Locating Exterior Defects on Hardwood Logs Using High Resolution Laser Scanning Liya Thomas 1, Ed Thomas 2, Lamine Mili 3, and Clifford A. Shaffer 4 1 and 4: Department of Computer Science 3: Dept. Electrical and Computer Engineering Virginia Tech Blacksburg, Virginia, USA 2: US Forest Service Princeton, West Virginia, USA June 20, 2005
Accurately locating defects allows operators to improve product value Expected savings would be $1.2 billion Fewer trees need to be harvested Helps strengthen domestic wood products industry
Definition: Manually or automatically detect and classify the location, shape, size, type, etc. of external or internal defects of softwood or hardwood logs and stems. Categories: External vs. Internal Softwood vs. Hardwood CT/X-ray, MRI, Ultrasound, Microwave, Laser Scanning Detection methods on hardwood and softwood very differentDetection methods on hardwood and softwood very different
Most research groups focus on internal Various systems over a few decades Large and accurate data Problems and difficulties
External defect detection is relatively new Data include digital images and 3-D laser-scanned surface profile Data do not contain information about log internal structure
External Defect TypesOver-grownKnot SoundKnot HeavyDistortion Adven-titiousKnot MediumDistortion Adven-titiousBranch Adventitious Knot Cluster Wound ExternalDefects UnsoundKnot
Log Sample Collection 3D Data Acquisition Radial Distance Image Defect Feature Extraction Contours Detection
Problem Statement No system available Existing technologies Systems for softwood sawing are not directly applicable. The system relies on laser-scanning equipment, which is safe to operators and at a reasonable cost. Log defects should be identified in the presence of bad data (outliers).
Focus of This Research 1. Examine the modeling of circle, ellipse, and cylinder 2. Surface fitting using GM-estimator 3. Defect detection based on contour levels derived from robust radial distances 4. Numerical methods for solving nonlinear equations 5. Presently we use the iteratively reweighted least- squares (IRLS) method together with QR decomposition and Householders reflections for numerical stability.
Methodologies and Algorithms Robust estimation: circle, ellipse, cylinder fitting using GME to generate appropriate reference surface in presence of missing data and severe outliers Radial-distance extraction with respect to reference to provide a foundation—radial- distance image—for subsequent tasks Radial-distance analysis through contouring to extract information that may help reveal the presence of defects
Experimental Results New and challenging research New robust Generalized-M Estimator with projection statistics to fit circles to log cross- section data Radial-distance images are obtained, based on which contour images are generated Probability of detection of 81% for the most serious defect classes, and 19% of defects falsely detected
Data: with missing data and severe outliers Circle fitting: robust GME algorithm with projection statistics Outlier removal: confidence intervals Preliminary Results in Robust Regression
A 3-D Presentation of Detection Results
Issues to Be Addressed More Data, More testing System integration Identify defects with bark patterns but no surface rise Classify defect types Link detection information with internal defect modeling system
16 Thank you! Liya Thomas: Ed Thomas: Lamine Mili: Clifford A. Shaffer:
17 Extra Slides
Log surface topology of a red oak. Note the missing data sections, both due to the size of this log and the supporting equipment during the scanning, as well as outliers that outlines the shape of supports but not part of log surface data.
Circle and Ellipse Fitting GME Algorithms(1) Radial-distance image from Circle FittingFrom Ellipse Fitting
Contour image (Circle Fitting)Contour image (Ellipse Fitting) Contour Levels of Radial Distances, #480 Circle and Ellipse Fitting GME Algorithms(2)
Haralick, Watson, et al.: Topographic Primal Sketch Tian & Murphy, Rao & Schunck: Oriented Texture Analysis Kass et al.: Active Contour Model
Illustration of an abstract external log defect Along Log Length Along Cross Section l l h w w Border Line at the Base
f(p, x + ) + e = 0 (x 1 – p 1 + 1 ) 2 + (x 2 – p 2 + 2 ) 2 – p e = 0 Circle-Fitting GM-Estimator … …
Circle-Fitting Functions
Circle-Fitting Functions: Projection Statistics