A framework for biometric recognition using non-ideal iris and face

Abstract

Off-angle iris images are often captured in a non-cooperative environment. The distortion of the iris or pupil can decrease the segmentation quality as well as the data extracted thereafter. Moreover, iris with an off-angle of more than 30° can have non-recoverable features since the boundary cannot be properly localized. This usually becomes a factor of limited discriminant ability of the biometric features. Limitations also come from the noisy data arisen due to image burst, background error, or inappropriate camera pixel noise. To address the issues above, the aim of this study is to develop a framework which: (1) to improve the non-circular boundary localization, (2) to overcome the lost features, and (3) to detect and minimize the error caused by noisy data. Non-circular boundary issue is addressed through a combination of geometric calibration and direct least square ellipse that can geometrically restore, adjust, and scale up the distortion of circular shape to ellipse fitting. Further improvement comes in the form of an extraction method that combines Haar Wavelet and Neural Network to transform the iris features into wavelet coefficient representative of the relevant iris data. The non-recoverable features problem is resolved by proposing Weighted Score Level Fusion which integrates face and iris biometrics. This enhancement is done to give extra distinctive information to increase authentication accuracy rate. As for the noisy data issues, a modified Reed Solomon codes with error correction capability is proposed to decrease intra-class variations by eliminating the differences between enrollment and verification templates. The key contribution of this research is a new unified framework for high performance multimodal biometric recognition system. The framework has been tested with WVU, UBIRIS v.2, UTMIFM, ORL datasets, and achieved more than 99.8% accuracy compared to other existing methods.