Defect detection on hardwood logs using high resolution three ...

2004 International Conference on Image Processing (ICIP)DEFECT DETECTION ON HARDWOOD LOGS USING HIGH RESOLUTION THREE-DIMENSIONAL LASER SCAN DATALiya Thomas'. Lumine Mili', Cl@ord A. Shuffer3, Ed Thomas4'.3Department of Computer Science, 'Bradley Department of Electrical and Computer Engineering,Virginia Polytechnic Institute and State University, Blacksburg, VA 24061, USA. 4NortheastemResearch Station, USDA Forest Service, 241 Mercer Springs Road, Princeton, W V 24740, USA.{ lithomas/lmili/shaffer} @vt.edu, ethomas@fs.fed.us.ABSTRACTThe location, type, and sevcrity of external defccts onhardwood logs and stems are the primary indicators ofovcrdll log quality and valuc. External defects providehints about the intcmal log charactcristics. <strong>Defect</strong> datawould improve the sawyer's ability to process logs suchthat a higher valued product (lumber) is generated. Usinga high-resolution laser log scanner, we scanncd anddigitally photographed 162 red-oak and yellow-poplarlogs. By means of a new robust estimator that performscircle fitting, a residual image is extracted from laser scandata that are corrupted by extreme outliers induced by thescanning equipment and loose bark. The residualsprovide information to identify defects with heightdifferentiation from the log surface. Combining simplcshape definition rules with the height map allows mostsevere defects to be detected by determining the contourlevels of a residual image. In. addition, bark texturechanges can be examined such that defects not associatedwith a height change might he detected.1. INTRODUCTIONOver the last few decades, a broad variety of scanningtechnologies have emerged for wood processing. Severalscanning and optimization systems are on the market thataid in the sawing of logs into lumber. Among them are thedefect <strong>detection</strong> and classification systems of logs andstems. <strong>Defect</strong> <strong>detection</strong> on hardwood trees and logs canbe categorized into two areas: internal and external<strong>detection</strong>, to determine defects inside logs and on a log'ssurface, respectively. Currently, most available scanningsystems are external models that use a laser-line scannerto collect rough log profile information. These systemswere typically developed for softwood (pine, spruce, fir)log processing and for gathering information aboutexternal log characteristics [l]. Optimization systemsthen use this profile information to better position the logon the carriagc and improvc thc sawyer's decision-makingability. Adding exlemal defect information to theoptimization process is a natural extcnsion of currenttechnology.Our rcsearch foocuscs on determining a method oflocating external defects on log surfaces. To accomplishthis we employed a commercially available TriCamscanning system with four laser units [Z], which gcncratedhigh-rcsolution profilc images of the log surface io threedimensions. Such an image is then processed to determinethe location of the most severe defects: overgrown knots,rotten knots, holedgouges, and removed branches. Thesetypes of defects usually are associated with a significantsurface rise or depression depending on the dcfect type.The image is processed using a new robust statisticalmethod to fit a series of circles to the log data. Byanalyzing the radial residuals, defects characterized by aheight change from the surrounding log area can belocated.More specifically, a typical log size is 10 to 20 inchesin diameter and 8 to 16 feet long. Resolution of thescanned data is about 0.04x0.78 inz per data point.Typically, each line of log data can be approximated by aclosed curve resembling a circle or an ellipse. Hence, oneof the problems that we are dealing with is to tit aquadratic curve or surface to the recorded log data. Ittums out that these data are corrupted by gross errors asbark on logs often becomes loose, forming flakes.Furthermore, the supporting structure underneath the logblocks the scanner, causes missing data, and the shape ofthe structure can he seen in the scanned images.Statistically, measurements with large errors, known asoutliers, can be regarded as observations that deviate fromthe pattern formed by the majority of the data set.Consequently, classical estimators based on the leastsquaresmethod cannot he used here to carry out curve orsurface fitting because they generate incorrect estimates inthe presence of outliers. We need instead to resort torobust statistics as initiated by Huber [3] and further0-7803-8554-3/04/$20.00 02004 IEEE243

developed by Hampcl et al. [4]. Based on this theoreticalframework, we dcveloped a new generalized M-estimatorto tit circles to log data, which is able to downweight alltypes of outlicrs, hence bounding their influence on theestimates. Thesc fitted circles allow us to extractresiduals, dctcrmine contour levels, and finally detect andidentify log defects. Let us stress that the contour levelsrevealcd themselves as a powerful tool for shaping thetopology of the log surface and thereby, allowing us topinpoint any atypical elcvation or depression on the logsurface by means of appropriate statistical methods. As afuture work, pattern recognition methods will be carriedout for defect identification and classification once thetaxonomy of the defects has been completed.The paper is organized as follows. Section 2describes the robust fitting procedures for estimating logshape. Scction 3 characterizcs several severe externaldefects on hardwood logs. Scction 4 presents examples ofrcsults that we have obtaincd to illustratc the potential ofthis approach in dctccting surface dcfccts. Finally Section5 contains concluding remarks and future work.under the log. As evidenced from Fig. 1, the classicalleast-squarcs estimator failed to provide a good fit fromsuch corrupted data set. Obviously, a robust estimatorthat suppresses the outlicrs is needed here.The nonlinear form of thc circlc equation promptedus to develop a new robust estimation method that is anoutgrowth of the one proposed by Mili et al. [6]. It is ageneralized M-estimator (GM-estimator) that fits anonlinear rcgrcssion model of the circle given bywhere e = [pl, p?, pi]' is the parameter vector containingthe center coordinaies (pi. p2) and the radius pi of thecircle and where &=[xl, xJT is the two-dimensionalmeasurcmcnt vector containing the data in a cross sectionfrom the scanner. The robustness of this GM-estimatorstcms from the fact that it not only bounds the influenceof the model errors, e, but also of the errors innicasurements, ~=[ql, qd'. This is achieved byminimizing an objective function exprcsscd as2. LOG IMAGE PROCESSINGSevere extemal defects that correspond to rises ordepressions on the log surface can be observed from thethree-dimensional log surface image. This suggests thatwe may extract the height change on the log surface fromits 3-D image to determine the defect location. We applyrobust statistics techniques to fit quadratic CUNCS to crosssections in the presence of a high noise level in the logdata, and a new generalized M-estimator was developed.Orthogonal distances between fitted cwes and the logsurface data, or radial residuals, are estimated, whichreflects height change from the surrounding log area.<strong>Defect</strong>s characterized by surface rises or depressionscould thus be located.whcrc2.1. Circle Fitting and Outlier SuppressionSince logs are natural objects that are approximatelycircular or elliptical along the cross sections, weexperimented with fitting circles and ellipses to the logdata, which all together form a reference surface, orvinual log, needed for defect <strong>detection</strong>. 2-D quadraticcurve fining is a problem in nonlinear regression [SI. Oneof the main problems that we had to overcome is thepresence of extreme outliers and missing data in the logdata cross-sections. See Fig. 1 for such an example.Missing data are due to the blockage of the log by thesupporting structure while extreme outliers are irrelevantdata caused by both dangling loose bark and the scanningsystem, duplicate data generated by scanner calibrationerrors and unwanted data from the supporting structureFigure 1 -A log data cross-section with outliers markedand with a fitted circle obtained by means of the newrobust GM- estimator displayed in continuous line.Shown also is the least squares fit in dashed line.244

4. SIMULATION RESULTSUsing the gray-level image, shown in Figure 2, we cangenerate a contour plot (Fig. 4), where it is possible todisccm the areas containing likely defects based on heightinformation alone. We developed an algorithm to generatercctangles that enclose arcas with contour curves at thehighest level. The arcas are selected dcpending on theirsizcs. Our simulation showcd that the majority of themost protruding or depressing dcfccts with a diameter atIcast 3.5 inches is largely detected. However, therectangles as in Fig. 4 often enclosc a portion, e.g., acomer, of the cxtemal defect. Further analysis is requiredto correctly map the exact dcfect area.circle fitting, the resulting residual images from ellipsefitting, however detailed, arc noisier. Cylinder fittinggives a coarse hut smooth picture; circlc fitting providesmore details, but introduces certain types of noise; andcllipse-fitting presents the most-detailed residual image,however noise level is the highest as well.The generation and initial processing of the residualimage is not the final step of this work. At this point, onlylog unrolling and height analyses methods have beenexamined. Currently we are dcvcloping an algorithm thatdetermines whether an area of interest contains a sawnknot, by locating the approximately straight line segmentin a cross-scction. Texture analysis methods will also heinvestigated to dctcct defects without a significant heightchange, which are not detected using the contour method.Further, a preliminary study was conducted to extractfeatures of external defect types from randomly chosendefect samplcs. These features will bc used to train arobust clustering and classification system for the defectclassification.6. REFERENCESFigure 4-A portion of a contour plot generated from thcgray-scale residual image tn Figure 2. The warmer thccolor, the higher level the contour. The rectangle framesthe highest contour which encloses the defect.<strong>Defect</strong>s with little surface variation are not covered inthe highest or lowest contour curves and thus cannot beenclosed in the rectangles. Such defects include smallovergrown knots and sound knots, and severe or mediumdistortions. In general these are less severe than big knotsor dcep cracks on log surface.5. CONCLUSION AND FUTURE WORKThis paper describes our approach to detecting surfacedefects in hardwood logs. Robust estimation and filteringtechniques are well suited to this application. The currentprograms can process an entire log-data sample that maycontain large amounts of missing data and severeabnormal data (outliers) due to the nature of datacollcction. The robust methods and procedures in theseprograms apply common heuristics to remove outliers.The original log data are processed and transformed into asharper and cleaner representation of the entire log. Thequality of the gray-level image lays a solid foundation forthe remaining defect-<strong>detection</strong> process. Contour levels ofthe residuals make it possible for further narrowing downthe potential defect areas.We may analyze residual images generated fromcircle-, ellipse, and cylinder- fining methods and combinethe results to help better identify defects. Compared to[I] M. Samson, “Method for assessing the effect of knots in theconversion of logs into stmctllral lumber,” Wood and FiberScience, Vol. 25, No. 3, pp. 298-304, 1993.[2] Perceptron, Forest Products Division, “Mill Wide Scanningand Optimization,” Perceptron Inc., Farmington Hills, MI,1999.htto://~.usnr.comluerceutro~Droducts.htm.[3] P. 1. Huber. Robust Statistics. John Wiley, 1981.[4] F.R. Hampel, E.M. Ronchcni, P.J. Rousseeuw, and W.A.Stahel. Robust Slatislics: The approach Bused on InjlocnceFunctions. John Wiley, 1986.[5] W. Gander, G.H Golub, and’R. Strebcl, “Fitting of circlesand ellipses-least squares solution,” Tech. Rep. 217,lnsitiut fur Eisenschafiliches Rechnen, ETH Zurich,Switzerland, url: A~:~/fi~.inf.ethz.chldoc/tech-~~uo~~2~~/ ,1994.[6] L. Mili, M.G. Cheniae, N.S. Vichare, and P.J. Rousseeuw,“Robust State Estimation Based on Projection Statistics,”IEEE Trans. on Power Systems, Vol. 1 1. No. 2, pp. , 1996.[7] R.M. Haralick and L. Shapiro. Computer and Robol Vision,Volume 2. Addison-Wesley. 1992.[SI Y. Leedan and P. Meer, “Heteroscedastic Regression inComputer Vision: Problems with Bilinear Constraint,” Int.Joumal of Computer Vision, Vol. 37, pp. 127-150, 2000.246