Unsupervised Registration Network for Fundus Image Mosaicking

In this paper, we propose a novel unsupervised registration network (FunUR) for fundus image mosaicking. The FunUR contains a keypoint detection module (KDM) and an image registration module (IRM). The KDM learns to extract the matched keypoint pairs in a self-supervised manner and provides a similarity criterion for the IRM training, while the IRM can effectively learn the long-range correlation and regress a quadratic transformation modeling the spherical shape of the retina with a large viewpoint change. We validate the proposed FunUR on the public FIRE dataset and the private ROP dataset, and compare it with other state-of-the-art methods. The experimental results demonstrate a significant improvement in robustness and accuracy compared with other methods.