An iterative image registration technique using local affine transformation

For registration between images, algorithms of point correspondence searching type and deformation field analyzing type have been proposed in the past. The former had a problem of increasing processing quantity and the latter had a problem of assuming small displacement or deformation. This paper proposes a cooperative image registration technique which absorbs finite displacement and nonrigid body deformation by iteratively applying affine transformations from coarse to fine for binary images expressed by sets of loci vectors of characteristic points. The process consists of four steps: (1) for each characteristic point of a reference image, a local affine transformation (LAT) weighted by Gaussian window functions is applied; (2) based on the least-squares method, LAT is optimized for each characteristic point of the reference image so that overlapping for a group of characteristic points of input images becomes best; (3) optimized LAT is applied for each characteristic point of the reference image and deformed reference image is generated; (4) iterating steps (1) through (3) while gradually decreasing the spreads of the Gaussian window functions, when the deformed image coincides with the input image, displacement between images is determined. We have applied this method to binary character images including large deformation and random-dot stereograms and obtained good results. Also, we evaluated the processing quantity by this method and showed that extension to gray scale images is easily formulated.