Retinal Image Registration Using Geometrical Features

Abstract

In this study, we have introduced an accurate retinal images registration method using affine moment invariants (AMI's) which are the shape descriptors. First, some closed-boundary regions are extracted in both reference and sensed images. Then, AMI's are computed for each of those regions. The centers of gravity of three pairs of regions which have the minimum of distances are selected as the control points. The region matching is performed by the distance measurements of AMI's. The evaluation of region matching is performed by comparing the angles of three triangles which are built on these three-point pairs in reference and sensed images. The parameters of affine transform can be computed using these three pairs of control points. The proposed algorithm is applied on the valid DRIVE database. In general (for the case, each sensed image is produced by rotating, scaling, and translating the reference image with different angles, scale factors, and translation factors), the success rate and accuracy is 95 and 96 %, respectively.

Fusion of SPECT and MRI images using back and fore ground information

“…MRI image has a bond. Also the main images are registered, otherwise used images should be registered by the registration procedures [10]. The proposed method is compared with Brovey, IHS, YCbCr, Laplacian Pyramid, DWT and Contourlet.…”

Registration of adaptive optics corrected retinal nerve fiber layer (RNFL) images

“…Cross-correlation has been used already to register images obtained using various AO assisted flood illumination retinal cameras [21][22][23]. Although the crosscorrelation approach can work well, its performance can be severely compromised by factors such as changes in the image intensity, noise, and reduced overlap area when there is large image motion [24]. In this work, we show that the phase correlation technique is more robust than cross-correlation for the estimation of translational motion [25].…”

Retinal image registration under the assumption of a spherical eye

“…Local methods are the most popular, usually utilizing keypoint feature correspondences (Tsai et al, 2010;Chen et al, 2010;Perez-Rovira et al, 2011;Zheng et al, 2011;Lin and Medioni, 2008;Tang et al, 2011;Hernandez-Matas and Zabulis, 2014;Hernandez-Matas et al, 2015) or retinal features such as vessel trees (Matsopoulos et al, 1999) or vessel bifurcations (Stewart et al, 2003;Chaudhry and Klein, 2008;Ryan et al, 2004;Matsopoulos et al, 2004). Recently, hybrid methods that combine both global and local cues (Reel et al, 2013;Gharabaghi et al, 2012;Adal et al, 2014) are becoming increasingly popular.…”