Research On Human Ear Recognition Based On Discriminative Representation And Feature Matching

As an important person authentication technique,biometric recognitions have been widely applied in surveillance applications,forensics and criminal investigations.Biometric system can provide much higher security solution than traditional personal authentication systems such as tokens or passwords.However,the tokens can be stolen,and the long passwords or secret codes are difficult to remember and can be forgotten.Moreover,with the requirement of more security systems for forensics and access control,immigration,and commercial applications,biometric systems have attracted increasing attention.Biometric traits can be grouped into two categories,i.e.physiological traits and behavioral ones.Physiological traits include face,fingerprint,ear,palm print,hand geometry,finger vein,and iris etc.On the other hand,behavioral traits involve gait,keystroke dynamics,signature,etc.In the last few decades,several of biometric recognition methods have been studied on face,fingerprint,and iris.Recently,ear print has received considerable research interests in biometric community due to its several prominent advantages.Human ear is large and visible for acquisition,stable through age and expressions,and it can be used for identical twins and triplets.Compared with other biometric traits,human ear has a stable structure with different ages.Also,the ear is insensitive to the variations such as make-up,glasses,and facial expression.In addition,the human ear is also easy to acquire with little person awareness and user cooperation.Furthermore,it has been proven that the left and right ears of the same person have some similarities but are not strictly symmetric.In summary,human ear satisfies all the required of biometric characteristics such as universality,uniqueness,permanence and measurability.Therefore,the ear recognition has received increasing researcher's attention.For ear recognition systems,several methods have been proposed focus on two aspects,i.e.feature extraction,and classification methods.However,these methods usually cannot achieve a desired performance and suffer from some limitations such as high feature dimensions,sensitive to scale and rotation variations,or limitation of training ear images.Moreover,some of these methods require images alignment and image registration,and affected by some occlusions as hair,earrings,glasses,or headphones.These have led us to propose different feature extraction approaches and feature matching methods for ear recognition problem.The main contributions of this thesis are four aspects as follows:1.Numerous methods have been proposed to extract geometric features for ear images.However,those features usually have relatively high feature dimension or are sensitive to rotation and scale variations.In Chapter 2,we introduce a novel geometric feature extraction method for better ear representation.The proposed method adopts the maximal and minimum ear height lines as feature vectors due to that the contour of outer helix is stable.This method shows that the minimum ear height line is helpful to characterize the contour of outer helix and the combination of maximal and minimum ear height lines also can achieve better recognition performance.In addition,we extract three ratiobased features.Overall,the final features we obtain have the advantages,including low dimension,be robust to rotation and scale variation.Moreover,our geometric features are complementary to appearance-based features.Improved recognition performance can be obtained by combining these two types of features.Experimental results on two popular databases,USTB and IIT Delhi,show that the proposed approach can achieve promising recognition rates.2.Compared with the global and geometric features,local features of ear images capture more local information of ears,which can benefit to improve the recognition performance.Those traditional ear recognition methods based on local features always need accurate images alignment or image registration,which has a great influence on the recognition performance.In Chapter 3,we present,evaluate and analyze the effect of local features and their fusion on aligned and non-aligned ear images to investigate an improved method for unconstrained ear recognition problem.In this regards,the discriminant correlation analysis algorithm is adopted to fuse different local features for improving ear image representation.In this way,each feature may provide complementary information for improving ear recognition and system performance.Experiments are conducted on the three public and difficult ear databases,IIT Delhi,USTB,and AWE databases.The experimental results show superiority of our proposed approach on ear verification and ear recognition.3.Prior works focus on the extraction of handcrafted features according to the characteristic of ears.Recent progresses on deep convolutional neural networks(DCNNs),especially the powerful representation ability has shown promising results in various computer vision tasks.Benefiting from that,we propose to extract deep features of ear images for representation.In Chapter 4,we first evaluate the recognition performance with single layer features extracted from the VGG-M model.To leverage the edge based information of the bottom features and the semantic information of the top features,we employ the Discriminant Correlation Analysis(DCA)for fusing different layer deep features.On the other hand,the adopted DCA performs dimension reduction for the sake of computation efficiency and storage memory caused by the high dimension of deep features.For recognition task,we propose to transform the recognition task to the binary classification problem by composing pairwise samples and solve it with the pairwise SVM owing to the lack of ear images per person.Experiments are conducted on four public ear databases: USTB collections I,II,and IIT Delhi collections I,II.The experimental results prove that the proposed method can achieve promising recognition rates and performs favorably against the state-of-the-art methods.4.Motivated from pairwise SVM which classifies the pairwise samples from same person to the positive one and ears from different persons to negative one,we explore metric leaning for the ear recognition problem in Chapter 5.The previous metric learning methods construct the pairwise constraints as a preprocessing step.They use the fixed pairwise constraints in training stage.Those methods suffer from drawbacks that the number of training pairs is limited,and some pairs are never used in training,which leads to that the trained model will under-fits the non-used training pairs.Therefore,we first propose a novel dynamically pairwise constrained based metric learning method for ear recognition.The proposed method learns the distance metric for several cycles to dynamically update the pairwise constraints.In each training cycle,this method selects the nearest similar and dissimilar neighbors of each sample to construct the pairwise constraints.The optimization problem then is solved by the iterated Bregman projections.In this way,the proposed method incorporates more pairwise constraints for training.The experimental results achieve a promising recognition rates and show superior performance compared with the state-of-the-art methods.