Research On Imperfect Bivine Iris Recognition Based On Combination Of Local Features And Global Features

With the overall improvement of people's consumption level in China,beef products are more popular in daily diet.However,when the contaminated beef cattle are involved into the food supply chain and fail to be tracked effectively,the residents'life will be threatened.As a result,it is of great significance to develop a technique to trace the sources of beef cattle accurately,so that the spread of contaminated beef can be efficiently controlled and the negative effect on human health can be largely reduced.In this thesis,we exploit two recognition methods for imperfect bovine iris image,which can be used to develop the traceability technology.These two methods are developed based on the theories and methods of global and local feature extraction,together with considering the characteristic of bovine iris.More details are as follows:(1)When the quality of bovine iris images is imperfect and the quantity of samples is perfect,this thesis proposes a novel two-dimensional linear discriminant analysis with locality preserving(2DLP-LDA)approach,which covers both within-class and between-class local geometric information.2DLP-LDA re-characterizes the within-class scatter matrix by employing locality preserving projection technique,meanwhile re-characterizes the between-class scatter matrix by introducing a Gaussian weighting function,making the samples from the same class moderately cluster while the samples from different classes properly distribute in the reduced subspace.Encouraging experimental results on SEU bovine iris database show the effectiveness of the proposed approach.(2)When neither the quality nor the quantity of bovine iris images is perfect,this thesis presents a discriminant analysis algorithm for single training sample image recognition based on virtual images and multi-manifolds(?-MDA).?-MDA selects horizontal 2DPCA,vertical 2DLDA,LBP and Gabor feature descriptor to fully extracte the global and local features of images.The proposed method assumes that virtual images come from different manifolds and calculates corresponding low dimensional manifold spaces of different classes.SEU bovine Iris database is used to verify the effectiveness of our method.(3)In order to verify the generalization ability of the proposed two methods,extensive experiments are conducted on FERET and CMU_PIE face databases.The experimental results show that the proposed two methods perform better than comparison methods,which demonstrates the good generalization ability of the proposed methods.