M-SIFT: a New Descriptor Based on Legendre Moments and SIFT

Zuo Xin,Dai Xiubin,Luo Limin

DOI: https://doi.org/10.1117/12.921074

2011-01-01

Abstract:There are many feature descriptors that are insensitive to geometric transformations such as rotation and scale variation. However, most of them cannot effectively deal with blurred image which is a key problem in many real applications. In this paper, we propose a new feature descriptor that combines SIFT descriptor with combined blur, scale and rotation invariant Legendre moment (CBRSL). The proposed method inherits the advantage of SIFT and CBRSL which leads to invariance for scale, rotation and blur degradation simultaneously. We also show how this new descriptor is able to better represent the blur and geometric invariant feature descriptor in image registration. The experimental results validate the effectiveness of our method which is superior to SIFT methods.