Font Size: a A A

Statistical Learning Theory-based Iris Recognition Study

Posted on:2006-01-18Degree:DoctorType:Dissertation
Country:ChinaCandidate:H Y GuFull Text:PDF
GTID:1118360182957621Subject:Computer applications
Abstract/Summary:PDF Full Text Request
This dissertation studies iris recognition based on statistical learning theory. Iris recognition possesses great potential in biometric security applications. It employs the unique texture patterns and features of the human iris to verify the identity of an individual by his or her iris of the eye, which is the colored ring surrounding the pupil. Based on the recent advancements in statistical learning theory and iris recognition, we further our study on iris feature extraction, iris feature fusion and selection mechanism, classifier design, and iris recognition system design.Firstly the paper summarizes current iris recognition methods. It analyzes their merit and demerit in presenting and classifying the characteristics of irises. Based on the analyses, we make the contributions described as the followings:The first contribution of this paper is iris feature extraction. We studied the variation details of iris texture and found the self-similarity. Consequently, the variation fractal dimension is proposed in chapter three. It is a new fractal dimension derived from traditional fractal geometry to better represent the variation details in iris texture. It works well even if some regions of iris are obscured or occluded by eyelashes and eyelids. The variation fractal dimension can also be applied to other image processing applications besides iris recognition;The second contribution is iris feature fusion and selection mechanism. Usually there is only one kind of features considered in former iris recognition methods. Wavelet transform and Gabor transform are most often chosen to get the features. While iris textures are so complex that it is inadequate to represent iris by only one transform in some cases. In chapter four we introduce steerable pyramid transform to get the orientation information of iris features and extract features from the subbands derived from a series filter bank. To get more precise fusion features by clearing redundant features, we propose a feature selection method based on multi-object evolution algorithm.The third contribution is iris classifier construction. The distance methods (including Euclidean distance and Hamming distance) and exclusive-OR are widely used in former iris recognition methods as the classifiers. To get higher correct recognition rate, we introduce learning mechanism to iris recognition in chapter five. Meanwhile, in the real world applications, the costs of False Rejection and False Acceptance are different and they should be treated differently. The FalseAcceptance should be punished greater than False Rejection in the applications with higher security demands. Therefore, we introduce and propose Non-symmetrical Support Vector Machine (NSVM) as our iris classifiers to meet the variant security requirements.The fourth contribution is IrisPassport system. A novel iris recognition system, namely IrisPassport, is proposed and built based on the theories and algorithms introduced above. We tested the system on public iris database CASIA and compared the result with other iris recognition systems. We introduce the structure and characteristics of IrisPassport system in chapter six.Experiment results show that the performance of IrisPassport can meet the requirements of real applications. We summarize all the works of the whole paper in chapter seven and indicate the parts that need further research in our next step.
Keywords/Search Tags:biometrics, iris recognition, variation fractal dimension, steerable pyramid, non-symmetrical support vector machine, genetic algorithm, statistical learning
PDF Full Text Request
Related items