Virtual Tool Mark Generation For Efficient Striation Analysis
Virtual Tool Mark Generation For Efficient Striation Analysis
Virtual Tool Mark Generation For Efficient Striation Analysis
doi: 10.1111/1556-4029.12435
PAPER Available online at: onlinelibrary.wiley.com
CRIMINALISTICS
Laura Ekstrand,1 M.S.; Song Zhang,1 Ph.D.; Taylor Grieve,1 B.S.; L. Scott Chumbley,1 Ph.D.; and
M. James Kreiser,2 B.S.
ABSTRACT: This study introduces a tool mark analysis approach based upon 3D scans of screwdriver tip and marked plate surfaces at the
micrometer scale from an optical microscope. An open-source 3D graphics software package is utilized to simulate the marking process as the
projection of the tip’s geometry in the direction of tool travel. The edge of this projection becomes a virtual tool mark that is compared to
cross-sections of the marked plate geometry using the statistical likelihood algorithm introduced by Chumbley et al. In a study with both sides
of six screwdriver tips and 34 corresponding marks, the method distinguished known matches from known nonmatches with zero false-positive
matches and two false-negative matches. For matches, it could predict the correct marking angle within 5–10°. Individual comparisons could
be made in seconds on a desktop computer, suggesting that the method could save time for examiners.
KEYWORDS: forensic science, tool mark comparison, computer simulation, screwdriver, statistics, striae
As a result of the recent concerns over the validity of impres- and marking angles produce different marks. Research continues
sion evidence analysis, many recent research efforts have to extend and improve these results.
focused on improving and reinforcing the scientific basis under- The relative initial success of automated analysis methods sug-
pinning firearms and tool mark analysis. Despite vastly varying gests that they could be used to streamline and enhance tool
focuses and approaches, initial studies appear to support the mark analysis in new ways. This article investigates the potential
existing conclusions of the forensics community that marks are of a new tool mark analysis approach based on the statistical
unique and reproducible. Petraco et al.,(1) for instance, used matching algorithm introduced in Chumbley et al.(3). This
established numerical pattern recognition algorithms to classify method seeks to link the tool tip directly to the evidence mark
75 marks in modeling clay from nine screwdrivers with an esti- through the generation of virtual marks. Both tool tip and evi-
mated error of under 3%. Moreover, Bachrach was able to use dence mark are digitized using 3D optical profilometry. A com-
preliminary 3D scans of bullet surfaces to qualitatively confirm puter simulation is used to generate virtual marks from the tip
the predicted surface features of the land- and groove-engraved geometry. The algorithm of (3) is used to evaluate the likelihood
areas of fired bullets (2). Chumbley et al.(3) used a statistical that a virtual mark is a match with a cross-section selected from
algorithm that compared correlations between subsets of mark the evidence mark. This method has the potential advantage of
profiles from 50 sequentially manufactured screwdriver tips to being able to rapidly generate virtual marks at a large variety of
the correlations resulting from random chance. This algorithm angles and twists and rapidly compare them to the evidence
could match screwdrivers to their marks with a 3% false-positive mark. Therefore, it could theoretically be used to exhaustively
rate, and the results supported the traditional conclusions of tool explore the possible match space and suggest the best matching
mark examiners that different screwdrivers, screwdriver sides, angle and twist to the examiner. The examiner could then focus
his or her traditional analysis on this orientation of the tool,
reducing the time required to perform the analysis and helping
to ease the burden on backlogged crime labs. This method
1 would also have the benefit of potentially reducing damage to
Ames Laboratory, Iowa State University, 2096 Black Engineering, Ames,
IA 50011. the tool tip caused during the generation of a large number of
2
Illinois State Police Retired, 3112 Sequoia Dr., Springfield, IL 62712. comparison marks in lead. Finally, as the virtual marks are
*Presented at the 65th Annual Meeting of the American Academy of derived directly from the tip geometry, error rates for the method
Forensic Sciences, February 18–23, 2013, in Washington, DC; also presented may be easier to estimate.
in part at the 42nd AFTE Annual Training Seminar; May 29–June 3, 2011,
This tool-focused approach is relatively new. Bolton-King
in Chicago, IL, and in part at the 2011 ASCLD Symposium, September
18–22, 2011, in Denver, CO. et al.(4) examined plaster casts of gun barrels but did not relate
†
Supported by Award No. 2009-DN-R-119 from the National Institute of them to the fired bullets. Geradts et al. suggested relating a scan
Justice and performed at the Ames Laboratory, which is operated by Iowa of the tool tip to the evidence mark in (5) but never imple-
State University under contract number DE-AC02-07CH11358 with the U.S. mented that approach. This is likely due to the difficulty experi-
Department of Energy.
enced in obtaining a 3D scan of the tool tip surface of sufficient
Received 7 Feb. 2013; and in revised form 4 June 2013; accepted 8 June
2013. accuracy. A profilometer must capture both the microscale surface
resembles a screwdriver tip, it is not as smooth as we might In contrast to the tips, the marked plates have relatively flat
expect. It has strangely “terraced” surfaces that suggest that part geometry and require only a minimal scanning procedure. The
of the microscale surface roughness has been lost in the spike additional light from the four fiber optic cables is unnecessary to
removal process. Therefore, it would be desirable to improve the obtain accurate results, and therefore, their light sources are
quality of the raw IFM scans. powered off. Gamma and color adjustments are made, and then,
To prevent the spike artifacts, an improved lighting technique the plate is scanned at 109 magnification. If spikes are present,
was developed. Two 150 W white light sources are positioned the gamma and color levels are readjusted and the plate is
around the IFM, and each is used to power a pair of flexible rescanned at 109 magnification. This is repeated until the plate
fiber optic lights. The ends of these cables are carefully and geometry is accurate. The high magnification can be used for
evenly placed about the screwdriver tip as shown in Figs 2 and iteration because the geometry is typically correct on the first or
4 to brighten the metal tip surface while avoiding the introduc- second scan.
tion of many specular highlights (glare spots) on the surface. The final resulting tip and plate scans at the 109 magnifica-
These highlights would cause imaging artifacts. Then, the tion have a spatial resolution of 0.804 lm. The tips and plates
gamma response of the IFM imaging system is adjusted until the for this study were digitized with a depth resolution of 2 lm
image of the tip surface contains no overexposed (or underex- (the IFM’s low setting). This setting was experimentally deter-
posed) areas. In these areas, the brightness or darkness of the mined to be sufficiently accurate for use with the statistical
surface is too extreme for the imaging system to correctly detect algorithm and appreciably increased the scanning speed. At
any surface details. Gamma correction instructs the IFM to the 109 magnification and the low setting, a plate or a tip takes
respond differently to a spatial change in the level of incoming c. 25 min to digitize.
light; the response becomes nonlinear (curved), so that small
changes in light level at the top and bottom ends of the detection
Data Preparation
range can be resolved. The IFM automatically compensates for
this gamma correction to accurately compute the depth. The To improve the quality of virtual mark generation and subse-
Alicona color correction tool is also used to ensure that the color quent statistical comparisons, the tips and marked plates require
response of the sensor is optimal for measuring the depth. These algorithmic cleaning to remove unneeded boundary regions and
2D image corrections affect the quality of the depth detection spike artifacts. Figure 5 presents the tip of 3(c) with its 2D tex-
because the focus variation technique relies upon 2D images to ture mapped onto its depth data both before and after algorithmic
compute depth. cleaning. From Fig. 5(a), we can see that the spike noise at the
After the adjustments to the 2D image, an iterative scanning ends is dark; the spikes occur where the screwdriver geometry
approach is used to ensure quality. The tip is scanned at 59 falls sharply away from the IFM detector. As this portion of the
magnification and the placement of the light cables is adjusted geometry is not useful for making virtual marks, we remove it
until the resulting scan is relatively devoid of spike artifacts. with basic image processing algorithms. Thresholding of the 2D
When this is satisfied, it is rescanned at 109 magnification to texture and the Alicona-computed quality map is used to remove
give the final geometry a finer spatial and vertical resolution. points that are dark and/or low in quality. The connected compo-
Figure 3(c) presents a representative result of using this lighting nents algorithm is then used to identify and keep only the main
technique. By visual examination, it was determined that the bulk of the tip geometry.
new scanning procedure yielded sufficiently clean geometry After the end noise is removed, the aforementioned spike
except for clusters of spikes at the ends of the tips. As the digi- removal algorithm is applied to remove any sharp spike artifacts
tized data is graphical in nature, the best, easiest, and quickest from the bulk of the tip geometry. First, a seventh-order polyno-
way to evaluate the quality of the data is with a simple visual mial is fitted to each column of data. Any point with a depth
examination. During this examination, several views are value 100 lm different from the value predicted by its column’s
inspected to ensure that the viewer is not fooled by the effects polynomial is removed. This process is repeated with the rows
of parallax. of the data. Finally, small holes are filled by linearly interpolat-
ing between the good points at their edges. The threshold value
and polynomial order for both the rows and columns were deter-
mined by experimentation.
The result of this cleaning for Fig. 5(a) is shown in Fig. 5(b).
The geometry resembles the edge of a screwdriver tip, and the
microscale surface roughness has the expected continuous
appearance. From a 3D visual examination of the cleaning
results for the tip sides, this cleaning procedure was determined
to function effectively. When tested on a 3.20 GHz desktop
computer, the entire cleaning process takes about 1–2 min.
Figure 6 presents a digitized marked plate before and after
algorithmic cleaning. In Fig. 6(a), the plate has dark edges with
spikes and no striations. These edges are the unmarked parts of
the plate. To remove these edges, the plate is first thresholded
using the 2D texture and the Alicona quality map in the same
manner as the tip geometry. Then, the x and y gradients of the
2D texture are computed and used to locate the edges of the stri-
FIG. 3––Spike noise removal. (a) Tip digitized with no additional lighting.
ated portion. Any data points beyond these edges are removed.
(b) Tip in (a) after algorithmic cleaning. (c) Tip digitized with new lighting The depth data is then median filtered to diminish the presence
procedure, prior to algorithmic cleaning. of any spikes in the striations. (Median filtering, unlike Gaussian
4 JOURNAL OF FORENSIC SCIENCES
FIG. 4––Close-ups of the tip lighting in the IFM, showing the four fiber optic cables. (a) Front view. (b) Side view.
and these small computations would add up to a large load that the direction of tool travel. The resulting flattened geometry is
would take a long time. An exhaustive search through the points extruded so that the camera can see it. This is performed using
would then have to be conducted to determine which of them two copies of the flattened geometry and directing OpenGL to
fell along the edge that produces striations. Finally, linear inter- draw triangles between them as they are pulled apart (a process
polation between the points would be required to resample the analogous to a game of cat’s cradle). The camera is then used to
edge with the appropriate resolution to correctly compare it with take a resampled depth image of this scene to retrieve the edge of
a cross-section from the marked plate. Therefore, performing the flattened tip. This edge is the simulated mark.
these computations on the CPU could take a very long time. Figure 9 shows a representative virtual tool mark made using
Nevertheless, the mark simulation is a computer graphics this process. The mark has very steep sides which arise from the
problem, and therefore, we can use a computer graphics library edges of the screwdriver tip geometry. These are not useful for
(such as OpenGL or Direct3D) to accelerate and simplify its the comparison and may confuse the comparison algorithm, so
solution. Computer graphics libraries provide easy access to the they are cut off. A trimming algorithm was developed to recom-
computer’s graphics processing unit (GPU). The GPU consists mend a cutting location to the user. This algorithm computes the
of a set of small processors known as shader processors (9). mean and standard deviation of the virtual mark, and then, mov-
Each shader processor can perform its own independent compu- ing in from the first and last points of the mark, finds the first
tations. This allows the GPU to process several points simulta- pair of points that have values within one standard deviation of
neously, greatly reducing the amount of time required for the the mean. The algorithm then repeats the process on the data
matrix multiplication process. A computer graphics library also lying between this new pair of start and end points to find a sec-
provides tools to simplify finding the edge of the flattened ond pair of start and end points that are closer together. The
screwdriver and resampling this edge. For this research, we sides of the box in Fig. 9 are these recommended points for
chose to use OpenGL as our graphics library as it is freely avail- trimming. Although the software user has the option to manually
able on the vast majority of desktops and laptops and runs under change these points, for this research, the autogenerated points
almost every operating system. were used for trimming as they were found to be reasonable
Figure 8 presents a simple analogy of how OpenGL and other through a visual examination. The entire virtual marking process,
computer graphics libraries work. The user defines a 3D object including the trimming suggestion, takes on the order of ten sec-
by giving OpenGL its x, y, and z values and positions and ori- onds on a 3.20 GHz desktop computer with a midrange graphics
ents it in the scene by directing OpenGL to multiply it by a ser- card intended for CAD applications.
ies of translation and rotation matrices. Likewise, the user
positions a simulated camera in the scene. OpenGL simulates
the process of taking a picture of the scene by multiplying the Results and Discussion
points of the object by a projection matrix to flatten them. Then,
Virtual Mark Verification
it resamples this image of the scene so that the new points are
arranged precisely in a rectangular grid. This results in a 2D pic- To validate the ability of the tool mark simulator to faithfully
ture of the scene that can be displayed on the user’s computer reproduce a mark from the tool tip geometry, three sets of artifi-
screen. Because the depth of the points in the scene are perfectly cial tool tip data were generated from standard geometric func-
known (they were given by the user), the camera can also pro- tions and used to generate virtual marks. Figure 10 presents the
duce a rectangularly sampled depth image, which is a 2D array cross-sections of these geometric functions compared with their
of the depth values of the points closest to the camera. virtual marks. The cross-section of each tool was generated from
To simulate the marking process using OpenGL, we defined 2000 samples of the geometric function (triangle, semicircle, or
the x-axis of the scene as the direction of tool travel and then trapezoid). This profile was copied 1000 times to form an
placed the screwdriver in the scene with the edge at the desired extruded tool which was then placed in the simulation scene
orientation to this direction. The screwdriver points are then mul- with the plane of the geometric cross-section perpendicular to
tiplied by a simple projection matrix that flattens the geometry in the direction of tool travel.
a small statistical study was carried out. Each side of each of the
six screwdriver tips was used to generate virtual tool marks in
increments of 5° beginning at 30° and ending at 90°. Each vir-
tual mark was then compared to the middle cross-section of each
marked plate using the statistical algorithm from (3). In the inter-
ests of better lighting, some of the plates were digitized in the
reverse direction; therefore, the virtual marks were also flipped
left to right and compared to the marked plates.
For this study, known matching virtual marks were defined to
be those made by the same side of the same screwdriver tip used
to make the physical mark and oriented in the same direction as
the scan of the physical mark. The correct orientation of the vir-
tual mark relative to the scan direction of the plate was deter-
mined by visual inspection. Known nonmatching virtual marks
were defined as those made with a different tip or tip side from
that of the physical mark. The assumption was made that differ-
ent sides of the same tip make their own unique marks because
the results of prior studies (such as [3]) have reinforced this
assumption. Moreover, known nonmatching virtual marks
included those made with the same side of the same tip that had
been flipped relative to the scan direction of the marked plate.
The flipped virtual marks of the other known nonmatches were
also considered known nonmatches.
Known Nonmatches
Figures 12–14 present the averaged T1 statistic values result-
FIG. 10––Comparisons of geometric standard tools to their virtual marks. ing from comparisons between 45°, 60°, and 85° marked plates
(a) Triangle function. (c) Semicircle function. (e) Trapezoid function. (b), (d), and known nonmatching virtual marks. T1 values are the
(f) Errors for (a), (c), and (e), respectively. measures of match likelihood produced by the algorithm of (3).
A T1 value near zero or a negative T1 value indicates a low
Due to the nature of the resampling process in OpenGL, the match likelihood, while a higher positive T1 value suggests a
resulting virtual mark could not be directly compared with the ori- high match likelihood. As part of the algorithm relies on taking
ginal tool cross-section. Although the virtual marking simulation random samples from the mark profiles, the T1 values can vary
samples the edge at the same resolution as the underlying tool for the same comparison. To reduce this fluctuation, the T1 val-
data, the camera may not be perfectly aligned with the tool data. ues used in this study were actually the average of 200 such T1
Therefore, there is some offset in where the sampling occurs along values for each comparison. In these figures and those that fol-
the geometric function. To correct for this, the inverse function of low, the T1 values are plotted in box plots. The box denotes the
the geometric cross-section was applied to the virtual mark and the middle 50% of the spread, and the dark line represents the med-
offset estimated as the median of the errors between the sample ian. The whiskers extend from the box to the nearest data points
locations. The geometric function was then sampled with this off- within 1.5 times the length of the box from its upper and lower
set and compared with the virtual mark. edges. (Thus, these are standard box plots.) Any points outside
The resulting root mean square errors were 0.0011%, 0.0055%, of this range are considered outliers and plotted with a circle.
and 0.0004% for the triangle, semicircle, and trapezoid, respec- In Figs 12–14, the large majority of the T1 values cluster
tively. These errors are negligible for the application of comparing around zero. As expected, there is no clear relationship between
tool marks. Moreover, they agree with similar results for 2D the boxes or medians and the angle of the virtual mark. The
OpenGL resampling found in (10). Therefore, the algorithm is lower whiskers fall near 1, and the upper whiskers fall near 1.
judged to accurately reproduce the geometry. The large majority of the outliers lie between 2.5. Previous
Visual inspections of the virtual marks also seem to reinforce experiences with this algorithm (such as the results in [3]) sug-
the validity of the tool mark simulation. Figure 11 presents a gest that the outliers between 2.5 are reasonable for the func-
representative visual comparison between a physical mark and a tioning of this algorithm. Nevertheless, there are some negative
known matching virtual mark. Both of these marks were made outliers of larger magnitude that cause concern.
with the same side of the same tip at an angle of 45°. As the fig- A visual examination of the comparisons that yielded negative
ure shows, while the two profiles have minor differences, the outliers below 2.5 revealed them to be caused by a weakness
visual resemblance is uncanny. Many of the features have been in the statistical algorithm. Figure 15 presents one such compari-
transferred faithfully from the tip to the plate. Nevertheless, there son. Here, the boxes denote the regions within the two profiles
is some minor spike noise present on the digitized tip that has with the maximum correlation. The algorithm first locates these
entered the virtual mark. regions and then randomly selects a common distance and direc-
tion relative to these regions as the location for a pair of valida-
tion windows. In the case of Fig. 15, the algorithm does not
Statistical Results
have many viable choices for this rigid shift. Because the maxi-
To evaluate whether the proposed method shows promise for mum correlation windows are at opposite ends, it cannot choose
reliably matching an evidence mark to the tool tip that created it, a distance far to the right or left. To the right, it runs into the
EKSTRAND ET AL. . VIRTUAL TOOL MARK GENERATION 7
FIG. 11––Known match comparison. (a) Digitized tip. (b) Virtual mark at 45°. (c) Digitized plate marked at 45°. (d) Middle cross-section of (c).
FIG. 12––Averaged T1 values from statistical comparisons of 45° physical FIG. 13––Averaged T1 values from statistical comparisons of 60° physical
marks to known nonmatching virtual marks. marks to known nonmatching virtual marks.
end of the virtual mark data, and to the left, it runs into the end opposite end problem was not the source of any false positives
of the physical mark data. As there is more data on the right or false negatives, and future work will modify the algorithm to
than the left, it will most likely put almost all of its validation prevent this problem.
window pairs there. In that small region, the virtual mark slopes
strongly upward, while the physical mark slopes downward.
Known Matches
Because these opposing slopes dominate the rigid shift validation
window correlations, the T1 statistic becomes strongly negative. Figure 16 presents the results for known matches with plates
A visual inspection of the other strongly negative comparisons marked at 45°. There is a clear trend in the T1 value distribu-
revealed that they were all caused by similar opposite end prob- tions as the angle of the virtual mark changes. Above 65°, the
lems. The opposite end problem could theoretically lead to high values cluster around zero in the same manner as those of a
positive T1 values and consequently false positives. The results known nonmatch. Below 65°, the distributions rise, peaking near
of the study were therefore visually checked to ensure that this 40° and 45°. At this peak, the distributions are actually very
8 JOURNAL OF FORENSIC SCIENCES
FIG. 14––Averaged T1 values from statistical comparisons of 85° physical FIG. 16––Averaged T1 values from statistical comparisons of 45° physical
marks to known nonmatching virtual marks. marks to virtual marks made with same side of the same tip.
tight, and every nonoutlier is well above 2.5. Nevertheless, there FIG. 17––Averaged T1 values from statistical comparisons of 60° physical
are two outliers near zero for 40° and 45°. Upon inspection, the marks to virtual marks made with same side of the same tip.
same marked plate, designated P5A-45, was responsible for both
of these outliers. Indeed, the T1 values for P5A-45 were below
1.0 for every virtual mark angle.
Performance Analysis
Figures 17 and 18 present similar results for known match
comparisons with plates marked at 60° and 85°, respectively. Based on these known match and known nonmatch results, a
Like the results for 45°, these results also display clear peaks in simple algorithm was designed to detect matches and infer their
the T1 value distributions near the angle of the marked plates. In angle. A comparison set was defined as the results of comparing
Fig. 17, the peak is slightly broader and has a few outliers under a certain marked plate to the virtual marks (30°, 35°, 40°, …,
it. A closer inspection of this data revealed that this is due to 85°, 90°) from a certain side of a certain tip. The maximum T1
different marked plates having slightly different T1 peaks rather value was found for each set in the results. If this T1 value was
than a false nonmatch with a particular plate. In Fig. 18, the greater than 2.5, the comparison set was recorded as a match
peak is narrower; the outliers under this peak are due both to and the angle at which this T1 value was produced was recorded
another false nonmatch with P2A-85 and slight differences in as the angle of the mark. Every match recorded by this algo-
the peak locations. rithm was confirmed to be a true match, and only two known
EKSTRAND ET AL. . VIRTUAL TOOL MARK GENERATION 9
TABLE 1––Error (°) between the predicted angle of an identified match and within minutes and generate and compare virtual marks within
the true angle. seconds. Therefore, this approach could quickly narrow down
the range of possible angles for the examiner, allowing him or
Statistic 45° Plates 60° Plates 85° Plates Total
her to make and compare fewer physical marks. This would save
Mean 1.667° 4.583° 6.364° 4.375° time and reduce the potential damage to the tool tip. Moreover,
Median 5° 5° 5° 5° given the speed of the software, it could potentially form the
Min 10° 10° 10° 10°
Max 10° 5° 0° 10°
basis for a database of searchable virtual marks and physical
marks.
Future research will seek to confirm these results on a larger
angle error was found when examining the mark-making jig set of tools and media. Moreover, potential modifications to the
shown in Fig. 1. A slop of about 1° was discovered in the four- statistical algorithm to reduce or eliminate the weakness will be
screw pattern, and a slop of 2–5° was discovered between the investigated.
screwdriver bit and the holder. Both of these would cause the
screwdriver tip to rotate slightly about the negative y-axis when
Acknowledgments
pulling force is applied to the handle to make the mark. Moreover,
as more force is required to make the marks as the screwdriver The authors would like to express special thanks to Dr. Max
handle approaches vertical, this slop may also explain the trend of Morris and Ms. Amy Hoeksema of the Iowa State University
increasingly negative mean angle errors. Statistics Department for their advice on integrating the statistical
Despite these error sources, the algorithm could correctly pre- algorithm into the software and their guidance in analyzing and
dict the angle of the identified matches within 5–10°. This summarizing the results.
agrees with the conclusion of tool mark examiners that marks
from the same side of the same tool will match if they were
References
made at angles within 15° of each other.
1. Petraco NDK, Shenkin P, Speir J, Diaczuk P, Pizzola PA, Gambino C,
et al. Addressing the National Academy of Sciences’ challenge: a
Conclusions method for statistical pattern comparison of striated tool marks. J Foren-
sic Sci 2012;57(4):900–11.
This study has introduced and investigated a new tool mark 2. Bachrach B. Development of a 3D-based automated firearms evidence
analysis approach that seeks to establish a more direct link comparison system. J Forensic Sci 2002;47(6):1–12.
between a tool and its mark. Both sides of six screwdriver tips 3. Chumbley LS, Morris MD, Kreiser MJ, Fisher C, Craft J, Genalo LJ,
et al. Validation of tool mark comparisons obtained using a quantitative,
were used to generate 34 marks on lead plates at angles to the
comparative, statistical algorithm. J Forensic Sci 2010;55(4):953–61.
horizontal of 45°, 60°, and 85°. Both the tips and the plates were 4. Bolton-King RS, Bencsik M, Evans JPO, Smith CL, Allsop DF, Painter
digitized, and virtual marks generated from the tip geometry JD, et al. Numerical classification of curvilinear structures for the identi-
were compared to plate cross-sections using the statistical algo- fication of pistol barrels. Forensic Sci Int 2012;220:197–209.
rithm of (3). Using a simple analysis of the likelihood values, 5. Geradts Z, Zaal D, Hardy H, Lelieveld J, Keereweer I, Bijhold J. Pilot
investigation of automatic comparison of striation marks with structured
matches were identified with zero false positives and two false light. In: Rudin LI, Bramble SK, Carapezza EM, editors. Enabling tech-
negatives. Moreover, the approach could determine the correct nologies for law enforcement and security. Proceedings of SPIE 4232.
angle of the matched marks within 5–10°. Although a weak- Boston, MA: SPIE, 2000;49–56.
ness was identified in the statistical algorithm, it was not the 6. Alicona Imaging GmbH. Focus variation. English edition, 2012;
issuu.com/alicona/docs/alicona_focusvariation3_e_high_resolution
fault of the virtual mark approach and did not affect the results
(accessed May 21, 2013).
of this preliminary study. However, the two false negatives sug- 7. Bolton-King RS, Evans JPO, Smith CL, Painter JD, Allsop DF, Cranton
gest that partial marking phenomena may need to be incorpo- WM. What are the prospects of 3D profiling systems applied to firearms
rated into the simulation for future research. and toolmark identification? AFTE J 2010;42(1):23–33.
These results, although preliminary, support existing conclu- 8. Kidd JA. Comparison of screwdriver tips to the resultant toolmarks [mas-
ter’s thesis]. Ames, IA: Iowa State Univ, 2007.
sions of the forensics tool mark analysis community and hold 9. Owens JD, Luebke D, Govindaraju N, Harris M, Kr€uger J, Lefohn AE,
potential both for reducing damage to the tool and for streamlin- et al. A survey of general-purpose computation on graphics hardware.
ing an examiner’s workflow. The outcomes of this study support Comput Graph Forum 2007;26(1):80–113.
the conclusions that each side of each screwdriver makes its 10. Zhang S, Ekstrand L, Grieve T, Eisenmann DJ, Chumbley LS. 3D data
processing with advanced computer graphics tools. In: Schmit J; Creath
own unique marks and that these marks vary significantly with
K; Towers C.E, editors, Interferometry XVI: techniques and analysis.
the angle of the screwdriver. Moreover, the success of the virtual Proceedings of SPIE 8493. Bellingham, WA: SPIE, 2012;849310–7.
marking simulation suggests that tip geometry is faithfully trans-
ferred to the lead plate during marking, and therefore, the poten- Additional information and reprint requests:
tial exists to link a mark directly to a tool. The comparison of Song Zhang, Ph.D.
mark directly to tool geometry reduces the population under Mechanical Engineering Department
Iowa State University
study from all marks to all tool tips, increasing the possibility of 2096 Black Engineering
establishing reliable error estimates for the analysis. Finally, the Ames IA 50011
optical profilometer does not alter the tip in any way during E-mail: [email protected]
digitization, and the software can prepare the digitized geometry