3D Ear Shape Feature Matching Based On Graph Theory

Biometrics is a kind of science and technology using individual’s intrinsic behavioral orphysiological characteristics to personal identity recognition. It provides a highly stable andreliable approach to the identity recognition. It has become a new star of biometrics due to itsunique biological location and structural feature of human ear and has gaining more and moreconcern. Ear recognition has prospective potential application in the field of public safety,information security and so on; it increasingly attracted people’s attention.There are many problems needed to be solved if we want to develop a real robust andpractical ear recognition system, because the study of ear recognition is still on the stage ofexperiment and test so far. A number of experiment studies for ear recognition are performedin this paper, the minimum spanning tree and graph matching algorithms based on graphtheory are applied into the ear recognition, the main works are as follows:(1) The IsoRank algorithm for the graph matching is used to ear matching, the method ofglobal alignment is adopted by matching the neighbor topology to solve the ear keypointgraph matching problem approximately, and finally we will find the correspondence betweenthe nodes of the two graphs to obtain the maximal overall match between the two graphs tocomplete the ear recognition.(2) Combining the graph structure and ear shape feature, the path following algorithm isused accomplish ear recognition firstly by linearly interpolating between the convex andconcave relaxations, we will obtain a solution path of convex-concave minimization problems,and secondly following the solution path find the local minimum to obtain the correspondingbetween the ear keypoint graphs. Finally the ear recognition result would be estimated by thecombinational difference measure.(3) Firstly the similarity of ear local shape feature is utilized to align the keypoints of ear;the minimum spanning tree is represented by the similarity of the two ears; secondly the localshape feature from the keypoints is incorporated into the joint α-entropy and the minimumspanning tree is utilized to estimate the joint α-entropy. Finally the ear recognition isaccomplished by estimating the joint α-entropy to measure the degree of difference betweenthe ears.(4) The core algorithms of many academic paper in the ear recognition area areaccomplished, and we analysis and compare the experiment result with the three algorithms inthis paper respectively to verify the recognition precision and efficiency of the threerecognition algorithms in this paper.