Improved Gabor Filter Based Multi-modal Features Fusion Technology

Biometric identification technology develops with modern society. By the development of human society, the increasing of network and information put forward new requirement for the biological feature identification technology. Gabor filter contains the strong character of multi-direction and multi-scale, and feature extraction method with multi-direction and multi-scale has been widely used in face recognition, palmprint recognition, fingerprint recognition, iris recognition and other biometric recognition technology. However, the traditional Gabor filter was not good to extract the image features of local bending, and it does not have the very good ability for characterization of bending such as facial nose, eyes and other areas. This paper improved the traditional Gabor kernel functions and combined with feature fusion and multi-modal learning technology, as follows:Firstly, in this paper, we improve the traditional Gabor filter. The improved Gabor filter can not only extracte the multi-direction and multi-scale information in the image, but also has a very strong expression ability of bending at the edge of the image, so the classification performance achieve better.Secondly, in order to fully utilize multi-feature information in multiple parameters of Gabor filter, and to lay a foundation for each individual condition of multi-modal structure, this paper will utilize image with different froms under the different parameters of feature fusion to obtain single mode with lower feature dimension and small redundancy. The fusion feature combined with multivariate correlation method is proposed in this paper to carry out multi direction characteristic information of the Gabor filter, and can effectively solve the feature fusion problems: high dimension or applicable situation difficult to promote.Finally, this paper introduces multi-modal discriminant learning analysis which combined with multi characteristics of Gabor filter, and it can achieve better recognition. In the real world, the image can often be described from multiple-modal, so it has a collection of samples or characteristics. Now, more and more applications need to identify the data, and these data usually come from different modal. Therefore, the recognition problem based on multi-mode is popular and has extensive application foreground and research topics. However, the traditional feature extraction method, such as LDA, solve problem is from a single mode of problem to, so it can not be used in multi-modal identification problem directly. In recent years,much multi-modal learning methods focused on mining discriminant features within multi-view but neglect of differential information effectively between multi-modal. In the proposed multi-modal learning framework, samples from every mode are mapping in one common discriminant subspace, and in this subspace, samples from the same and different modes in the same classs are gathering, heterogeneous samples are apart from each other.In this paper, the experimental verification is carried out in the HK PloyU palmprint database, FRGC face database and AR face database, and final experimental results show the feasibility of our innovative work.