| With the rapid development of information society, the traditional authentication methods(such as ID, passwords, keys, etc) have been unable to meet the requirements of information security. While biometrics technologies(such as fingerprint, iris, face, ear,etc)have become a new authentication method, because of its advantages of universality,uniqueness and stability, etc. As a new biological characteristics in the field of biometrics, ear recognition has been the research hot spot because of its unique physiological position and structure characteristic. At present, the selection of the appropriate feature extraction and classification method is most important in the process of ear recognition. Because Gabor wavelet transform can well capture image local information of the time domain and frequency domain in the space, it is widely used to extract the image feature.While there is the defect of too high dimension of feature extraction. Linear discriminant analysis(LDA) is a common and effective method of feature extraction and classification. It can achieve the goal of maximizing the expanse of inter class variances and minimizing the expanse of intra class variances,moreover, it can obtain the optimal feature projection space. But there is LDA small sample size problem which is caused by singularity of intra class scatter matrix.To solve the problem of small sample size of ear recognition, ear recognition based on Gabor features and improved linear discriminant analysis(ILDA) is deeply studied in the paper, the main work is as follows:Firstly, the local feature extraction method based on Gabor is introduced, and the principle of two-dimensional Gabor wavelet transform is comprehensively analyzed. Then according to the corresponding relation between image clarity and energy character of the local Gabor filter in the composition of five dimensions and eight directions, some local Gabor filters are selected to construct the local Gabor filter bank. Therefore the impact caused by high feature dimension of Gabor wavelet transform is reduced.Secondly, feature extraction and dimensionality reduction algorithm based on the linear subspace is expounded, then the fundamental principles and advantages and disadvantages are deeply studied about the traditional principal component analysis(PCA) and the linear discriminant analysis(LDA). In addition, this part focuses on the causes,the impact and the previous solution of the small sample size problem of LDA algorithm.Finally, on the basis of local Gabor feature extraction, the improved LDA(ILDA) is proposed to feature extraction and feature dimension reduction of Gabor feature again. Basedon in-depth study of the fundamental principles and defect of the traditional LDA, the appropriate classification criterion function is used, then the intra class scatter matrix is redefined in terms of convex hull algorithm,which can ensure that the data class information after feature dimension reduction is not lost.And the method of dimension reduction before projection is adopted, which reduce the complexity of the algorithm.Moreover, feature after dimension reduction is projected to null space and non-null space of intra class scatter matrix,which can ensure the integrity of the classification and identification information.Minimum distance classifier is used to classify and identify ear images, the experiment results show that ear recognition based on Gabor feature and ILDA can effectively avoid the influence of the human ear small sample problem, and obtain higher recognition accuracy rate. |
