Research On Iris Recognition Based On Support Vector Machine

Traditional statistical theory aims at the asymptotic theory when sample size istend to infinity. However, in many practical cases, samples are limited. Most ofexisting methods based on traditional statistical theory may not work well for thesituation of limited samples. Statistical Learning Theory (SLT) is a new statisticaltheory framework established from finite samples. SLT provides a powerful theoryfundament to solve machine learning problem with small samples. Support VectorMachine (SVM) is a novel powerful machine learning method based on SLT. SVMsolves practical problems such as small samples, nonlinearity, over learning, highdimension and local minima, which exit in most of learning methods, and has highgeneralization. Currently, being the optimal learning theory for small samples, SLTand SVM is attracting more and more researcher and becoming a new active area inthe field of artificial intelligent and machine learning. Compared with conventionalmachine learning methods, SVMs has many advantages. Especially for a giventraining set, both theory and practice prove that SVM has the best generalizingability than any other methods till now. This dissertation will focus on extensionsof SVM including extensions of the concept of kernel function, SVM classificationalgorithms and some application field.To SVM, it should deal with not less than two samples at the same time(that isto optimize their' s corresponding lagrange multiplication),because the problem ofquadratic programming' s equation restriction would not let us optimize only onevariable. The best benefit of minimal optimization is that it can solve the matter ofthe smallest scale of the quadratic programming by the way of analysis, so that wecan avoid the iterative algorithm. Of course, the "minimal optimization" like that isimpossible to insure the result is the last one by the way of optimizing the lagrangemultiplication, but it can make the objective function to stripe forward. Then webegin to optimize other lagrange multiplications until all multiplications are accordwith the KKT term. Thus, there are two matters to solve in this algorithm: one ishow to optimize two variables, the other is to decide which lagrange multiplicationis to optimize first.Because the iris recognition has the features including: it is easy to acquire, itsfeature is apparent, it is unique and it is not easy to fake. Moreover, the irisrecognition has high accuracy and it is long-time steady and difficult to fake, sothey are the generally accepted promising biological recognition technology, andthey have important theoretical and social economy meaning. We make research ofthe iris recognition technology, and their breakthrough is certain to bring greatbenefit for society. In this dissertation, according to the geometric andphysiological characteristic of human iris, we have made some research of irisrecognition system, including auto acquisition, pre-processing of images, edgeslocation, feature extraction and pattern recognition.In the dissertation, by the research of the theory and the algorithm of thesupport vector machine, we combine them with the technology of iris recognition.In order to increase the training speed of SVM, an improved Sequential MinimalOptimization(SMO)learning algorithm is presented. The method of choosingparameter C and ε are dealt with.In the procedure of the experiment, we first collect the iris image through ourlab's Iris recognition System in order to create the database of iris, and then topreprocessing, so that we can get the shape of stripe image after the specification.After that ,we make use of the method of PCA to extract the feature aiming at thedimention reduction, then we get one vector with 40 dimentions in correspondingwith one training sample. At last, we begin the training and the testing, and theaccuracy 94.3% has been achieved.Even if we have obtained much achievement, there are still a series ofproblems need to solved, and then will be the emphasis and direction of our futurework. They include:1,The problem of choosing the parameters. Now, there aresome methods for choosing it, but they are not very accuracy, so the method ofattempting and patch is the main measure. Therefore that is a main problem of thetheory of SVM to study and plan. 2,As the data of iris is added, we must tomulticlassificate.