Study On Human Ear Recognition By Algebraic Methods

Human ear recognition is a biometrics recognition technology which sprang up in the 1990s. It has drawn more and more attention because of its high reliability and stability in the academic field. This dissertation analyses the problems of human ear recognition based on currently existed method. This dissertation studies the extraction and matching methods in ear recognition in-depth using the ear databases of USTB. In detail, the main jobs and contributions are as follows:(1) Aiming at the problem of rotating distortion problem to be recognized, the ear recognition algorithm based on local gray scale grad feature point is studied. First, the ear gray-scale image is divided into several sub images, and then these sub images are equally divided into several sub regions. The location information in sub region of which contains the maximum gray-scale sum of all points in this region will be used to build the feature matrix of the ear image. At last, the recognition will be realized by way of matching the feature matrix in space directly. The experiment results indicate that the algorithm has more strong robustness for the rotating distortion ear images.(2) Due to illumination and pose factors, the ear image space is nonlinear. It will be approximated as linear in regular recognition method, and it will affect the recognitioon accuracy. Aiming at the nonlinear structure hidden in the image, the ear recognition method based on generalized discriminant analysis (GDA) is studied. The recognition accuracy of GDA increases 4.2% compared with Fisherear. At the same time, the kernel parameter selection is studied.(3) Aiming at the requirement of the storage space and speed, a low resolution ear recognition method is studied. We first break up the human ear images into different layers by using Gaussian pyramid, and then extract the features of each layer based on Generalized Discriminant Analysis (GDA), finally, we calculate the cosine distance between the test samples and make classified recognition by the threshold value. The experimental results show that the system recognition performance is the best when the human ear image reduces to 36 x 24, the recognition accuracy achieve 99.41%, and the recognition time is 23.6ms, so the requirements for a real-time biometric identification system are met.(4) The multi-biometrics recognition method of ear and face fusion is studied. The experimental results show that the recognition accuracy increases 12.5% compared with single face recognition. The multi-biometrics recognition method of ear and face effectively improves the recognition accuracy of system, and expands the system application scope and adjustability.(5) We accomplished the ear recognition system on-line based on the kernel algorithm of local gray scale grad feature point, firstly, ear images were located and cropped based on Adaboost algorithm, then the ear recognition will be realized by way of matching the feature matrix in space directly, and ear classification by threshold to give recognition effect, finally satisfactory effects were obtained in the laboratory.