Study On Preprocessing Approach And Feature Extraction In Palmprint Recognition

Biometrics combines biology technology and information technology to exploitphysical features in human body such as fingerprint, iris, face and palmprint or behavioral features such as gait and signature to identify a person so as to replaceor enhance traditional personal identification methods. This new technologycan be applied to homeland security, social welfare, financial and e-commerce. Biometric is becoming the highest potential technology in this era.Palmprint recognition exploits the effective information on our palm to separatedifferent persons. Different palms have totally different palmprints. In other words, palmprint is unique. Therefore, it can be used for personal identification. The advantages of palmprint recognition are direct, friendly, convenient, unique and apply widly.This dissertation concentrates on the research of the key technologies and algorithms in palmprint identification system. According to the detail research of other biometrics, and considering the characteristics of palmprints, we propose a series of efficient palmprint preprocessing, feature extraction algorithms. In detail, this dissertation concentrates on the research of the following key.Firstly, in the aspect of the preprocess of palmprint image,we focused on the essential technologies of palmprint identification of color adjustment,we proposeda statistical method to filter out splash, Based on this technologies, this paper had done some preprocessing to the palmprint images. After translated the original image into binary image, found two key positioning points in the palmp rint image using automatic detectingmethod. Then the palmprint image was rotated according to the two key positioning points. Finally, a sub-image of the rotated palmprint image was cut for further abstracting of palmprint features.Experimental data obtained through extensive simulations shows that our scheme is efficient.The progressof palmprint identification system as follows:(1) Filter out splash;The definition of HSB data of each sample is a vector V_i(h_i,s_i,b_i),Similar sampling by the typical characteristic of a vector V_mean(h_mean, s_mean, b_mean), The various components of the vector is all samples of the component statistics meanin HSB space, such as h_mean = 1/n sum from i=1 to n h_i. Calculating the similarity (?)_i betweensampling and V_mean with the formula (1) .Then calculated according to a group of similarity (?)_i to calculate the average difference similarityβ.Whereβis a similarity to the decision-making limits of tolerance and that the final determination Similarityγdecision threshold:γ=1-β(3)Each pixel image T_i(h_i,s_i,b_i),for the T_i and V_mean similarity with (7.15), if (?)_γ≥γ,then the point of splash, otherwise the normal colour .(2)Palmprint images will be converted to binary image;(3)According to the values of the binary image, a palm-edge detection to determine the location;(4)According to the two anchor points and their midpoint,make a Cartesian coordinates;(5) Rotating original palmprint image to a new place in the Cartesian coordinatesystem;(6)A sub-image of the rotated palmprint image was cut.Sencondly , a new image feature extraction and recognition method based on image principal component analysis (I-IMPCA) was proposed. The method constructed covariance matrices using original image matrices directly and provideda sequentially optimal image compression mechanism. I-IMPCA performed image principal component analysis (IMPCA) twice: one was in horizontal directionand the other in vertical direction. Also, I-IMPCA suggested a feature selection strategy to select the most discriminative features. K-nearest neighborhood(KNN) algorithm was used to construct classifiers.IMPCA:Let U is n dimensional vector ,A is m×n dimensional image,A will be through a linear projection to transform in U:B = AU (4)Where B is a projection eigenvector of image A,projection eigenvector by the spread of the situation to decide on the best projection axis, U,Using the following criteria:J(U) = trS_u (5)Where S_u is a class spread matrix of the projection eigenvector B ,tr is the trace of class spread matrix . Maximize the criteria is to seek a axis projection of U, we have the largest spread between categories after the projection .WhereE is mathematical expectation,G_t is the inter-spread matrix of image .There are M Samples,A_j(j = 1,2,..., M) is NO. j m×n dimensional image matrix. The average of them is A =1/M sum from j=1 to M A_j.SOThrough (5), (7), (8),we have(9) is called the guidelines of generalized inter-spread.Vector U_opt is called the best projection axis, the physical meaning is, the value of the spread between classes of the eigenvectors in U_opt axis is largest.In fact,the best projection is the largest eigenvalue of the corresponding eigenvector of the inter-axis matrix spread G(t).Generally speaking, the more types of samples circumstances, there is only one optimal projection axis is not enough, and the need to find a group of projectionaxis u₁,u₂, ...,u_p satisfying the following conditions:In fact , orthogonal projection axis u₁, u₂,..., u_p are the eigenvector corresponding to largest eigenvalue of G_t before p.For a given palmprint image A,we have component projection characteristics:b_k = Au_k; k = 1,2, ...,p (11)So m×p dimensional image matrix B = [b₁, b₂,..., b_p] is the Eigenmatrix of image A.I-IMPCA:performed image principal component analysis IMPCA twice: one was in horizontal direction and the other in vertical direction. Also, I-IMPCA suggested a feature selection strategy to select the most discriminative features. K-nearest neighborhood (KNN) algorithm was used to construct classifiers. Image principal component analysis IMPCA was in vertical direction is looks as image matrix to buy in horizontal direction.Firstly,get the eigenmatrix of image B,then performed IMPCA to B^T ,find the best projection axis V, extract feature C^T- B^TV.Similarly, the definition of the inter-spread matrix H_t and mean matrix B^T as follow:we haveFind the projection axis V = [v₁,v₂,...,v_q] which make J(V) = V^TH_tV get max. In fact,v₁,v₂,...,v_q are the eigenvector corresponding to largest eigenvalue before p of H_t. SoWhere C is q×p dimensional image matrix,.GeneralLy,get p, q far less than m, n, so the dimension of C is far less than m×p dimensional of B and m×n dimensional of A.