Development of Multirate Filter – Based Region Features for Iris Identification

Abstract

The emergence of biometric system is seen as the next-generation technological solution in strengthening the social and national security. The evolution of biometrics has shifted the paradigm of authentication from classical token and knowledge-based systems to physiological and behavioral trait based systems. R & D on iris biometrics, in last one decade, has established it as one of the most promising traits. Even though, iris biometric takes high resolution near-infrared (NIR) images as input, its authentication accuracy is very commendable. Its performance is often influenced by the presence of noise, database size, and feature representation. This thesis focuses on the use of multi resolution analysis (MRA) in developing suitable features for non-ideal iris images. Our investigation starts with the iris feature extraction technique using Cohen −Daubechies − Feauveau 9/7 (CDF 9/7) filter bank. In this work, a technique has been proposed to deal with issues like segmentation failure and occlusion. The experimental studies deal with the superiority of CDF 9/7 filter bank over the frequency based techniques. Since there is scope for improving the frequency selectivity of CDF 9/7 filter bank, a tunable filter bank is proposed to extract region based features from non-cooperative iris images. The proposed method is based on half band polynomial of 14th order. Since, regularity and frequency selectivity are in inverse relationship with each other, filter coefficients are derived by not imposing maximum number of zeros. Also, the half band polynomial is presented in x-domain, so as to apply semidefinite programming, which results in optimization of coefficients of analysis/synthesis filter. The next contribution in this thesis deals with the development of another powerful MRA known as triplet half band filter bank (THFB). The advantage of THFB is the flexibility in choosing the frequency response that allows one to overcome the magnitude constraints. The proposed filter bank has improved frequency selectivity along with other desired properties, which is then used for iris feature extraction. The last contribution of the thesis describes a wavelet cepstral feature derived from CDF 9/7 filter bank to characterize iris texture. Wavelet cepstrum feature helps in reducing the dimensionality of the detail coefficients; hence, a compact feature presentation is possible with improved accuracy against CDF 9/7. The efficacy of the features suggested are validated for iris recognition on three publicly available databases namely, CASIAv3, UBIRISv1, and IITD. The features are compared with other transform domain features like FFT, Gabor filter and a comprehensive evaluation is done for all suggested features as well. It has been observed that the suggested features show superior performance with respect to accuracy. Among all suggested features, THFB has shown best performance.