The Dual Bootstrap Iterative Closest Point Algorithm with Application to Retinal Image Registration. Charles V. Stewart Chia-Ling Tsai Badrinath Roysam A new generalization of the Iterative Closest Point (ICP) registration algorithm is introduced and incorporated into a complete algorithm for registering retinal images. ICP works by iterating two steps: matching points based on the current transformation estimate and refining the estimate based on the matches. It requires a good initial estimate. By contrast, the Dual-Bootstrap ICP algorithm only requires an initial estimate that is a "toe hold" on the correct alignment - accurate only locally, over a small image region, and perhaps using a lower-order transformation than is needed to accurately align the entire images. The algorithm iteratively "bootstraps" both the region over which the model is accurate and the chosen transformation model, using a robust version of standard ICP restricted to the bootstrap region during each iteration. The covariance of the transformation estimate controls both bootstrap processes. The algorithm is designed to handle several dfficult circumstances in registration: (a) data from which relatively poor initial estimates can be reliably obtained, (b) data that has structural (geometric) complexity such as multiple proximate curves and surfaces, (c) data that has many missing or extraneous points, and (d) image pairs that have low overlap. In registering retinal image pairs, the Dual-Bootstrap ICP algorithm is initialized from similarity transformation estimates obtained by matching individual or pairs of vascular landmarks, and it aligns images based on blood vessel centerlines to produce quadratic transformations. On tests involving approximately 6000 image pairs, it successfully registered 99.5% of the pairs containing at least one common landmark, and 100% of the pairs containing at least one common landmark and 35% overlap or higher. This near flawless performance enables a variety of applications. Department of Computer Science, Rensselaer Polytechnic Institute, Troy, NY 06/24/2002 cs-02-09
The Dual Bootstrap Iterative Closest Point Algorithm with Application to Retinal Image Registration.
Charles V. Stewart
Chia-Ling Tsai
Badrinath Roysam
A new generalization of the Iterative Closest Point (ICP) registration algorithm is introduced and incorporated into a complete algorithm for registering retinal images. ICP works by iterating two steps: matching points based on the current transformation estimate and refining the estimate based on the matches. It requires a good initial estimate. By contrast, the Dual-Bootstrap ICP algorithm only requires an initial estimate that is a "toe hold" on the correct alignment - accurate only locally, over a small image region, and perhaps using a lower-order transformation than is needed to accurately align the entire images. The algorithm iteratively "bootstraps" both the region over which the model is accurate and the chosen transformation model, using a robust version of standard ICP restricted to the bootstrap region during each iteration. The covariance of the transformation estimate controls both bootstrap processes. The algorithm is designed to handle several dfficult circumstances in registration: (a) data from which relatively poor initial estimates can be reliably obtained, (b) data that has structural (geometric) complexity such as multiple proximate curves and surfaces, (c) data that has many missing or extraneous points, and (d) image pairs that have low overlap. In registering retinal image pairs, the Dual-Bootstrap ICP algorithm is initialized from similarity transformation estimates obtained by matching individual or pairs of vascular landmarks, and it aligns images based on blood vessel centerlines to produce quadratic transformations. On tests involving approximately 6000 image pairs, it successfully registered 99.5% of the pairs containing at least one common landmark, and 100% of the pairs containing at least one common landmark and 35% overlap or higher. This near flawless performance enables a variety of applications.
Department of Computer Science, Rensselaer Polytechnic Institute, Troy, NY
06/24/2002
cs-02-09