Iris Recognition Method Based On Independent Component Analysis

Based on the iris biological recognition technology is a new person verification and identification technology, it has the broad application prospect in the finance, national defense, information security and other fields.Because of iris own uniqueness, the stability, may gather, merits and so on non-offensive, the iris recognition has become increasingly current biometric a research focus.The iris recognition system is composed of the iris image gathering installment and the iris recognition software , in the iris recognition software aspect, this article key to has conducted the research based on the ICA iris imagery processing algorithm, primary coverage next four points:(1) Before in the characteristic extraction iris preprocessing, first defers to this topic-based group already the traditional method which has carried on the localization to the iris image, the normalization, revolves, image intensification.In view of the ICA algorithm and iris own characteristic, to pretreated proposed three different improvement methods:Based on partial scrap image improvement method: Divides into eight partial scraps after the entire 64×512 normalized image separately to make the ICA characteristic extraction again to each partial scrap, through to eight regions the eyelid, the eyelid, the illumination influence is different, we have established the different weight to the eight regions.The piecemeal improvement may code separately each kind of iris partial information, this caused the image change to limit in each partial scrap region.The experimental result indicated made the ICA transformation in the partial space to obtain the better recognition effect.The correct recognition rate enhanced 2%.Based on the interest region the improvement method: This is the different method pretreats in the existing iris preprocessing.According to the ROI (region of interest) to divide four regions, selects about the iris region each 64×64 and about each 32×32 small region, these four regions have not changed the iris primary information, and contains the enough texture information to use in the characteristic extraction, reduced the computation order of complexity enormously. According to the area, the degree of disturbance, was four regions has been established the different weight.The experimental result indicated, although this method on drops 1.66% slightly in the recognition rate, but its training time decay near 10 times, the recognition time reduced near 6 times.Based on wavelet transformation two characteristic extraction improvement method: First uses the Daubechies wavelet base to 64×512 the normalized iris image carried on 2 level of tower wavelets to decompose, has selected low-frequency band LL2 and intermediate band LH2, HL2, LH1 and HL1 five innertube regions carries on the ICA processing separately, and according to the wavelet coefficient threshold value, has carried on the filter to the intermediate band, effective elimination image noise.The experimental result indicated that, on this criterion frequency domain two characteristic extraction method, its training time reduced 24s, the recognition rate enhanced 3.33%.Compares the experiment and analyzes, we choose the third improvement method to take the preprocessing method.(2) In the ICA characteristic extraction aspect, according to had the article introduction, the design realizes has taken the iris platform using negentropy fixed point algorithm the core algorithm, and through analyzes PCA and the ICA similarities and differences, improved the original algorithm with the PCA/ICA method.The ICA method extraction is based on the higher order mutually the independent characteristic, it seeks both the statistical independence and the non-Gauss's ingredient.This article introduced in detail the ICA theory and the algorithm. We use the suited matrix budget---Matlab to realize the FastICA algorithm, sped up the ICA independent base vector extraction speed.ICA is PCA from two step statistical analysis to the higher order statistical analysis development.PCA the original high dimensional data will project to lowers dimensional data space, and retention data main information.Compares with ICA, PCA is based on the signal two step statistical property analysis method, it separates between the signal is non-correlated; But ICA is based on the signal higher order statistical property analysis method.In ICA"the sphering"front, first carries on the PCA processing, may remove the inferior ingredient effectively, reduces the dimension, the elimination noise, obtains independent base vector better expression iris characteristic.The experimental result indicated that, uses PCA/ICA, its recognition rate enhanced 4.31% equally, its recognition time reduced 19.4ms equally.(3) In the later characteristic match period, this article introduced the cosine measure, the BP neural networks, and the support vector machines three kinds of sorters, and use in the iris recognition to realize.Compares the experiment and analyzes, the support vector machines method has the highest recognition rate.Cosine measure sorter: Take the vector included angle cosine as the measure, in the commonly used most close neighbor separator, was most suits ICA the sorter, in this article ahead of time experimental majority of use this separatorBP neural network: By the input level, the concealment level and the output level constitution, this article uses 3 network architectures, the independent base vector integer which ICA withdraws 30 is the input level, the category which in the training regulations contains several 50 is the output level, the concealed level supposes is 40.In the experiment its recognition rate has achieved 96.5%.support vector machines: Lowers vector's most superior classified planoid through the solution, causes the classified risk upper boundary to be smallest, it essentially is solves one to have the massive restraint two plan question.This article uses the libsvm software package, in the experiment its recognition rate is highest, is 98.5%.(4) Finally , through the massive experiment comparative analysis, we designs the final experimental platform core algorithm uses PCA/ICA the fast fixed point algorithm, the pretreatment elects for two characteristic extractions which decomposes based on wavelets, the final sorter selects support vector machines, designed according to this on matlab has realized based on the ICA iris recognition simulation software.The experimental effect is satisfying.In summary, this article research based on the ICA iris recognition methods, through designing and choice the methods of iris processing, characteristic extracting,and characteristic matching, finally designs the software platform.But because of the rush and my limited level, the software system also has many places need to improve and to consummate. Such as the iris illumination model establishment, the big sample storehouse experiment, and the ICA algorithm optimization and the most superior independent base vector extraction, these will all need the following endeavor, to strive for realized the iris product marketability.