Research On Key Technology Of 3D And Multi-view Human Ear Biometrics

Human biometrics identifies the identity of an individual from colony by utilizing individuals'biological characteristics, including human physiological and behavioral characteristics. So far, diversified biometrics technologies such as face, iris, fingerprint, palm print which are based on physiology and gait, handwriting which are based on behavior have been researched and used in some degree. Recently, ear recognition becomes a newly emerging biometrics trend. Ear is not affected by facial expressions, skin color, and cosmetics. The appearance of the auricle is nearly invariant by aging. At the same time, ear recognition can work in a non-invasive way which is easily accepted. In addition, on the ear itself, ear shape holds abundant structure features and special position features. And according to statistics, ear still holds uniqueness even for 10,000 people of large sample data set. All these merits make ear biometrics be a new promising technology.Currently, ear recognition works in oversea and domestic concentrated on the ways of 2D front view image and 3D range image. The appearance of ear is so sensitive to pose variations that if test samples hold different pose from the training samples, existing 2D ear recognition methods will lose its function. Moreover, the current methods of 3D ear recognition are offline and need high cost laser range device to capture 3D ear data. These methods not only need high quality ear range image, but a well registrated color ear image for additional information.Different from previous works, this work focuses on multi-view 2D ear recognition and low cost 3D ear point clouds recognition. At the same time, based on two views, ear shape feature extraction and 3D ear reconstruction are made attempts in this work. The research content of this work involves many aspects of 2D and 3D ear recognition: ear shape feature extraction, 2D multi-view ear recognition, 3D ear reconstruction based on multi-view geometry, 3D ear registration, 3D ear recognition using strip laser scan point clouds models,ear sampling and recognition device.The main contributions of this work are the following:Provide multi-view ear shape feature extraction method. Our methods not only extract the geometry shape features and affine invariant high-order moment features by calculating Tchebichef coefficients from the front view of ear, but also extract ear shape features from the backside view. Our methods overcome the deficiency of existing methods that extract ear geometry features only from the front view.Provide a new multi-view ear recognition method which is based on null kernel discriminate analysis B-Spline pose manifold. We first adopt null subspace kernel discriminate (NKDA) method to extract multi-view ear non-linear features and form a discriminative space. Then in this space, we utilize B-Spline interpolation to construct all subjects'pose manifolds. Thus ear identity can be recognized by determining the minimum project distance from test sample points to pose manifolds.Made some attempts on automatic and semi-automatic 3D ear reconstruction based on multi-view ear images. Use corner detection algorithm to get ear features, discuss the problem of using random sample consensus (RANSAC) algorithm to solve fundamental matrix for automatic ear feature matching. Match ear features according to auricle shape semi automatically. Epipolar correction is done by using camera calibration information and 3D ear shape is reconstructed after density matching. Reconstructed 3D ear model can be made texture mapping.Present an improved Mesh PCA method for 3D ear model pose normalization. And also provide a set of neural network 3D ear model registration strategy based on improved Mesh PCA. Labor why traditional Mesh PCA is uncertainy and present an improved method. Based on this, introduce neural network to determine the rigid transformation matrix adaptively.Present a new 3D ear strip point clouds matching method based on improved ICP registration and local surface reconstruction. Utilizing improved ICP method to do registration iteratively, and calculating root mean square (RMS) distance after registration. Then, we reconstruct local surface using full quadric surface form after registration and calculate the algebra distance from corresponding points to this local surface. Finally, combine the RMS distance and the algebra distance to form feature vector for 3D matching.Present a new 3D ear recognition method based on slice curves of point clouds model. Utilize a group of planes which are parallel to principle axis to cut 3D ear model to get slice curves. Thus the problem of 3D shape matching turns to be the one of 2D slice curves matching. Calculate the slice curves curvature strings and cross angle chain code strings. Utilize the corresponding points distance difference of longest common curvature substring and the angle difference of cross angle chain code strings to form feature vector for 3D matching.Based on statistical significance and samples matching scores distribution, according to small data set recognition performance, evaluate the accuracy of verification performance of presented ear recognition method and predict the cumulative matching characteristic curves in real world large data set.Adopt cheap strip laser, servo motor, and color camera to construct strip laser 3D ear model acquisition device。Reconstruct 3D ear shape by triangular principles. Do 3D strip point clouds model preprocessing task. With OpenCV and OpenGL, develop 3D ear recognition software which is based on improved ICP and 3D local surface reconstruction method.