Experimental investigation of feature descriptors for retinal image registration

Retinal imaging is an important test for diagnosis of eye diseases and treatment evaluation. One of the steps in eye fundus image processing is image registration. It is inevitable in order to eliminate geometrical differences, introduced during imaging with different settings or pursuing follow up disease screenings. One of available strategies for image alignment is feature-based approach where feature descriptors have an important role in registration process. The quality of feature descriptors affects feature matching performance and overall results of image registration. In this paper we present a comparison of various feature extractors in tandem with conventional, bio-inspired or deep neural network-based local feature detectors applied for retinal image registration. Comparative evaluation of descriptors has been carried out using Fundus Image Registration Dataset, measuring Euclidean distance between ground truth points after image alignment. We present the results showing the performance of various feature detector-descriptor pairs applied for retinal image registration.