A Robust Outliers’ Elimination Scheme for Multimodal Retina Image Registration Using Constrained Affine Transformation

摘要

This paper proposes a robust outliers' elimination scheme for the registration of multimodal retina images. Our proposed constrained affine transformation Least Trimmed Squares (CAT-LTS) method has been designed to deal with image registration problems where the putatively matched feature points has a very large fraction of wrong matches. The constrained affine transformation allows all combinations of transformations such as scaling, rotation, translation and shear but disallows reflection. We use the Scale-Invariant Feature Transform (SIFT) feature points and Partial Intensity Invariant Feature Descriptors (PIIFD) to obtain the putatively matched feature points. We show that our proposed scheme when applied to the application of registering color fundus to enface optical coherence tomography (OCT) images significantly outperforms other outliers' elimination methods, namely the m-estimator sample and concensus (MSAC) and Random sample consensus (RANSAC) methods.