Abdul AwwalAutomatic Alignment & Imaging TeamIntegrated Computer Controls System
National Ignition Facility (NIF)Lawrence Livermore National Laboratory
November 18-19, 2004
Two-Step Position Detection for NIF Automatic Alignment
Presentation toLLNL CASIS WorkshopLivermore, California
The work was performed under the auspices of the U.S. Department of Energy National Nuclear Security Administrationby the University of California Lawrence Livermore National Laboratory under Contract No. W-7405-ENG-48.
UCRL-PRES-208084
UCRL -208084
• Introduction
• Problem Statement
• Background
• Two step algorithm examples: simple to complex
• Results
• Challenges
Outline
UCRL -208084
NIF Facility flyover National Ignition Facility
UCRL -208084
National Ignition Facility
National Ignition Facility
MasterOscillator
CapacitorBay 3
LaserBay 2
TargetChamber
OpticsAssembly
ControlRoom
SwitchYard 2
UCRL -208084NIF-0202-0XXXXppt15/GHM/tr
UCRL -208084
• Automatic alignment uses
video images of fiducials to
find positions of laser beam
Automatic alignments need automated detection ofbeam position and automatic mirror adjustments
UCRL -208084
Problem: How do we detect positions when thefiducials vary in shape and size?
• Automatic alignment needs to detect position of beam
— Challenges
– Beams with different markers
– Same marker, but sizes vary due to defocus
– Circles of different sizes
– Variations relatively large
UCRL -208084
Alignment image contains circles and squares
UCRL -208084
Defocused spots exhibit wide variation in size andquality
UCRL -208084
Different markers identify different beams
UCRL -208084
Some alignment images vary from 35 pixels to200 pixels
UCRL -208084
How to choose the template?
• Centroid
• Template
Possible solutions
UCRL -208084
)),(exp(),(),(*),( yxyxyxyxCMF UUjUUFUUFUUH F-==
)),(exp(),( yxyxPOF UUjUUH F-=
[ ]myxyx
yxyxAMPOF
UUFdUUFcb
UUaFUUH
2),(),(
),(*),(
++=
CMF
POF
AMPOF
Algorithm background
UCRL -208084
cautocrosspos xxxx +-=
cautocrosspos yyyy +-=
Position is obtained from cross and auto-correlation and the original template location
UCRL -208084
CMF output shows cross-correlation location
Peakdetected
UCRL -208084
Problem
UCRL -208084
• Two circles, r1 < r < r2 and r3 < r < r4
• Square, d1 < d < d2
• Why ?
• By design
• Due to optics
Problem
UCRL -208084
Two step approach
• Step 1: Search for right radius
• Segment based on blob size— Segment 1, less than 800— Segment 2, between 900 and 1100— Segment 3, between 1200 and
1400
Segmentthe image
Search for r1 < r < r2DetectEdges
• Change to square, repeat step 1
UCRL -208084
Two step process
• Step 2: Find the position
• Change to square, repeat step 2
Correlate wholeimage withchosen radius
Find locationDetectEdge
UCRL -208084
Segmentation is performed based on pixel size
• Feature sizes of interest (area in pixels)
(1086, 1074, 1086, 836, 832, 836)
• Searching for squares (589.0, 158.5)
• Searching for circles at (52.0,158.0)
UCRL -208084
Movie of the circle match in the correlation planeshows correlation becomes a point at right radius
Correlation is alsoa circle
Circle becomessmaller as it nearsthe right radius
UCRL -208084
Search selected 16.5 pixels for circle radius
Correlation vs radius
0
0.5
1
1.5
2
12 14 16 18
radius in pixels
No
rmal
ized
co
rrel
atio
n
Circlecorrelation
UCRL -208084
Search selected square with dimension of 15.5pixels
Correlation vs side
0
0.5
1
1.5
2
2.5
14 16 18 20
square dimension in pixels
No
rmal
ized
co
rrel
atio
n
Squarecorrelation
UCRL -208084
Circles of two sizes and squares of a single sizedetected (after 3 searches and 3 correlations)
UCRL -208084
Example 2
• 5 circles, r1 < r < r2
— Where r1 = 35 and r2= 260
UCRL -208084
Pinhole images could vary from 35 pixels to 200pixels (radius)
UCRL -208084
Solution to complex problem: Example 2
• Successive approximation
• Search by matching
Section the imageat centroid
Estimate r1 < r < r2
Centroid
Correlate wholeedge image withradius range
Find location
UCRL -208084
Cross-section of images could provide an initialestimate of the radius such as 37 and 204 pixels
UCRL -208084
Edge detected image
UCRL -208084
Two typical final results show the edge and centerof the circular images
UCRL -208084
Refined algorithm
• Successive approximation
• Search by matching
Measure thepedestal
Estimate r1 < r < r2
Correlatesegmented imagewith range ofradius
Find location
Segment the image find ROI
More detailed in paper no. 5556-30
UCRL -208084
Box classifier
Data Cluster Classifier
Features
Classes
Featureestimation
UCRL -208084
Search space shows 3 classes, range is chosen tobound the class- box classifier
pedestal
radius
This causes 14 to belower limit of class III
Class I
Class II
Class III
UCRL -208084
Box classifier
if (pedestal lt 700) then minRad=5
maxRad=10elseif (pedestal lt 1000) then
minRad=11;maxRad=22
elseminRad=14 ;maxRad=27
end
Measure thepedestal area
Estimate r1 < r < r2
UCRL -208084
Detection of four defocused spots
UCRL -208084
Challenges
• Partially missing
• Blobs may be divided into two
• Successive approximation may fail if initial guess is incorrect
— automated guess correction by counting the right and wrong
moves
UCRL -208084NIF-0202-0XXXXppt15/GHM/tr
UCRL -208084
• Wilbert McClay, Jim Candy, Walter Ferguson, Scott Burkhart, Karl
Wilhelmsen (Automatic Alignment team)
• Erlan Bliss, Mark Bowers
• Eric Mertens, Patrick J. Opsahl (control room - data collection)
• Paul Van Arsdall (feedback)
Acknowledgement