Registering retinal images using automatically selected control point pairs

Describes a method of registering retinal images automatically. Control points are automatically identified in each image from blood vessel segments extracted from both images. The location of the optic nerve is used check the spatial similarity of control point pairs. The control point pairs are ranked with a similarity assessment that calculates a correlation of image intensity around each control point. Using a model of an idealized registration, the authors calculate the expected scaling factor between the images. Control point pairs that differ from this expected scaling factor are eliminated, with a bias against pairs with a low similarity assessment. Accurate registration is reported in 22 out of 23 image pairs. The registration error is related to the errors from the methods used to extract the vascular tree and to identify the location of the optic nerve