The Research On Recognition And Fusion Methods Based On Multispectral And 2D/3D Palmprint

The palmprint recognition has become an important technology in personal identification because of its advantages such as low resolution, low cost and stable structure features. Compared to face recognition, palmprint is not so easily affected by the expression and illumination. The palmrpint has more features than the fingerprint, so more and more reaserchers have done many works for palmprint recognition.The palmprint features are diversity. For example, there are different features among the different multispectral pamprint images; 2D palmprint images contain many texture information and 3D palmprint images have plenty of depth information. The different palmprint features have complementery and limitations. Thus how to effectively use the palmprint images and extract the palmprint features is an important issue, and is the main researches of this dissertation. The specific research topic and main innovations are as follws:1. For the multispectral palmpirint recognition, how to effectively fuse the plamprint images from different bands is an key issue. Based on this issue, we proposed an fusion strategy based on the image-based linear discriminant analysis(IBLDA) method. The experiemnt results demonstrate that the fusion strategy based IBLDA method can effectively fuse the multimodel biometrics. The experiment results demonstrate that a higher recognition rate can be achieved by using the IBLDA method than using the QPCA and MCPCA method. On the other hand, how to select the optimal bands from the total bands to do fuse is also an key point. In this case, we present a novel spectral band selection procedure to select and fuse the most representative bands to accurately perform palmprint recognition. Experimental results demonstrate that our spectral band selection procedure, not only can improve the total palmprint accuracy rates and also can reduce the time complexity.2. The palmprint images in different multispectral bands exsit redundant informations. How to effectively delete the redundant informations and improve the classification rates become one hot topic of research. The extended general color image discriminant(GCID) algorithm can eliminate the redundant information among the color component and obtain the discriminant information of the components. But using the GCID algorithm can lead to singluar of the inner-class scatter matrix. So we proposed one novel parallel GCID algorithm to solve this problem. Using the parallel GCID algorithm on four channels multispectr al palmprint database, on one side, can well solve the singluar of inner-class scatter matrix, on the other side, can obtain discriminative palmprint features. The experiment results verified the effectiveness of the parallel GCID algorithm, not only can reduce the reduunant information among the multispectral palmprint, but also can avoid the singluar of the inner-scatter matrix.3. 2D palmprint images contain plenty of texture information, but they are easily forged and affect by illuminations. 3D palmprint images contain the depth information of the palm surface and can overcome the shortcomings of 2D palmprint iamges. So it is siginificant to fuse the 2D palmprint and 3D palmprint to obtain higher classification rates. In this section we proposed two fu sion strategies for palmprint recognition. At first, we use the mean curvature image(MCI) of 3D palmprint as features of 3D palmprint, and use the Gabor feature of 2D palmprint images as features of 2D palmprint. Then, we perform the matching score level fusion for fusing the palmprint features and we use the PCA method for feature extraction. We carry out experiments on two palmprint databases and our method achieves satisfied recognition performance. Moreover, taking into account the rotation and translation of the palmprint images, we use the Daisy represention and SIFT feature to extract the direaction information of the palmprint images. The experiment result demonstrates that our method is robust to the rotation and translation of the palmprint images. Compaerd with the coding palmprint recognition method, our method can obtain better performance.4. By reducing the dimensions of palmprint images, w e can obtain the linear sparse representation and improve the effectiveness of palprint fusion. Compared with the conventional linear dimensionality reduction methods, the sparase representtaion learning method can obtain the sparse weight matrix. This case can reduce the time complexity and improve the classification rates at the classification stage. Thus in this paper we joint the sparse learning method with 2,1 norm, and we proposed a robust feature extraction algorithm for palmprint recognition based on 2,1 norm. This algorithm can discover the lower dimensional structure embedded in the high dimensional space and can achieve the structure information of the data points. Also it can effectively obtain the latent structure of the non-linear high dimensional data, which is better for dimensionality reduction and data analysis. The experiment results demonstrate that the proposed method based on 2,1 norm can obtain higher classification rate for palmprint recognition.In summary, this paper manily studied the palmprint recognition method s based on palmprint fusion. The main purpose is extracting more features from the palmprint images by using feature extraction and fusion strategies.The experiments results demonstrate that the feature extraction algorithm and fusion staetgies in this paper can improve the palmprint recognition performance effectively, which have established the foundations of palmpirnt applications.