Reserch On Iris Feature Descriptor And Recongniton Algorithms

With the development of information society and the increasement of socialsecurity requirements, the personal identitification technology based on biometrichas been received more and more attention. Iris recognition technology has thephysiological advantages of stable, unique, non-invasive, so it has the higherreliability and higher recognition rate. Meanwhile, it is also one of the typical andcomplex computer vision and pattern recognition problems, which makes it becomea research hotspot of biometrics recognition, and it has very important value ofacademic research. This dissertation focuses on the study of iris feature extractingand matching in terms of the texture characteristics of iris. The main work andcontributions are provided as follows:Aiming at the inherent cyclic structure of iris, the multi-scale iris localizationalgorithm is proposed. Firstly, the light spots are detected through the spot detectionoperator, and then filled with the bilinear interpolation method. Secondly, the iriscoarse localization is performed in order to detect the pupil’s approximate location,which can be used to cut iris image. Thirdly, the iris fine localization is implementedcombined with the Daugman’ calculus operator and Wildes’ two-step method. Finally,the iris is normalized by the rubber band model proposed by Daugman. Theexperimental results show that this method can locate iris fastly and accurately. Because the EMD and the local mean decomposition (LMD) cannot takeaccount of decomposition rate and smallest error, a fast and effective improvedmethod based on the PCHIP-LMD is proposed. According to the distributioncharacteristics of iris texture, the improved method is applied to decompose the irisimage line by line with the aim of generating the different scale component image.Useful component images for the iris recognition are found, and they arethresholded to get the feature image of iris. This method is well capable ofeliminating the high faequency noise, and extracting binary image features.Compared with EMD and LMD methods, this method has a faster speed, higherrecognition rate and better robustness.Because the local binary pattern (LBP) and the center-symmetric LBP (CS-LBP)of the iris recognition have the problems of high dimension and sensitive to noise,an effective improved method based on the statistical characteristics CS-LBP(SCCS-LBP) is proposed. Firstly, CS-LBP is applied to encode the normalized irisimage based on the distribution characteristics of iris texture. Secondly, in order tofurther reduce the feature dimension, the statistical characteristics of the encodedimage are calculated, and then the statistical results are thresholded to obtain thefeature image of iris. Compared with LBP and CS-LBP methods, this method hasthe advantages of lower dimension, higher recognition and better robustness.Based on the study of Gabor filter, the weighted Hamming distance usinganalytic hierarchy process (AHP) is proposed. Firstly, the feature image is generatedthrough different scales and directions of the odd symmetric Gabor filter. Secondly,the recognition correct and error rate are acquired by the least Hamming distance.Subsequently, the appropriate scale and direction are selected so as to calculate theweighted Hamming distance, and the weights are computed by AHP method. Finally,the weighted distance is applied to implement iris recognition. The experimental results show that this method has better recognition performance and robustnessthan other Gabor methods.