An automatic evaluation method for retinal image registration

The registration of retinal images is significant for image fusion, mosaicking, and retinal verification. Numerous methods have been proposed for the registration. The problem following this is how accurate an image pair has been registered. In this study, we propose an automatic method to quantitatively assess the registration of retinal images. To achieve the assessment, we use Canny edge detection combined with a target area mask that is generated based on vessel detection to detect the edges, and we then use edge dissimilarity evaluation as a criterion of the registration performance. The input is the registered images pair, which is preprocessed to normalize the intensity and contrast. Then we detect the retinal vessel and use the result to determine the target area through dilation. Canny edge detector then is used to detect edges inside the target area. Finally, to evaluate the dissimilarity of edges, modified Hausdorff distance is utilized to determine the final registration score. Experimental result shows that this method is robust in different light conditions or in anomaly situations.