Transcript Poster III

An Efficient Multi-Resolution GA approach to
Dental Image Alignment
Diaa Eldin Nassar, Mythili Ogirala, Donald Adjeroh, and Hany Ammar
Dental radiographs show 2D projections of teeth. Changes in the relative
positions of the x-ray tube, the film and the teeth result in non-identical
projections of the teeth.
Aspect of the Image
Alignment problem
To compensate for the possible geometric discrepancies between a pair
of subject/reference teeth images, the dental image alignment stage
transforms the subject tooth image so that it becomes aligned to the
reference tooth image. The decision-making stage then determines
whether two aligned teeth images are matched or not.
The MR-GA method
Location and Orientation
Features
information of edge points
The Automated Dental Identification System (ADIS)
provides an automated web-based environment for
matching unidentified deceased individuals to missing
or wanted persons. Matching is based on comparison
of dental records of missing or wanted persons
against those obtained from the remains of deceased
individuals.
For each submitted subject record, ADIS generates a
short “match-list” of candidates whose dental features
are significantly close to those of the subject. The
“short” match-list is then inspected by a forensic
expert who makes a final decision on the identity of
each unidentified individual.
This Research is supported by the Criminal Justice Information Services division
(CJIS) of the US Federal Bureau of Investigations (FBI), the US National Science
Foundation (NSF) award number EIA-0131079, and the US National Institute of
Justice (NIJ) award number 2001-RC-CX-K013.
 X  a
  
Transformation Affine:  Y    c
  
Model
 1  0
  
Similarity
measure
Search
strategy
b
d
0
e  x 
 
f  y
 
1   1 
Oriented Hausdorff
distance:
Multi-resolution Genetic
Algorithm
To estimate the values of the parameters a, b, c, d, e, f :
• We start with the coarsest level and generate two random
subpopulations of chromosomes that correspond to points in the affineparameter search space.
• Each chromosome has 6 genes, each gene is 7 bits, and each
subpopulation contains 8 chromosomes.
• Reproduction rate is 10%, crossover rate is 70%, and migration each
other generation at a rate of 30%.
• We decode each chromosome to its corresponding parameters value
using inverse linear mapping.
• We evaluate the goodness of a chromosome, using the edge-oriented
Hausdorff distance.
• Genetic search at any scale continues as long as there is progress and
the maximum number of generations at that scale is not exceeded.
• Otherwise, we advance our GA search to the next scale after adjusting
the horizontal and vertical translation parameters accordingly.
• In effect, the alignment parameters are progressively refined as we
move from coarser to finer scales, with fast convergence as a result of
excluding the poorest chromosomes early at the coarsest scale.
Empirical results show that the MR-GA approach is capable of timely
correcting misalignments of up to 18.