One-Sample Three-dimensional Ear Recognition

Ear recognition is one of the most common biometric recognition technologies.Compare with the face,fingerprint and other widely used recognition methods,people have lower sensitivity and higher acceptability to ear acquisition.Robust ear recognition system has great application prospects of many aspects,such as access control management,entry-exit management,law enforcement and criminal investigation.After years of academic research,three-dimensional ear recognition technology has made great progress.However,while the academia focuses on how to improve the recognition rate,the research on some key problems of practical application is still insufficient.Among them,the one sample per person(OSPP)problem is one of the problems often encountered with practical applications.In the case of OSPP,each individual to be identified has only gallery sample of one information source.However,the information contained in the gallery sample is difficult to cover the information contained in the probe sample obtained under uncontrolled scenes completely.Especially,the ear range images normally contain occlusion and pose changes in uncontrolled scenes.How to recognize ear in this case is a challeng problem that must be addressed in the application of OSPP three-dimensional ear recognition.To address the aforementioned problem,this dissertation carries out a research into the following three parts:1)Three-dimensional ear recognition combining local and holistic featuresTo reduce the intraclass variation by reducing the difference between the surface structure of the gallery and probe ear range images from the same individual,an evolutionary game theory based three-dimensional ear recognition method combining local and holistic features(EGTLHF)is proposed.Firstly,in the framework of evolutionary game theory,a two-step game method that can aligning probe and gallery ear range images and extract the optimal corresponding range images(surrounded by corresponding voxels orthogonally projected to the same region at the bottom of the ear range image bounding box)from them is proposed for matching local feature.Then,this local feature matching method is combined with a global feature matching method based on voxelization technology to recognize ear.Experimental results show that EGTLHF method achieves high recognition rate under partial occlusion and pose variation.2)Three-dimensional ear recognition combining local,block joint structure and holistic featuresTo identify the range image more effectively by using the relative pose information between the surface patches,a three-dimensional ear recognition method combining local,block joint structure and holistic features(LBJSHF)is proposed.Firstly,a block joint structural feature descriptor:a pairwise surface patch cropped using a hemisphere cut-structured histogram of an indexed shape descriptor(PSPHIS),which can represent the neighborhood information of two keypoints by dividing the local surface patches of the two keypoints with the plane perpendicular to and passing through the end point of the line segment connecting the two keypoints,is proposed.Then,the EGTLHF method is optimized by reducing the earning of the candidate strategy pairs that make the PSPHIS descriptor mismatched in the two-step game method.The experimental results show that PSPHIS descriptor can identify the relative pose between surface patches and is not easily affected by the variation of ear pose,and LBJSHF method achieves higher recognition rate than EGTLHF method and the existing method in ear recognition task under partial occlusion and pose variation.3)Three-dimensional ear recognition based on constrained corresponding keypoint screeningTo effectively and efficiently match features from the range image by reducing the range of searching matching features robustly,a constrained corresponding keypoint screening based LBJSHF method(CCKSL)is proposed.Firstly,the optimal layout is classified according to the number of visible and invisible ear landmarks.Then,by establishing the corresponding relationship between the keypoints from ear range images and regions(regional cluster)on the directional rectangle that roughly faces the image for each class of optimal layout,the corresponding keypoints are obtained.Finally,ear recognition is performed by combining corresponding keypoints and LBJSHF method.The experimental results show that CCKSL method can not only achieve high computational efficiency and recognition rate,but also use ear size and physiological position information to reduce the range of searching matching features robustly.The study of this dissertation is of significance in addressing the OSPP problem in three-dimensional ear recognition,and it is valuable for widely used ear recognition methods and other biometrics research in similar cases.