Regional Gaussian descriptors and their application to object recognition
Yishu Liu · Jisuanji gongcheng yu sheji · 2007
Feature extraction is a challenging problem in the field of pattern recognition.Features mainly fall into two classes: boundary descriptors and regional descriptors.Gaussian descriptors,which have attractive properties and show promising performance,belong to the former class.This paper is an improvement and an extension of Gaussian descriptors.A series of novel invariants called regional Gaussian descriptors are constructed based on region.The method includes defining a regional Gaussian potential function(RGPF) and then constructing 8 regional Gaussian descriptors by calculating the averages of RGPF's along 8 circles with the same center.Some pro-perties of regional Gaussian descriptors including the invariance on translation,rotation,scaling changes and reflection are studied.A central advantage of these new features over Gaussian descriptors is that they are insensitive to noise and edge variations.To support our new theory,an algorithm for object recognition is designed based on regional Gaussian descriptors and numerical experiments are con-ducted,and experiment results give an encouraging high recognition rate.