3D Reconstruction and Registration for Retinal Image Pairs

Structure analysis of the optic-nerve head (ONH) and retinal image registration are important for diagnosing eye diseases. Thus, a 3D reconstruction method and a retinal image registration method are proposed in this paper. The former reconstruction method exploits disparity map estimation as a depth cue to disambiguate the relative positions and then uses the guided filter to obtain the final depth map. Using the depth map, a red-cyan anaglyph can be computed and the 3D structure of the ONH can be reconstructed. The latter registration method exploits and establishes the vasculature relationship between retinal images as a mosaic cue to be used for feature extraction and matching, and then divides the vessel segmentation image into four quadrants and obtains the Grid-based Motion Statistics (GMS) feature correspondence technique for each region. Using the regional GMS feature correspondence, the transformation parameters can be accurately identified and a good registration result can be ensured. Experiment results demonstrate the benefit of the proposed reconstruction and registration method.