Research On Homographic Point Sets Optimization For Non-rigid Images Registration

Image registration is a basic task in computer vision, pattern recognition and image processing. Recently, most classic homonymy point Sets extraction and matching algorithms for image matching problems center around the rigid deformation. However, there is a widespread and growing need of the non-rigid image registration. Aiming at the optimization of homonymy point sets, some fundamental theories and key technologies such as point sets uniform distributed in nonlinear scale space, local deformation invariant feature description and matching, local affine invariant geometric constraints, non-rigid geometry deformation are systematically studied based on the framework of image registration techniques.In candidate homographic point sets detection, for demands of non-rigid image registration, a method based on non-linear scale space and uniform distribution strategy are proposed, in order to extract the feature set of points which is uniform distributed along the line of the image. This method uses bilateral filter instead of Gaussian filter to build the non-linear scale space, and then extract a specified number of candidate homographic point sets based on uniform distribution strategy in both scale space and image space.In terms of local deformation invariant feature description, an oriented non-rigid deformation local invariant feature descriptor is proposed in order to better specialize the direction of the local feature and divide local support area. We first estimate the elliptical neighborhood of each feature point and indicate the main direction according to its second moment matrix, and then normalized the local support area to eliminate differences in the anisotropic deformation. Finally, we embed the local area into 3D space and construct a descriptor in terms of a heat kernel signature, which is invariant to deformation.In homographic point set matching and image registration, an image registration algorithm based on local features and geometric constraints is proposed in order to get accurate homonymy point sets. This method uses GIH feature and o-Da LI feature as the local describetor which is deformation invariant, and built the local affine invariant geometric constraint based on triangulation which is constructed by K-nearest neighbor algorithm and the Delaunay algorithm. Based on the utilization of local deformation invariant features and geometric constraints, an objective function is constructed and the corresponding match matrix is sloved, then the two images are registrated by the model of thin plate spline. The successful use of our algorithm applied in medicine and remote sensing image registration achieved good results.