The Research Of Iris Feature Representation Based On The Method Of Statistic

As a kind of identity recognition method, biological recognitiontechnology is more and more emphasized recently. Biological recognitiontechnology is a kind of pattern recognition that combines informationtechnology and biology technology to recognize a person.Because the iris recognition has the following features: it is easy toacquire, its feature is apparent, it is unique and it is not easy to be faked.Moreover, the iris recognition has high accuracy and it is long-timepermanent and difficult to fake, so it is the generally accepted promisingbiological recognition technology. Many methods have been proposed andimproved to be effective, some of which have been developed into appliedsystems.An iris recognition system always consists of the following modules:image auto acquiring, image preprocessing, feature extraction andclassification. In this paper, we have made some research of iris featureextraction and classification according to the geometric and physiologicalcharacteristic of human iris, several algorithms on iris feature extraction andclassification are proposed based on the analysis of traditional iris recognitionsystem. We bring forwards a novel algorithm of iris representation based onthe method of statistic. The experimental results show that the method hashigh precision and the recognition results are satisfactory.The main jobs we do in the paper are as followed:1) Analysis the composition of the iris system, especially discuss theways for the basic algorithm of iris feature extraction and classification;2) First, we locate the iris edges and iris normalization,then the wavelettransform was used to obtain the stable blow frequency sub-band of theimage in relatively low dimensions.The wavelet transform is a kind of local region transform of time andfrequency, and it can validly extract the information from signal. It can makemulti-scale analysis of the function and signal by stretch and translationfunction, and it can solve many difficult problems that can't be solved byFourier transform, so it is called "Mathematical Microscope". As the result,we extract the iris feature by wavelet transform;3) Up to now, many methods of feature extraction have been proposed.Most works on the iris feature extraction have been done by linear transform(such as Gabor and Wavelets).However, one disadvantage that the lineartransforms inherently possess is that the base vectors are fixed independentlyof any data. Therefore, conventional methods for iris feature extractionshould select the parameters (e.g., spatial location, orientation, and frequency)for fixed bases. We put forwards a novel algorithm of iris representationbased on the method of statistic. These are Principal Component Analysis(PCA), Fisher's Liner Analysis (FLD) and Independent Component Analysis(ICA).PCA is a very effective feature extracting method in pattern recognition.The basic idea of the method is that high dimensional space of iris images istransformed to a group of new orthonormal bases by Karhunen-Loeve (K-L)transform, of which low dimensional and linear space is made up. In thispaper, the total scattering matrix of the set of training samples is used to formthe production matrix and its feature vectors are used as iris space. Becausethe bigger eigenvalue related to the eigenvector, the greater its contributionrate to the iris images. The eigenvectors related to small eigenvalues arediscarded in this paper in order to compress the dimension of iris featuresmore effectively. Consequently, every iris image is projected to the space thatis extended by the eigenvectors related to bigger eigenvalues, and the vectorsof projected coefficients are just the extracted iris feature.Pure PCA method can obtain the best features in a representation sense,but these features are not very suitable for classification purposes. FLDmethod makes good use of label information to get the best features forclassification purposes. However, the dimensionality of the original irisimage is too high to use FLD method directly. Thus, the paper adoptedPCA+FLD algorithm: Fisher which first reduced the dimension of the irisimage via PCA, and then use FLD to obtain the best projection direction forclassification purposes. Experiment results show that the PCA+FLDalgorithm has better performance in iris recognition.We have researched how to extract iris feature by ICA method,contrasted which feature is better to represent iris. The speed of training ICAiris base is accelerated by FastICA method, at the same time, we select betteriris base using some methods, thus can use as few and better base as possibleto represent iris, and reduce the dimension of iris feature space in effect.After extracting the feature, the recognition rates of the three methodshave been compared using the minimum distance classifier;4) An improved back propagation (BP) algorithm was introduced totrain the neural network for recognition. Comparing with the traditional BPalgorithm, the improving algorithm can promote the learning rate andcredibility of the algorithm. It adopts learning rate adaptive adjustable andmomentum method strategy. Adaptive adjustable strategy can effectivelyavoid the network out of the local minimum and into the global minimum,while the momentum method lower the sensitivity to details of error curvedsurface and efficiently restrains from falling into the local minimum. Finally,the features of the images extracted by PCA were sent into the improved BPneural network for classification and recognition.The experiment results indicate there are more robustness and accuracyin our system. This paper researches the theories and technologies in irisidentification field and provides a feasible foundation of theories andtechnologies for the research of auto iris identification system deeply.