The Research Of Ear Recognition Based On Computer Vision

Ear recognition is a relatively novel biometrics technology, which pertains to the use of an individual's ear to determine or validate identity. Ear recognition is not only a beneficial supplement for other biometrics technology, but also can be solely used at some identification occasion. The development of ear recognition technology will enrich and perfect the field of biometrics technology, and provide new challenges for the computer vision and pattern recognition.The ear image pre-processing and edge detection are researched in this dissertation. In the ear image edge detection, an improved canny edge detection algorithm based on wavelet transform and median filtering is proposed. Experimental results show that the accuracy of edge detection is improved evidently, and a much better edge detection effect is obtained.Second, the ear feature extraction method is researched, which mainly describes the linear subspace and non-linear manifold learning. The ear feature extraction based on Linear Discriminant Analysis (LDA) algorithm is researched. Because there exists the small sample size problem in the process of actual ear algebraic feature extraction, this dissertation does experiments with the LDA algorithm using PCA to reduce the dimension of the samples. According to the experiment results, the ear features extracted with the algorithms is effective.Then,the ear feature extraction based on non-linear manifold learning is researched. The classical manifold learning methods have Isometric Feature Mapping (Isomap), Local Linear Embedding (LLE) and Laplacian Feature Maps (LE).Through using different classifiter experiments on the different earbases,we discover their characteristics on the application of ear recognition.At last, an improved algorithm is proposed according to the researching of traditional LLE: LLE algorithm based on the geodesic distance. We prove feature equation of these modified LLE algorithms and apply them to ear recognition. Through doing experiment on the different earbases,we verify that they can reach better recognition performance comparing with LLE.