Alain Beauchamp, PH'D'

1
(No Transcript)
2
Alain Beauchamp, PH.D.
Large Database
3
Today I would like to

• Tell you about the research that we have
  conducted with 3D since we last reported to you.
  Relative to
• 3D correlation performance in large databases of
  bullets
• The effect of various bullet composition /
  materials (e.g. lead copper jacketed) on the
  results
• Predicting correlation performance when lands are
  removed from consideration (e.g. emulate damaged
  bullets)
• Assigning a match probability to a 3D correlation
  score
• Make the case that 3D is better than 2D
• but 3D 2D combined is even better.

4
Presentation Overview

• Background
• 2D and 3D data
• Correlation Studies and Results - 2 Parts
• Large database analysis
• Simulated damaged bullets correlation

5
Background Three Types of Data

• Raw data (2 types)
• 2D image
• 3D topography
• from the BulletTRAX-3D acquisition system
• Interpreted data
• Simulation of a comparison microscope by the
  MatchPoint image analysis station

6
Background Raw Data (From BulletTRAX3D)
3D topography image
2D image
Grey Scaled Levels relative heights White ?
Peaks Black ? Valleys
7
Background Interpreted Data (using MPO)
8
Part 1 Large Database Analysis

• Goal
• Database general information
• Acquisitions
• Correlation Scores
• Correlation performance
• Visual analysis

9
Large Database Analysis

• Goal Measure the performance of the correlation
  on BulletTRAX-3D
• Pristine bullets
• Constraint More than 2000 acquisitions in 2
  months
• 4 calibers
• 160 to 400 pairs of known matches per caliber
• 2 materials/compositions
• Copper and lead
• 2D, 3D and combined 2D/3D correlations

10
Database General Information
Pittsburgh bullets database (Allegheny County
Coroners Office Forensic Laboratory Division)
11
Acquisitions

• The operator had no skill related to Forensic
  science or image processing
• No experience on BulletTRAX-3D and IBIS
• All 2136 bullets done in 2 months on 3 systems in
  parallel
• 20 minute acquisition per bullet
• --------------------------------------------------
  --------------------------------

Test bullet

Land1 Land2 Land3
Land4 Land5
Land6
Land1 Land2 Land3
Land4 Land5
Land6
Reference bullet
12
Correlation Scores

• For every possible pair of LANDS
• Compute all possible LAND-to-LAND correlation
  scores
• Get a correlation score matrix
• For every phase (column)
• Sum up the scores
• Get the Max Phase score
• --------------------------------------------------
  --------------------------------------------

• Phase
• 1 2 3
  4 5 6
• LAND (Reference)
• 1 184 31 19 11
  16 24
• 2 92 11 9 12 5 12
• 3 36 13 11 13 14 18
• 4 42 20 13 13 6
  31
• 5 87 68 11 9 22
  21
• 6 96 12 13 22 32 16
• -------------------------------------
  -------------
• MaxPhase 537 155 76 80 95
  122

Best Phase The phase with the highest sum of
LAND-to-LAND scores ( 1) Max Phase the sum
of LAND-to-LAND scores at the Best Phase ( 537)
13
Correlation Performance

• Computed from the top 10 in the sorted MaxPhase
  score list
• For each reference bullet
• Correlation against the rest of the database
• Sort the scores from the highest to the lowest
• There is a hit if the known match is in the top
  ten
• The performance is the percentage of hits
• For the 2D/3D combined performance,
• There is a hit if the known match is in the top
  ten in either the 2D or 3D score list

14
Correlation Performance (Contd)
Number of hits in the top ten
From 2D to 3D 15
From 3D to 2D3D 10
2D3D, Copper 85 2D3D, Lead 60
15
Visual Analysis (9mm Luger, Copper)
3D Rendering 2D(texture)
Rank3D Rank2D 1st
3D
2D
16
Visual Analysis (22 Long Rifle, Lead)
Rank3D Rank2D 1st
3D Rendering 2D(texture)
3D
2D
17
Why Some 3D Images May Look Fuzzy to You?
32 Levels of Gray

500 Levels of Gray
gt
18
Part 1 Summary

• 3D and 2D provide complementary information
• For 4 calibers and 2 materials/compositions
• Correlation
• 3D correlation is better than 2D (by 10),
• 3D 2D correlation is better than 3D alone (by
  15)
• Visualization
• Better with a combination of 3D 2D data
• 3D2D correlation performance (top 10)
• Copper (85), Lead (60)

19
Part2 Simulated damaged bullets correlation

• Goal
• Correlation Scores
• Database
• Correlation performance

20
Simulated damaged bullets correlation

• There are 2 problems with damaged bullets
  correlation
• Deformation
• Missing lands or lands that cannot be acquired
• This can be simulated from pristine bullets
• Goal Measure the effect of the missing lands
  (only) by using correlation results from pristine
  bullets
• We can use the correlation results already
  available from the large database study.
• Remove some LAND-to-LAND scores in the
  computation of the Max Phase score

21
Simulated damaged correlation scores
Test (Pristine bullet)
Land1 Land2 Land3
Land4 Land5
Land6
Land1 Land2
Land3 Land4 Land5
Land6
Reference (Damaged bullet)
--------------------------------------------------
--------------------------------------------------
-------
Phase 1
2 3 4 5 6
LAND (ref) 1 184
31 19 11 16 24 2
92 11 9 12 5 12 3
36 13 11 13 14
18 4 42 20 13
13 6 31 5 87
68 11 9 22 21 6
96 12 13 22 32 16
-----------------------------------------
--------- MaxPhase 537 155
76 80 95 122 MaxPhase
403 124 54 55 84 79
2 Lands Permutations (1,2), (1,3), (1,4),
(1,5), (1,6) (2,3), (2,4), (2,5), (2,6) (3,4),
(3,5), (3,6) (4,5), (4,6) (5,6)
22
Database
50 sister pairs (100 bullets) picked randomly
from the 9mm Copper database (848
bullets) Simulated Damaged1 vs. Pristine2
Permutations 1 Acquired Land 6 Lands 6 2
Acquired Lands 6 Lands 15 3 Acquired Lands 6
Lands 20 4 Acquired Lands 6 Lands 15 5
Acquired Lands 6 Lands 6 6 Acquired Lands
(Pristine) 6 Lands 1 1) Reference bullet
2) Testing bullet
23
Correlation performance
pristine-pristine

• 3 lands over 6 is enough to get within 5 a
  Pristine-Pristine correlation performance.

24
Overall Conclusions

• Large Database 2000 bullets\4 calibers\2
  compositions
• 3D2D (combined) is better than 2D and 3D alone
  for correlation and visualization
• 3D2D correlation performance
• Copper (85), Lead (60)
• Simulated damaged bullets correlation
• - 3 Lands over 6 is enough to get within 5 of a
  Pristine-Pristine
• correlation performance.

25
(No Transcript)