Remote sensing image registration based on KICA-SIFT descriptors

Abstract

This paper presents a method to construct efficient and distinctive descriptors for local image features based on Scale Invariant Features Transform (SIFT), namely, Kernel Independent Component Analysis Scale Invariant Features Transform (KICA-SIFT). KICA-SIFT is a improved version of the conventional SIFT for the two reasons: first, the improved SIFT descriptors are relative invariant to affine transformation, second, the Kernel Independent Component Analysis (KICA) is applied to obtain the independent components of the descriptors to improve the accuracy and speed of matching. It is can be used to register two remote sensing images that with large geometric and intensity variations. Experimental results for remote sensing image registration show the proposed method improves the registration performance compared to the related methods.