Matching deformed and occluded iris patterns: A probabilistic model based on discriminative cues

The complexity and stability of the human iris pattern make it well suited for the task of biometric recognition. However, any realistically deployed iris imaging system collects images from the same iris that exhibit variation in appearance. This variation originates from external factors (such as changes in lighting conditions) as well as the subject's physiological responses (such as pupil motion or eyelid occlusion), which make reliable recognition a difficult task. We discuss pre-processing techniques which isolate and normalize the iris pattern, and survey existing techniques for iris pattern matching. The standard iris matching algorithm aligns the iris pattern in a way that accounts for relative rotations of the eye. However, we assert that the matching performance becomes more robust when the matching algorithm explicitly models and estimates pattern deformation and partial occlusion, and uses this information to determine the final match score. The novel work in this dissertation is divided into three main areas (i) the analysis of a variety of bandpass filter bank iris pattern representations and optimization of such a representation for its ability to discriminate between iris classes, (ii) the derivation and application of fusion correlation filters to generate distortion-tolerant similarity cues between iris patterns, and (iii) the design and implementation of an iris-specific probabilistic model on the hidden states of deformation and occlusion, capable of estimating the posterior distribution over these states for a given iris pattern match comparison. Fundamentally, this dissertation combines signal processing techniques with probabilistic machine learning techniques to solve the iris recognition problem. The performance of the proposed algorithm is compared to that of the standard iris matching algorithm on three datasets: one from the Chinese Academy of Sciences (CASIA), one from the NIST Iris Challenge Evaluation (ICE), and one collected by the authors at Carnegie Mellon University (CMU). In our experiments, we demonstrate the superior accuracy of the proposed technique using both real and artificially distorted iris data. Finally, we analyze of the computational cost of our iris recognition algorithms in terms of complexity and processing time.