelectronic technologies.

IS&T is an international nonprofit society whosegoal is to keep members aware of the latest scientificand technological developments in the fields ofimaging through conferences, journals and otherpublications.

SPIE—The International Society for OpticalEngineering, P.O. Box 10, Bellingham, WA 98227-0010 USA. Tel: +1 360 676 3290. Fax: +1 360 6471445. E-mail: [email protected].

IS&T—The Society for Imaging Science andTechnology, 7003 Kilworth Lane, Springfield, VA22151 USA. Tel: +1 703 642 9090. Fax: +1 703 6429094.

© 2003 SPIE. All rights reserved.

EIONLINEElectronic Imaging Web

Discussion ForumYou are invited to participate in SPIE’sonline discussion forum on ElectronicImaging. To post a message, log in tocreate a user account. For options see“subscribe to this forum.”

You’ll find our forums well designedand easy to use, with many helpfulfeatures such as automated emailnotifications, easy-to-follow ‘threads,’and searchability. There is a full FAQ formore details on how to use the forums.

Main link to the new ElectronicImaging forum:

http://spie.org/app/forums/tech/

Related questions or suggestions canbe sent to [email protected].

Reference Code: 4646



Deflectometric inspection of diffuse surfacesin the far-infrared spectrumJan W. Horbach and Soeren Kammel, Institute for Measurement and Control Systems, University of Karlsruhe

Deflectometry is an opti-cal method for the inspec-tion of specular freeformsurfaces such as var-nished car-body parts.Bumps, dents, waviness,scratches and coating de-fects can be detected us-ing this method. Up tonow, deflectometry wasrestricted to specular sur-faces. The specular re-flection from known pat-terns was analyzed, as theshape of the surfacecauses distortions in thepattern that provide infor-mation about the localsurface curvature.Clearly, it would be ad-vantageous to accomplishsuch an inspection priorto the costly varnishingprocess.

This is now possible,using infrared technol-ogy. Even beamless ob-jects like unpainted moulded or stamped partsare specular for far-infrared radiation. We de-veloped a pattern generator that emits such ra-diation (see Figure 1), the reflection of whichis observed using a thermal camera. We canrealize a phase shift by performing a lateralmovement of the pattern. By analyzing fourimages of the reflected phase-shifted pattern(see Figure 2) our algorithm can calculate thelocal curvature of the inspected object. It usesthe fact that the curvature of a spherical mirroraffects the focal point of the optical mapping.As the fringe pattern is shifted, each point onthe surface is illuminated with changing bright-ness. The contrast between shifted images atthe same point is influenced directly by thesharpness of the mapping and therefore by thelocal curvature of the reflecting point. This

contrast information is visualized in the result-ing image (see Figure 3).

In the example shown, a bump has been de-tected. As the part has not been painted yet, ahuman observer cannot see this defect. How-ever, since the defect would become visibleafter painting, the part must be rejected. To geta sense of the overall performance of our sys-tem, we compared our results to that of tactilemeasurements. These tests showed that, usingour setup, infrared deflectometry is able to de-tect bumps down to a minimum height of 15µmgiven a lateral dimension of the bump of about2mm. Because our method is sensitive to localcurvature, even flatter defects can be detected,if their lateral dimension becomes smaller.

We have shown that our method can be ap-plied to a moving assembly line if constant

Figure 1. Setup for infrared deflectometry:infrared radiation is emitted as a pattern that islaterally shifted. The radiation is reflected fromthe part and observed by the infrared camera.

Figure 2. Four images of the reflected phase-shifted pattern. The phase shift is90 ° between consecutive images, used to calculate the result.

Figure 3. Resulting image: the local curvature ofthe surface is represented by the grey value. Thedetected bump is marked with a circle.


membership/ei/pdfs/ei15-2.pdf june 2005 vol. 15, …gsharma/papers/imagingarithei...to download the...

Documents