Research On Defocused Iris Recognition Methods

Every person has iris. Being one of the biometric features, the iris of each individualis unique, highly reliable and non-duplicable, which makes the iris recognition be one ofthe most promising biometric identification techniques. Traditional iris recognitionalgorithms obtain good performance only on the focused images captured in anenvironment where the camera focus length is fixed and the Depth of Field (DOF) is small.The highly constrained imaging condition which requires the good cooperation from usersmakes the practical application of iris recognition face big challenges. Therefore, thepopularization and application of the idea about collecting iris images in a more naturalenvironment that extends the DOF is the bottleneck of iris recognition research and istherefore the research hotspot.Specifically targeting the key problems that the performance of recognition willdecrease when the fixed focus iris recognition system works in a less constrainedenvironment by extending the DOF and therefore introducing the defocused imagesoutside the DOF, this thesis tries to propose the defocused iris recognition methods that arerobust to optical defocus. The proposed methods differ from existing techniques that relyon special hardware to extend the DOF, computationally expensive algorithms to restorethe defocused images prior or get the aids from other biometrical features. The maincontributions of this thesis are as follows:(1) For the problem that optical defocusing increases intra-class variation from thesame individual during iris recognition, the robustness of iris features for varying degree ofimage defocusing when the images are captured outside the DOF of cameras isinvestigated for iris recognition system that extends the DOF. Then three methods forstable feature based feature extraction and recognition are proposed, which locates thestable features in coding level, feature level and pixel level respectively.Firstly, a defocused iris recognition method that is based on the multi-sampling imagesequence stable code bit (SCB) is proposed. Such method determines the stable feature’s code representation bits for recognition by comparing the codes of the defocused iris imagesequence.Secondly, a defocused iris recognition method that is based on the stable phase vectordisplacement (SPVD) is proposed. Such method determines the stable phase features forcoding and recognition by comparing the displacement of the phase vectors in complexplane from a single iris image.At last, a defocused iris image recognition method that is based on the stable darkpoint region (SDPR) is proposed. Such method determines the stable dark point regions forfeature extraction, coding and recognition by comparing the region dark point featurecontrast from a single iris image.The experimental results show that, compared to the classical Gabor full code (GFC)based recognition method, the proposed methods reduce the intra-class variation offeatures and achieve lower false reject rate and the equal error rate decreases by14.81%、10.65%and11.11%respectively, which fully illustrate that the proposed methods increasethe robustness of defocused iris feature effectively by excluding the non-stable featuresthat are sensitive to optical defocusing. SCB method achieves the best accuracy among thethree proposed methods, however, its requirement of multi-sampling registration increasesthe hardware cost and the complexity of the image sampling system. Therefore, SPVD andSDPR that are based on single sampling registration have the advantages that are moresuitable for practical applications.(2) For the problem that optical defocusing blurs iris and reduces effective featuresduring iris recognition, this thesis proposes a stable feature fusion (SFF) solution. Theproposed method achieves the recognition by statistical studying the distribution rule of thestable features in the spatial domain and frequency domain in the cases of varying imagedefocused degree, determining the rules to extract the stable dark region features andFourier low frequency phase features through single image registration and fusing them inthe matching score level finally. The experimental results show that, compared to themethod that is based on single feature before fusion or the method that is based onperiocular and iris biometrics fusion (PIBF), the proposed method achieves betterrecognition performance with the reduced equal error rate by7.87%at most, which fullydemonstrate that the SFF method, which extract diversified stable features, effectivelyextend the iris’s distinguishable features, and illustrates that it’s beneficial to introduce theidea about fusion for increasing the defocused iris recognition performance. (3) By analysizing the law that how the shape of iris texture evolves along with thechange of defocusing degree and encoding the evolution in the descriptor called the texturecontrast degree, The SPVD method builds a functional relationship model between thephase feature (Gabor filter response) and texture contrast degree. The built model isvalidated by the experiment results obtained from a practical system, which show that thevariation rules of the phase feature and texture contrast degree are fitted for the theoryanalysis.(4) Carry out a well evaluation of the performance of the defocused iris recognitionmethods. Due to the practical requirement for iris recognition system that extends the DOF,it is necessary to guide users to stand in the allowable range for collecting the images sothat the whole process is comfortable and fast for users. This thesis evaluates the effects onthe recognition performance from different iris images sampling range (Iris imagedefocused depth), which supply the theoretical evidence for the choice of defocused irisrecognition methods under different situations.(5) For the problem that DOF extended iris recognition system increases the falseclassification rate on the defocused images with sparse content, a method to evaluate thedefocused depth of the defocused iris images based on the feature pixel is proposed.According to the different pixels’ contribution to the evaluation of defocused depth for irisimage, the weights measuring the importance of the features are assigned adaptively so thatthe pixels corresponding to the features play more important roles while the pixels that arenot so relevant play less important roles. The experimental results showed that theproposed method not only increases the sensitivity to the defocused depth evaluation fordefocused image with sparse content but also for the iris image around the focused point,which illustrates that the proposed method can provide technical support for the fieldapplication of iris image collection system whose DOF are extended.