Image Registration for NIF Optics Inspection · 2007-07-26 · Image Registration for NIF Optics...
Transcript of Image Registration for NIF Optics Inspection · 2007-07-26 · Image Registration for NIF Optics...
Image Registration for NIF Optics Inspection
Presentation toCASIS, Signal and Imaging Sciences Workshop
Judy Liebman, Laura Kegelmeyer, Marijn Bezuijen
November 18-19, 2004
UCRL-PRES-208127
This work was performed under the auspices of the U.S. Department of Energy by University ofCalifornia, Lawrence Livermore National Laboratory under Contract.
camera at target chamber centerfocuses on each of the final optics
Final Optics Inspection Setup
Current Final Optics Registration
In-focus
Out-of-focus
Out-of-focus
Object growing?
Future Registration Adventures
• Optic lifetime reporting – follow optichealth through online and offline imaging
Current inspection
Previous inspection
Initial offline inclusion map
?
Initial offline damage map
?Future offlinedamage map
?
Registration Problem Characteristics
• 35MB images
• Lots of noise!
•Transformation type:
outliers perturbation errors
current future
Choosing Main Registration Approach
• Use whole image data directly
• Use features extracted from image
1. Feature Detection
2. Feature
3. Finding Optimal Transformation
Standard Feature-Based Registration Steps
(xa, ya)(xa, ya)(xa, ya)(xa, ya)(xa, ya)
?
A B
Find t g T to minimize: t(A) - B
(xb, yb)(xb, yb)(xb, yb)(xb, yb)(xb, yb)
Matching
2) Feature Matching
• Challenge: high % of outlying points leads to breakdownpoint
• Expecting a small transformation?• Distances between features in different sets < distances between adjacentfeatures in single set?
• Evaluate guess using expected matching %
• For larger transformations, reduce feature density using apriori knowledge
Feature Registration Without Feature Matching?
• Points vote to concur on thebest transformation
• Comparing two points onlysupplies x and y shift
• Similar feature consensusand clustering techniquesuse complex features to findrotation, scaling and shifts
• Adapt to use pairs ofpoints from each image?
Point setA
Point setB
One pointvotes for allpossible shifts
Incrementtransform spacewith votes
Find and apply peak shifts!
Current Final Optics Solution: Iterative RegistrationPoint set A Point set B
1. Use voting method toroughly find large shiftsbetween images
2. Use featuremeasurements andlocations to find optimalmatch
Optics through focus
Measurementsof single
feature
Dates of inspections
Measurementsof single
feature
Probabilistic Transformation Without Feature Matching
1. Choose random transformation
2. Evaluate transformation
3. Decide whether to accept new position• If evaluation sum > previous accept• If evaluation sum < previous accept with low probability
4. Iterate
01Distribution of Offsets
Distribution of Magnifications
Pointset A
Pointset B
0Distribution of Rotations
Developed by Marijn Bezuijen
Conclusion: Registration Now and Then
1. Vote for largeshift
2. Matching & leastsquares findssmall additionalshifts andmagnification
1. Vote for completetransformation usingsubset of features
2. Probabilistic method:midsize transformationusing all features
3. Matching & leastsquares finds smalloptimal transformation
Evaluate ResultsEvaluate Results
Evaluate Results
3) Finding Optimal Transformation
Finding shifts and scaling is a linear & separable problem:
•Two point sets:
• Want to find optimal scaling and shifting in each direction:
• To include rotation implement analytic least squares solution
(xa1, ya1)
(xa2, ya2)
(xa3, ya3)
A B(xb1, yb1)
(xb2, yb2)
(xb3, yb3)
Ax = b
xa1 1
xa2 1
xa3 1
xa4 1
x_scaling
x_shift =
xb1
xb2
xb3
xb4
x_scaling
X_shift = (ATA)-1AT b