Iris Feature Descriptor And Recognition Methods Based On 2D-Gabor Filtering

Iris recognition shows advantages over other biometric recognition technologies in terms of uniqueness, stability, non-invasive and anti-falsification, which earns it a lot attention from academia and industry because of its application prospect in military, security, e-business and so on. To solve the practical problems in iris recognition equipment development and industrialization, this thesis provides an in-depth exploration on iris feature descriptor and recognition algorithms based on Gabor filter. It also make improvements in structuring of Gabor filter banks and parameter selection, iris feature extraction, local texture analysis, feature fusion and multimodal biometric recognition.Our research and findings are as follows:(1) In feature extraction with 2D-Gabor filter, different feature descriptors reflect different qualities of iris textures. Thus there are different filter bank structures and parameter selection. We constructs a multi-channel Gabor filters based on features of energy distribution of iris textures on different filter orientations. Based on detailed analysis of frequency, scale and orientation of filter banks, a method of filter parameter selection, the Energy-Orientation Feature(EOF) is proposed. After tests in several iris databases, our proposed EOF is of better distinguishability and with advantages of magnitude and direction, that is to say, the feature distribution is smooth without any grid edge points and insensitive to illumination changes.(2) Circular scope of Gabor filter bank is incapable of multi-scale analysis of variable ratio. To solve that problem, an iris recognition method of combining multi-sector texture features and weighted fusion is proposed. Such method first constructs a sector region of interest(SROI) according to the distribution of occlusions like eyelids and eyelashes, and then optimizes the normalization scale of every SROI to avoid the local optimum of the EOF extraction. Later, we extract EOFs of iris images on optimum scale. Finally, weighted fusion of all EOFs is carried out for multi-sector encoding and Hamming distance is applied for feature matching. Experiments on CASIAV1.0 and JLU3.0 iris databases show that optimization of normalization scale enhances the distinguishability of EOFs and multi-sector feature fusion can further improves the accuracy and robustness of the recognition system, especially in the case of many occlusions.(3) To speed up the feature extraction and improve the distinguishability of EOF, an iris recognition method based on energy-orientation histogram feature(EOHF) is proposed. EOF extraction is performed with multi-orientation and single scale Gabor filter bank instead of multi-orientation and multi-scale filters to shorten the extraction time and the encoding length. Then, the point-by-point matched EOF is converted into block-by-block matched EOHF with histogram transform. Finally, Euclidean distance is used for feature matching. The experiment results show that EOHF is of better distinguishability than EOF and is less affected by local noise. Moreover, it takes less time but is on high accuracy, so this method satisfies the requirements for real-time iris recognition.(4) The distinguishability of EOHF decreases with the increase of number of samples. In view of this problem, a secondary iris recognition method based on local texture feature is proposed. EOF extraction and histogram transform are applied in the first and second recognition respectively. In the first stage, all the samples are extracted for EOFs and rough classification. In the second stage, histogram transform and EOHF encoding are performed for the remaining misrecognized samples in the first stage. The number of misrecognized ones is far fewer than the totals, so the number of samples is not a problem to histogram transform, in other words, the accuracy of second recognition is ensured because of the absence of effect caused by the sharp number increase of samples on EOHFs.(5) Unlike score level and decision level fusion, feature level fusion demands all the features extracted from unimodal traits with high distinguishability, as well as homogeneity and compatibility, which is difficult to achieve. Therefore most multimodal biometric researches focus on score level fusion, whereas few investigate feature level fusion. In this paper, we propose a face-iris recognition method based on feature level fusion. This method constructs a common feature extraction model of face and iris traits to produce unimodal facial and iris features with homogeneity and high efficiency. One-to-n multimodal biometric recognition based on feature level fusion of high accuracy is proposed with Multi-PCA and SVM fusion strategy. In this sense, it overcomes the shortcoming of Gabor filter which is incapable of face-iris feature level fusion.The experimental results demonstrate that this method can not only effectively extract face and iris features, but also provide higher recognition accuracy. Compared with some state-of-the-art fusion methods, the proposed method has a significant performance advantage.