Retinal image registration based on the feature of bifurcation point

Retinal image registration is the process of matching or superimposing two or more fundus images obtained from different spaces, times or devices, which plays an important role in treatment and diagnosis of fundus diseases. This paper presents an algorithm for retinal image registration based on point correspondence. First the vascular tree is extracted by using gradient divergence. Then eight neighborhood search algorithm and Harris corner detection are combined to extract feature points. Next an improved local feature descriptor, which combines Euclidean distance with corresponding geometric information, is computed to get matching pairs. After that, random sampling consistency algorithm is used to purify point pairs. Finally, transformation matrix, which is computed according to matching point pairs, is used to register input images. Experimental results on public fundus image database FIRE show that our approach is comparable or superior to SIFT-based methods in terms of efficiency and speed.