USING MULTI-SCALE AUTO CONVOLUTION MOMENTS TO GET IMAGE AFFINE INVARIANT FEATURES

Abstract

This paper includes two important works. First of all, the complete mathematical proof procedure of Multi-Scale Auto convolution was summarized and the simplified geometric proof the procedure was proposed.  Secondly, Multi-Scale Auto convolution moments were adopted to describe images’ maximally stable extremal regions to get affine invariant features of images.  In the second job, the Multi-Scale Auto convolution moments of the image features were calculated on each feature’s MSER region to form image features’ descriptors, and then the image feature matching was performed.  In order to verify the validity of the second job, the proposed algorithm were compared with the SIFT algorithm and MSER SURE algorithm.  Simulation experiments show that, for affine transformed images, the feature matching accuracy of the second job is much higher than the classical SIFT algorithm and the MSER SURE algorithm, which indicates that using Multi-Scale Auto convolution moments on the MSER regions could get effective affine invariant image features.