Robust Iris Recognition Method Under Non-ideal Conditions

Human iris has a unique, stable, non-invasive characteristics, iris recognition technology is one of many biometrics technologies with the highest recognitionrate. With the rapid development of iris recognition technology, current researchfocuses on iris recognition under non-ideal conditions. The quality of acquired iris images degrade seriously and contain high-light reflection, eyelids or glass cover, elastic deformation, motion blur and noise pollution, which bring great challenges to the iris recognition. In order to resolve the interference and shielding factors of noise pollution and corruption on iris recognition, this thesis focus on integration of multiple features or multiple modal cues for robust feature extraction,while concern about iris recognition method based on image indexing for large-scale data.Firstly, this thesis proposes an iris recognition method based on ordinal encoding of Log-Euclidean covariance matrix, which calculate cross-correlation relations among various features under the framework of Log-Euclidean covariance matrix, then use ordinal measure to encode the feature vectors extracted from neighboring regions and finally the Hamming distance is employed to measure the similarity. The thesis presents a robust method using low-rank matrix recovery theory,and proposes a novel feature descriptor, ordinal relation matrix(ORM), which transform the iris recognition problem into the low-rank matrix decomposition of joint ORMs from probe image and training images while using Lagrange multipliercoefficient method to optimize the objective function. Then the thesis presents arecognition method fusing periocular data and iris data. The global statistics andlocal feature descriptors respectively extract the color structure information of periocular image and character of iris texture, and their matching values are integrated by using a weighted sum rule at score level fusion strategy. Finally, the method adapting to large-scale data for iris recognition is presented, employing a simple but efficient hash based template coding for image indexing, the test image directly compare with the candidate subset after indexing for improving the matching efficiency.The proposed methods are evaluated on UBIRIS.v2,CASIA-IrisV4and ND-Iris-0405dataset. Finally, the thesis summarizes the proposed algorithms and propose the research directions in the future.