Fusing face and iris for identity recognition inherits the ease to use of face recognition and the high precision of iris recognition. It can provide comfortable and secure user experience, which will promote the wider application of biometrics. It is in urgent need of the effective face and iris image preprocessing methods to improve the performance of the system due to lots of low quality images captured in unconstrained environment. In this thesis, we study some key problems in face and iris image preprocessing on the identity recognition systems that fuse face and iris. Several robust face alignment and iris segmentation methods are proposed to deal with illumination variations, occlusion and noise. Our main contributions are summarized as follows: (1) A robust face alignment framework is presented by using the half-quadratic (HQ) minimization. Based on the framework, we discuss the similarity between the iteratively reweighted least squares (IRLS) algorithm and the multiplicative form of HQ, and reveal the underlying relationship between the $\ell^1$ loss and the Huber loss. These insights provide a better understanding of the sparse error based face alignment, and are instructive for the future development of face alignment. (2) Based on image decomposition, we propose a coarse-to-fine approach for single-sample face alignment. In the coarse alignment stage, the lightness components of test images are estimated by a shared lightness dictionary. In the fine alignment stage, more accurate alignment results are obtained by using the reflectance components. With the benefits of complementary information used in two stages, the algorithm can deal with large illumination and pose variations. (3) A series of iris segmentation methods are developed by using the information of iris boundaries. A set of visual features including intensity, gradient, texture and structure information are used to construct class-specific iris boundary detectors (LBDs). LBDs improve the accuracy of iris boundary detection and eyelid detection on low quality images. We further exploit the shape and structure information of iris boundaries by using boundary segments and pupillary regions, which improves the robustness of boundary detection and speed up iris localization. (4) Based on the appearance of the neighborhood around a pixel, we propose a new iris segmentation method. The multiscale deep convolutional neural networks (CNNs) are utilized to learn the most distinguish...
修改评论