TY - GEN
T1 - Image registration by minimization of mapping complexity
AU - Myronenko, Andriy
AU - Song, Xubo
PY - 2009
Y1 - 2009
N2 - The criterion for the correct spatial alignment is a key component in image registration. We formulate the registration problem as one that finds the spatial and intensity mappings of minimal complexity that make images exactly equal. We do not assume any parametric forms of these functions, and estimate them within variational calculus. We analytically solve for non-stationary intensity mapping, eliminate it from the objective function and arrive with a new similarity measure. We name it the Mapping Complexity (MC) similarity measure, because it achieves the optimum when intensity and spatial mappings are of minimal complexity. Due to its general formulation, the similarity measure works both for complex intensity relationships (e.g. multimodal registration) and for spatially-varying intensity distortions. Our similarity measure can be interpreted as the one that favors one image to lie mostly within a span of the leading eigenvectors of the kernel matrix, where the kernel matrix is constructed from the second image. We introduce a fast algorithm to compute the similarity measure. In particular, we introduce a fast kernel vector product (FKVP) algorithm, which is of general interest in computer vision. We demonstrate the accuracy of the new similarity measure on several mono- and multi-modal examples with complex intensity non-uniformities.
AB - The criterion for the correct spatial alignment is a key component in image registration. We formulate the registration problem as one that finds the spatial and intensity mappings of minimal complexity that make images exactly equal. We do not assume any parametric forms of these functions, and estimate them within variational calculus. We analytically solve for non-stationary intensity mapping, eliminate it from the objective function and arrive with a new similarity measure. We name it the Mapping Complexity (MC) similarity measure, because it achieves the optimum when intensity and spatial mappings are of minimal complexity. Due to its general formulation, the similarity measure works both for complex intensity relationships (e.g. multimodal registration) and for spatially-varying intensity distortions. Our similarity measure can be interpreted as the one that favors one image to lie mostly within a span of the leading eigenvectors of the kernel matrix, where the kernel matrix is constructed from the second image. We introduce a fast algorithm to compute the similarity measure. In particular, we introduce a fast kernel vector product (FKVP) algorithm, which is of general interest in computer vision. We demonstrate the accuracy of the new similarity measure on several mono- and multi-modal examples with complex intensity non-uniformities.
UR - http://www.scopus.com/inward/record.url?scp=70449556154&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=70449556154&partnerID=8YFLogxK
U2 - 10.1109/CVPR.2009.5204345
DO - 10.1109/CVPR.2009.5204345
M3 - Conference contribution
AN - SCOPUS:70449556154
SN - 9781424439911
T3 - 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2009
SP - 17
EP - 24
BT - 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2009
PB - IEEE Computer Society
T2 - 2009 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops 2009
Y2 - 20 June 2009 through 25 June 2009
ER -