Due to the intrinsic drawbacks of our human biometric traits, traditional single modal biometric systems cannot achieve both high accuracy and high usability, which prevents the widespread applications of biometric recognition. Human eye region not only can be captured easily, but also contains rich biometric traits, especially, iris texture and periocular pattern. The fusion of them is promising due to the high uniqueness of iris texture and the high usability of periocular pattern. This thesis covers the key problems in the fusion of periocular and iris biometrics, and the objective is to establish a high-performance and user-friendly multibiometric system. In particular, the main contributions are summarized as follows: 1. In the general framework of eye image preprocessing, based on the geometry characteristics of iris region: 1) A robust noisy iris boundary localization method is proposed by combining the top-down and bottom-up segmentation. In detail, two types of energy are designed to evaluate the localization results derived from these two strategies, and a ratio based decision level fusion scheme is proposed to obtain the better result; 2) An adaptive level set based deformable iris segmentation method is proposed. At first, a distance map is generated via an isotropy sigmoid function by considering the spatial relationship of iris center and other pixels. After this, a Semantic Iris Contour Map is generated by combining the distance map and the gradient map as the edge indicator for level set based segmentation. At last, a convergence criterion and a means of updating the parameters are designed carefully for robust and accurate curve evolution; 3) An ellipse fitting and geometry transform based method is proposed for off-angle iris texture correction. In detail, iris inner boundary is fitted as an ellipse based on the localization result of iris segmentation, and rotation transform and scale transform are performed to correct the deformable iris texture based on the parameters derived from the ellipse. All these methods improve the robustness and accuracy of eye image preprocessing. 2. Based on the analysis of local descriptor based representation and its graph modeling, this thesis proposes a hierarchical feature selection method with two layers. In detail, structured sparsity learning is adopted to reduce the redundancy of descriptors in local regions in the first layer. After that, AdaBoost learning is performed to select the most discri...
修改评论