The Research Of Face And Iris Fusion Recognition At The Feature Level

Multimodal biometric fusion recognition which aims at identification or verification bythe use of multiple physiological characters or human behavioral characters is very hot inrecent years at biometric recognition area. These physiological characters include fingerprint,palm print, face, iris, voice etc., and behavioral characters includes gait, signature,keyboarding and so on. Comparing with unimodal biometric systems, multimodal biometricsystems perform higher accuracy and reliability. That’s because using two or more biometrictraits for person authentication can reduce false accept rates and false reject rates. And at thesame time, it’s more difficult to forge several biometric traits. What’s more people can useother traits if some one trait is ineffective. This paper chooses face and iris as experimentsamples, because these two biometric traits are non-contact acquisition, and can be obtained atthe same time. On the one hand, they can ease user’s vigilance; on the other hand, the systemappears quickly and conveniently.Depending on the information fusion level, multimodal biometric system can be dividedinto four levels: sensor level, feature level, matching level and decision level. At the featurelevel, feature vectors are extracted by related algorithms, then weighted or un-weightedconcatenation are done on these vectors. Comparing with other fusion levels, the feature levelcan retain more information, have more failure tolerance, anti-jamming capacity, systemefficiency optimization. So this paper chooses feature level fusion.First of all, we propose a new iris localization algorithm. Iris localization is important forfusion recognition. Some existed algorithms can not meet time-consuming and accuracy inboth. This paper proposed a combination of Fourier descriptors and the least square methodfor iris localization. When locating the inner boundary of iris, an improved tracking algorithmwas introduced to obtain point-sequences by tracking the edge of the binarization pupil. AfterFourier descriptors transforming, the least square method was finally used to fit the innercircle. When locating the outer boundary, the gray gradient information on the left and rightpart of the iris was appropriately used. Based on this information, the outer boundarypoint-sequences were extracted. Similarly, the outer circle can be fitted by the least squaremethod at last. Experimental results show that the method can overcome the eyelashes andeyelid occlusion, and also improve the localization speed and accuracy. The results can meetthe needs of real-time iris recognition system.Then, we have made a detail study of iris segmentation and face illuminationcompensation. Statistics have done on the CASIA iris database show that90%of the irises are blocked by the eyelash and eyelid at upper and down part, while only4%at left and right. Sowe segment the left and right part instead of the whole iris. Experiment shows that theoptimization segment angle is75°which lead to a better recognition result. Face isvulnerable to the impact of uneven illumination which will lead to negative impact to therecognition results. To address this problem, this paper introduces the homomorphic filteringalgorithm to illumination compensation. Homomorphic filtering is an image processingmethod that combines frequency filtering and gray-scale transformation. Based on imageillumination and reflection models, the method is aiming to improve the image quality bycompressing the scope of the brightness and enhancing the image contrast. Experiment resultsshow that after illumination compensation, not only the quality of the uneven illuminationface samples, but also the whole system has good improvement.This paper have made a detail derivation of the PCA, LDA and LPP algorithms at thefeature extracting stage and get the conclusion that PCA and LDA algorithms are consistent inmathematical principal. PCA and LDA mainly concern about the global information of theimage, and the subspace of which characterize the difference between any two images. Thus,we introduce the LPP algorithm which concerns about the local details. Experiment resultsshow that the recognition rate is higher than the other two by using LPP algorithm.At last, this paper proposed an improved fusion algorithm. Different from the traditionalfusion algorithms, this algorithm is based on the property features, not the original features.The property features are a set of features produced on the original features. They have twotypes of distribution, random distribution and normal distribution. At the fusion stage, weconcatenate the property features that marked the same tag. Then we use the KNN algorithmto evaluate the ultimate result. Experiment results show that by using property features thesystem performs better.In summary, this paper has researched the fusion recognition of face and iris at thefeature level. Proposed a new iris localization algorithm, have done a lot of experiments atface illumination compensation and iris segmentation. For PCA, LDA and LPP algorithms,we have deep derivation and comparison. Proposed an improved fusion algorithm based onproperty features. Large quantities of experiments show that the proposed algorithm canfurther improve the performance of the system.