Relative Moments and Their Applications to Geometric Shape Recognition

Shape recognition is a very important problem in computer vision.Geometric shapes can be divided into two classes: regions and single pixel structures.Hu's invariant moments are often used as features for characterizing shapes of regions.However when applied to structures,Hu's moments lose scale invariant.Chen proposed a method based on region boundaries to speed up the calculation of Hu's invariants.In this paper we extend Hu's and Chen's methods to define 10 relative moments,which are independent of position,scale and orientation.The formula of invariant relative moments we proposed can be applied to both regions and structures.Experiments show that the invariant relative moments recognition are effective to shape of regions,closed and unclosed structures.This method is also more convenient than Fourier Descriptors.