Iris recognition is a novel method for personal identification based on the similarity between the features of two iris images, which is a mission-critical technology, having many applications in customs, banking, network, public security, welfare distribution, etc. In iris recognition systems, how to represent the feature information embedded in an iris pattern is a key factor to the system’s performance. Although a number of methods have been proposed for iris representation, a general systematic framework has not been established. In this thesis, we attempt to address this issue. Our contributions include: 1)Based on the fact that the iris texture is radially distributed, we propose to use the spatial distribution of Fourier frequency energy for iris image quality assessment. Both defocused and motion blurred iris images can be recognized and excluded. 2)Inspired by the biological system’s response function to visual signal, we propose a general framework of iris feature representation based on ordinal measures. 3)A novel dissociated multi-pole filter is developed to extract the non-local ordinal measures of iris images. This method improves the information contents and robustness of iris features and breaks the bottleneck of the state-of-the-art iris recognition methods because they are based on only local ordinal measures. Based on the novel method, we achieve higher accuracy and lower computational costs simultaneously. 4)A novel robust direction estimator is proposed based on directional diffusion and directional filtering, to extract the robust directional features of gradient vector field of iris images for iris recognition. 5)We represent iris images using graphs, i.e. regarding the image blocks as nodes and the histogram of local binary pattern as the attributes of the nodes. And a fast graph matching algorithm is proposed to measure the similarity between two iris images. 6)Based on the zero-crossings of wavelet transform, we segment iris images into blob regions. We regard the centers of the blobs as the control points and their moments as the associated attributes. Then a string matching algorithm used in fingerprint minutiae matching is exploited to align two blob patterns and the number of matched blob pairs is used to measure their similarity. We propose a cascased classifier to integrate different iris recognition methods, which improves iris recognition accuray with little extra computational cost. 7)The ordinal measures based iris feature representation model is successfully extended to palmprint and face recognition.
修改评论