A Robust Feature-Based Method for Mosaic of the Curved Human Color Retinal Images

Accurate registration is essential for montage synthesis, change detection, and design of computer-aided instrumentation. This paper describes a robust local feature-based for automatic mosaic of the curved human color retinal images. The kernel of this method is the m space scale invariant feature transform (mSIFT). The mSIFT algorithm is designed to overcome the SIFT's drawback that detects less features in the flat regions. Using the mSIFT algorithm, second-nearest-neighbor strategy, inlier identification, bilinear warping and multi-blending techniques, pairs of the curved color retinal images can be mosaicked to create panoramic images. Experiments show that the proposed method works well with the rejection error in 0.2 pixels, even for these cases where the retinal images without enough discernable structures, in contrast to the state-of-the-art algorithms.