Adaptive Registration of Very Large Images
Brian Jackson & Ardy GoshtasbyWright State University
Characteristics of Very Large Images
• They often have local geometric differences.• Finding correspondences
is nontrivial.•Management of time
and space is a challenge.
6/23/2014
Mapping Function• Break down the overall transformation
into a global and a number of local transformations.• Represent the global transformation by an
affine.• Represent each local transformation by an
affine.• Blend the global and local transformations
into a smooth mapping function.
6/23/2014
Coarse-to-Fine Matching• Reduce resolution sufficiently so
local geometric differences become negligible.• Find global transformation.• Increase resolution and globally
align.• Subdivide image domain, find
local correspondences, and find local transformations.
6/23/2014
Affine Parameters under Scaling
6/23/2014
Combining Global and Local Affines
• Mapping function at a level is obtained from a blending of local affine transformations.• A local affine is obtained from an estimate of the
affine at previous level and a refinement to the estimation.
6/23/2014
Blending Function•Rational Gaussian
6/23/2014
Preliminary Results
6/23/2014
6/23/2014
6/23/2014
Computational Requirements
• Computationally, it takes a couple of seconds on a Windows PC with Intel i7 processor to register one Mega-pixel images.• It takes 30 seconds to register ten Mega-pixel images.• Computation time increases linearly with image size.• The time to find correspondence between points in
images is about the same as the time to calculate the mapping function and register images.
6/23/2014
Limitations•It cannot register multimodality images.•Nonlinear geometric difference between corresponding blocks is ignored.•Image discontinuities due to occlusion are ignored.
6/23/2014
Top Related