Biometric Recognition

In the past decade, research on 3D face analysis has been extensively developed, and this study briefly reviews the progress achieved in data acquisition, algorithms, and experimental methodologies, for the issues of face recognition, facial expression recognition, gender and ethnicity classification, age estimation,

etc.

, especially focusing on that after the availability of FRGC v2.0. It further points out several challenges to deal with for more efficient and reliable systems in the real world.

The two phase sparse representation (TPSR) method has achieved promising face recognition performance. However, this method has the following flaw: its recognition accuracy varies with parameter

M

and at present there is no means to automatically set it. As a consequence, it becomes the bottleneck to apply the TPSR method to real-world problems. In this paper, we propose an improvement to TPSR (ITPSR), which can choose a proper value of parameter

M

for obtaining the optimal performance. Extensive experiments show that the proposed ITPSR is feasible and can obtain excellent performance.

Using 3D information is expected to handle challenges in 2D face recognition and improve system performance. Extracting pure facial part in face point cloud is usually the first step in a 3D face recognition system, which was mainly operated by manual in most previous studies. In this paper we propose a fully automatic approach for pure face extraction from 3D point cloud. Considering that 3D face point cloud can often be sensed in combination with color information, we use random forest classifiers to classify skin points and non-skin points in 3D point clouds. Usually there will be a few holes in the obtained skin point cloud, which mainly correspond to eyes, mouth, moustache, etc. We propose an approach based on nearest neighbor search method to fulfill the holes. Experiments show that the proposed approach can extract pure faces with different sizes, poses and expressions under various illumination conditions.

To enhance the discriminant power of features in face recognition, this paper builds a novel discriminant criterion by nonlinearly combining global feature and local feature, which also incorporates the geometric distribution weight information of the training data. Two formulae are theoretically derived to determine the optimal parameters that balance the trade-off between global feature and local feature. The obtained parameters automatically fall into interval [0, 1]. Based on the parameter formulae, we design an efficient cross iterative selection (CIS) algorithm to update the optimal parameters and optimal projection matrix. The proposed CIS approach is used for face recognition and compared with some existing methods, such as LDA, UDP and APD methods. Experimental results on the ORL and FERET databases show the superior performance of the proposed algorithm.

This paper applies a novel clustering method in the image set-based face recognition, called the Muti-Exemplar Affinity Propagation algorithm (MEAP)[11]. The new method is extended from the affinity propagation (AP). It is a muti-exemplar model which constructs a two-level mapping:

ϕ

1

between the feature points and the exemplars, and

ϕ

2

between the exemplars and the super-exemplars. In this paper, we just use the first-level mapping result, i.e the subclasses to take part in the subsequent face recognition. The experiment results in different databases indicate the excellence of our method and the robustness to face occlusion.

In this paper, we propose a novel face recognition method by using vector projection, which uses vector projection length to evaluate the similarity of two image vectors in face image vector space. The projection length of a test image vector on direction of a training image vector can measure the similarity of the two images. But the decision cannot be made by only a training image which is the most similar to the test one. The mean image vector of each class also contributes to the final classification. Thus, the decision of the proposed vector projection classification (VPC) approach is ruled in favor of the maximum combination projection length. The performance of the proposed VPC approach is evaluated using two standard face databases; a comparative study with the state-of-the-art approaches illustrates the efficacy of the proposed VPC approach.

With the increasing demand of surveillance camera-based applications, the very low resolution (VLR) problem occurs in many face application systems. Traditional two-step methods solve this problem through employing super-resolution (SR). However, these methods usually have limited performance because the target of SR is not absolutely consistent with that of face recognition. Moreover, time-consuming sophisticated SR algorithms are not suitable for real-time applications. To avoid these limitations, we propose a novel approach for VLR face recognition without any SR preprocessing. Our method based on the linear combination coefficients of non-local image patches is the same regardless of image resolutions inspired by the learning-based face SR method. Experimental results show that the proposed VLR face recognition method is high in recognition accuracy and robust in resolution variations.

In this paper, a new algorithm named Adaptive Weighted Label Propagation (AWLP) which explores the complementary property among sub-patterns from the same face image is proposed for local matching based face recognition. The proposed AWLP first partitions the face images into several smaller sub-images. Then, multiple similarity graphs are constructed for different sub-pattern sets. At last, in order to take correlation among different sub-patterns into account, the graphs obtained by various sub-pattern sets are combined and the procedures of label prediction and graph weight learning are integrated into a unified framework to propagate the class information of the labeled samples to unlabeled ones. Moreover, a simple yet efficient iterative update algorithm is also proposed to solve our AWLP. Extensive experiments on three face benchmark databases show that AWLP has very competitive performance with the state-of-the-art algorithms.

A novel approach for pupil center location is presented in the paper. It is based on the hypothesis that both of pupil centers are of symmetry about a perpendicular bisector of a face, and uses the center of the face region detected and one of the eye region located to determine another unknown pupil center. To reduce the effect that the center of the face region detected deviates from the perpendicular bisector, the center of the face region is disturbed and two constraints,

i.e.

radial and angular, are used. Thus the pupil center candidates are obtained. Then peak detection, modified least trimmed squares, and PCA-reconstruction-error-minimum are used to select the optimal one. Experimental results show that the approach can be used for pupil center location of faces under pose variations.

Disguised face recognition (FR) is considered as one of the difficult and important problems in FR field. Rather than disguised modeling, a disguised face recognition algorithm based on local feature fusion and geometry coverage is presented in this paper. Local binary pattern (LBP) and local phase quantization (LPQ) is firstly applied to extract the binary and phase statistics features which are robust to the disguised mode, then hyper sausage neuron based on biomimetic pattern recognition (BPR) theory is adopted to construct high-dimensional geometry coverage of different classes, which makes full use of continuous characteristics of identical class face features while avoids the interruption of the disguised mode. Experiments on AR face database and disguised face database established by police face combination software show that, compared with the state-of-the-art methods, the proposed recognition algorithm can achieve high recognition results under disguised conditions.

In order to extract the robust thermal infrared facial features, a novel method based on the modified blood perfusion model and the improved Weber local descriptor is proposed. Weber local descriptor (WLD) is able to extract a wealth of local texture information, which computes not only the differences between the center pixel and its neighbors but also the gradient orientation information describing the direction of edges in the local area, so it is suitable for texture-based thermal infrared face recognition. In order to make full use of local authentication information, an improved Weber local descriptor is proposed to extract the local features from the blood perfusion image. For improved Weber local descriptor, the Isotropic Sobel operator instead of the traditional method is used to compute the orientation and build more stable feature histograms. Experimental results show that the proposed method could achieve better recognition performance compared to the traditional methods.

Face recognition is one of the research hotspots in the area of computer vision and pattern recognition which has a wide application perspective. In this paper, a research on the classical Local Binary Pattern (LBP) is made and an improved algorithm named double-circle LBP is proposed, which can further enhance the rotation invariant characteristic of LBP. Since LBP descriptor based on the block has good recognition effect, this paper further proposed the strategy of "multiple blocks+middle block" in double-circle LBP descriptor, which can effectively solve the problem that the information around the original block line cannot be extracted completely. Finally, experiments are conducted on Orl, Yale and Extended YaleB face databaseds by comparing the recognition rate by using original LBP and its improved algorithms. The results show that double-circle LBP descriptor and "multiple -block+middle-block LBP descriptor can greatly improve the recognition rate.

We design a real-time face recognition system based on IP camera and SRC algorithm by way of OpenCV and C++ programming development. Meanwhile, we do research on the IP camera and rewrite some function of SDK so that OpenCV can process the video frame. First, AdaBoost algorithm is used to detect face in each frame, and then LBP is used to extract the feature of texture. Finally, we obtain the result by SRC algorithm. Experimental results show that the system can deal with real-time video and have robustness to the illumination.

Most of facial expression recognition systems have a low recognition rate for non specific facial expression. Therefore, a new method of facial expression recognition is proposed based on classification tree. According to the differences in expressions, we classify 7 kinds of expressions from coarse to fine. And at each layer in the classification tree, we set feature vectors into different regions, and extract the most feature vectors for classification by LDA. Experimental results show that the proposed method can achieve a recognition rate of 82.38% on JAFFE database, which verifies the effectiveness of the proposed algorithm.

Stimulated by multi-task learning method, this paper proposes an algorithm of Feature Selection based on Multi-Task Learning (FS-MTL) for ethnicity and gender recognition with face images. The proposed FS-MTL selects the common features which are shared by multi-tasks are based on the sparse optimization solution of group Least Absolute Shrinkage and Selection Operator (LASSO). Compared with either the classic feature selection algorithm or the single task feature selection, the proposed algorithm can get higher recognition rate through sharing the related information among tasks. At the same time, the stability analysis is introduced to feature selection. With given stability metrics, the results of experiments show that features selected with the proposed algorithm are more stable.

To support the relevant research on face analysis tasks, face image databases with annotated ground-truth are necessary. Although there are many face databases with large amount of images available with increasing research on face analysis, there are few open large face databases with the coordinates of multiple Facial Feature Points (FFPs) provided. In this paper, we build up a large FFP database combining several existing face databases through mapping the known coordinates of the available FFPs to a unified FFP model. The unified model is established based on multiple principles through very thorough analysis of the existing models. The FFPs are mapped to the protocol model with 7 different algorithms. As a result, we obtain a large face database of 70 FFPs labeled with various gender, ethnicity, age and expressions. This new database can be widely used in other relevant researches.

This paper proposes a novel method for localizing the center of pupils. Given a face detected in an image, it first empirically initializes the eye regions in the face, and locates the pupils within the eye regions by using an improved isophote curvature based method. It then updates the eye regions according to the detected pupil centers. In the updated eye regions, the pupil centers are also refined. The above process iterates until the detected pupil centers have sufficiently high consistency with the eye regions. Compared with previous methods, the proposed method can better cope with faces with varying pose angles. Evaluation experiments have been done on the public BioID database and a set of self-collected face images which display various pose angles and illumination conditions. The results demonstrate that the proposed method can more accurately locate pupil centers and is robust to illumination and pose variations.

Affine Scale Invariant Feature Transform (ASIFT) is robust to scales, rotation, scaling and affine transformation. It could be used for face recognition with pose variation. However, ASIFT requires large data. Could we reduce the data of ASIFT and preserve the face recognition performance? In this paper, we propose an effective face recognition algorithm to combining the structural similarity (SSIM) and PCA-ASIFT (PCA-ASIFT&SSIM).First, we reduce ASIFT dimension using principal component analysis and get PCA-ASIFT. The PCA-ASIFT’s discriminative capability drops because of the dimension reduction. It brings about more false SIFT matching. We further introduce the SSIM to reduce the false matching. The experimental results show the efficiency of the proposed approach.

With the development of E-Commerce, biometric based on-line authentication is more competitive and is paid more attentions. It brings about one of hot issues of liveness detection recently. In this paper, we propose a liveness detection scheme to combine Fourier statistics and local binary pattern (LBP). First, The Gamma correction and DoG filtering are utilized to reduce the illumination variation and to preserve the key information of the image. Then the Fourier statistics and LBP are combined together to form a new feature vector. Finally, a SVM classifier is trained to discriminate the live and forge face image. The experimental results on the NUAA demonstrate that the proposed scheme is efficient and robust.

To overcome the crucial problem of illumination, facial expression and pose variations in 2D face recognition, a novel algorithm is proposed by fusing global feature based on depth images and local facial feature based on Gabor filters. These two features are fused by residual combined with collaborative representation. Firstly, this approach extracts Gabor and Global feature from 3D depth images, then fuses two features via collaborative representation algorithm. The fused residuals serve as ultimate difference metric. Finally, the minimum fused residual corresponds to correct subject. Extensive experiments on CIS and Texas databases verify that the proposed algorithm is effective and robust.

Fake fingerprint detection technology is used for detecting spoof fingerprint attacks in biometric systems. In this paper, an improved software-based fake fingerprint detection approach using wavelet analysis and local binary pattern(LBP) is proposed. Firstly, wavelet analysis is applied to get the denoised image and residual noise image. Then both two images are divided into blocks of the same size to calculate the histogram of LBP as features, which provide more texture information than the features in original wavelet-based method. Finally, support vector machine(SVM) is used for classification. The average rate of accuracy of the proposed approach is 88.53% for all datasets in LivDet2011, and 88.98% in LiveDet2013, while the winner in LivDet2011 is 74.41%, and the winner in LivDet2013 is 86.63%.

Fingerprint is a popular biometric trait for designing an automatic human recognition system. These systems are commonly benchmarked over fingerprints of the urban population whereas their practical deployment involves majority of rural population. Living standards of the rural population is not as high as urban ones. They are mostly involved in hard work and less careful about their skin conditions. Therefore, it is desirable to explore the average quality of fingerprint and the performance of automatic fingerprint recognition system for rural population. This paper analyses the (1) age-group and gender wise quality of fingerprint and (2) recognition performance under cross scanner settings. To justify the analysis, 41400 fingerprints are collected from 1150 participants living in rural areas and actively involved in physically hard work. Participants are from age group of 18 to 70 years. Samples have been collected in two phases with a gap of two months with the help of three different fingerprint scanners. Every participant has provided multiple fingerprint samples in each phase on all three scanners.

Fingerprint recognition technology as a promising high-tech has been widely applied in many fields. Fingerprint matching is one of the most important aspects. The biggest challenge is how to improve the recognition performance, when fingerprint images are with low quality and nonlinear deformation. An improved fingerprint match algorithm is proposed in this paper, which is based on key minutiae and optimal statistical registration. First, this algorithm not only combines the global matching with local matching, but also uses the optimal statistical idea to evaluate the best parameters of rotation and translation between two images. Second, this paper uses local greedy method to get the corresponding key minutia pairs. Experimental results show that the proposed algorithm can rival those advanced algorithms in the world.

The study of the artificial fingerprint detection has lasted for a decade. With full prior knowledge of the spoof attack, researchers extract discriminative features and apply some two-class classifiers to detect the spoof. However, we don’t know the materials of fake fingerprints used by the attackers in the real world. It means that the traditional evaluation is not scientific. In this paper, we proposed to measure the security of fake fingerprint detection systems by the inter-operability performance across various materials. Fake fingerprints made of various materials have diverse feature distributions. The traditional binary SVM over-fits the training negative data. We proposed a novel model named one-class SVM with negative examples (OCSNE) to solve the problem. In order to simulate the real environment, we modified the structure of the Liveness Detection Competition 2011 (LivDet2011) database accordingly. The experimental results on the LivDet2011 modified database showed OCSNE outperforms the traditional SVM.

In order to solve the problem of discontinuous palmprint mainline, this paper presents a novel algorithm which can make the discontinuous mainline continuous. The algorithm consists of four steps: energy expansion based on the Gaussian function, calculation of texture probability distribution, acquisition of texture pixel values based on the exponential function and determination of the iteration. Compared with traditional methods, the distinct advantage of the algorithm is the palmprint mainline restoration with directional properties. Therefore, wrong palmprint mainline restoration could be reduced effectively. The results illustrate that the proposed algorithm is feasible and valid.

As a new biometric technique, finger vein recognition has attracted lots of attentions and efforts from researchers, and achieved some progress in recent years. A survey of progress in finger vein recognition is given in this paper. It mainly focuses on three aspects, i.e., the general introduction of finger vein recognition, a review of the existing research work on image acquisition and feature extraction methods. We finally present the key problems and future directions in order to enlighten finger vein recognition research domain.

This paper concerns about the performance evaluation of finger vein verification algorithms. A finger vein verification competition was held by the author of this paper in the 9th Chinese Conference on Biometrics Recognition. The competition itself is called PKU Finger Vein Recognition. This competition aims at studying the state-of-the-art performance of finger vein verification algorithms, mainly in China. This competition used a general recognition algorithm evaluation platform called RATE developed by the author. RATE provides systematic solutions for database management, benchmark generation, and equipped with a distributed computation system for algorithm evaluation. This paper will discuss RATE platform, the data sets and test protocols of the competition, and the analysis of the competition results.

This paper presents a method to calculate capacity of hand vein features by coding divided sub-images. An image coding model is proposed in this paper, including gray level inertia moment extraction and feature coding. The proposed method is tested on a database of 1000 images from 100 individuals built up by a custom-made acquisition device. The experiment results indicate that the capacity of hand vein features supports over 100 thousand individuals.

The spatial feature interdependence matrix (SFIM) has been proposed for face representation, which encodes feature interdependences between local patches. However, not all patches are equally important for classification purposes. For finger-vein identification, patches that contain vein lines contribute more to classification. Inspired by this, we propose a weighted SFIM based on probability and direction (PDSFIM). Both the probability and direction of vein lines in a patch are integrated into the SFIM. The experimental results demonstrate the superiority of the proposed method after comparison with various state-of-the-art methods.

In this paper, a database of finger vein(FV) and finger dorsal texture(FDT) region of interest(ROI) images named by THU-FVFDT2 is described in detail. The database is provided as an aid in studying fusion strategy of finger vein and finger dorsal texture images. Furthermore, on account of ROI extraction and manually coarse alignment during construction of the database, it facilitates using the database for research. Moreover, some algorithms of finger vein and finger dorsal texture authentication are tested in our database to ensure the reliability of the database and make research performed with the database as consistent and comparable as possible.

In finger vein recognition, the input image is generally labeled in accordance with the nearest enrolled neighbor. However, it is so rigid that it is inadequate for some cases. This paper explores a modified sparse representation method for finger vein recognition. In the method, each block in a finger vein image will be sparsely represented by dictionary textons, not simply labeled by the nearest enrolled block, and the representation coefficients of all blocks are arranged to be a two-dimensional histogram to model the image. As textons is learned from local vein pattern, not global vein pattern. Therefore, for encode global geometric information of finger vein pattern, the representation coefficient histogram is projected to different lines, and then connected in parallel to generate more powerful image features. Extensive experiments on the HKPU finger vein database show the effectiveness of the modified sparse representation method in finger vein recognition.

In order to improve the recognition accuracy with high speed, a palm vein identification method based on multi-direction gray surface matching is proposed. The algorithm extracts region of interesting (ROI) of palm vein image firstly. Then, it computes the multi-direction gray scale’s difference in the matching of surface of two ROI. The variances of the multi-direction grayscale difference surface are calculated and the minimum of variance is considered as the distance between two feature surfaces. At last, the algorithm decides whether these two images come from the same hand or not according to the distance. In the self-build palm vein database, the recognition rate of this method reaches 98.48% and the speed is 21.8ms. Comparing with other typical palm vein recognition methods, the proposed approach improves CCR and decreases FAR.

Great progress of iris recognition has been achieved in recent years driven by its wide applications in the world. This survey summaries the progress in iris image acquisition, segmentation, texture analysis, classification and cross-sensor recognition from 2008 to 2014. The core ideas of various methods and their intrinsic relationships are investigated to obtain an overview and insights in the development of iris recognition. The future research work to improve the usability, reliability and scalability of iris recognition systems is also suggested.

The parameter selection of SIFT operator is the premise and difficulty of feature extraction with SIFT. Based on

$a\!\!\!\!\!\!=$

analysis of the change regulation between each parameter of SIFT operator and the valid key points in detail, a variety of parameter selection ways to fit to extract iris texture features are put forward in this paper. A new set of feature matching method is designed and realized according to the features. According to the experimental results from three public iris databases, including CASIA V1.0, CASIA-V3-Interval and MMU, compared with classical SIFT method of Lowe, the method we proposed has been proven to increase by 2% to 5% in recognition accuracy. It shows that the method we proposed has strong robustness and high recognition ability.

We propose a novel fast robustness palmprint recognition algorithm based on the Curvelet transform and local histogram of oriented gradient (CLHOG) for the poor curve and direction description in the traditional wavelet transform. Curvelet transform is firstly used to obtain four images with the different scales. Then, an algorithm based Local Histogram of Oriented Gradient (LHOG) is designed to extract the robust features from those different scale images. Finally, a Chi-square distance is introduced to measure the similarity in the palmprint features. The experimental results obtained through using the proposed method on both PolyU and CASIA palmprint databases are robust and superior in comparison to some high-performance algorithms.

Eyelash occlusions pose great difficulty on the segmentation and feature encoding process of iris recognition thus will greatly affect the recognition rate. Traditional eyelash removal methods dedicate to exclude the eyelash regions from the 2D iris image, which waste lots of precious iris texture information. In this paper we aim to reconstruct the occluded iris patterns for more robust iris recognition. To this end, a novel imaging system, the microlens-based light field camera, is employed to capture the iris image. Beyond its ability to refocus and extend the depth of field, in this work, we explore its another feature, i.e. to see through the occlusions. And we propose to reconstruct occluded iris patterns using statistics of macro pixels. To validate the proposed method, we capture a unique light field iris database and implement iris recognition experiments with our proposed methods. Both recognition and visual results validate the effectiveness of our proposed methods.

The iris health evaluation emphasizes the detecting and analyzing of local variations in the characteristics of irises. Therefore, to accurately extract intestinal loop region of iris and objectively represent the texture feature is a prerequisite for gastrointestinal health evaluation based on iris. Based on the Canny operator approach, this paper presents adaptive Canny operator’s partition for the extracting of intestinal loop region. This paper presents the measurement method of gray level co-occurrence matrix for representing texture information of irregular intestinal loops and the calculating the 6 texture measure. As the input is to establish the support vector machine model, we solve the classification of different kinds of people. Experiments were performed in collected samples. The detection method can effectively extract different types of the iris intestinal loop region. At the same time, the classification model shows that the proposed texture feature works as a measurement of effective health evaluation basis.

Point light sources always reflect on color iris image captured by portable color iris image capturing device, which will affect the pupil contour extraction. A pupil’s contour extraction method of anti-light spot interference is given in this paper to solve the problem. Firstly, the expanded pupil region is unfolded into a rectangle image. Secondly, all the light spot regions in the rectangle image are positioned by projecting method of axial directions after binarization and median filtering. Thirdly, these regions are filled by image inpainting technique based on fast marching method. Then, pupil contour extraction can be launched. Next, those regions which are close to contour line are repaired so that the entire exact pupil contour is finally extracted out. In addition, the experiment about this method is conducted with 200 color iris images. The results show that this method has considerable validity and adaptability.

Gait is an important biometric feature to identify a person at a distance. However, the performance of the traditional gait recognition methods may degenerate when the viewing angle is changed. This is because the viewing angle of the probe data may not be the same as the viewing angle under which the gait signature database is generated. In this paper, we introduce the separable criteria into the couple metric learning (CML) method, and apply this novel method to normalize gait features from various viewing angles into a couple feature spaces. Then, the gait similarity measurement is conducted in this common feature space. We incorporate the label information into the separable criteria to improve the performance of the traditional CML method. Experiments are performed on the benchmark gait database. The results demonstrate the efficiency of our method.

We address the challenging problem of human pose estimation, which can be adopted as a preprocessing step providing accurate and refined humanpose information for gait recognition and other applications. In this paper, we propose a method and augmented Pose-NMS to process the human pose estimation in the consecutive frames based on a reasonable assumption. The poses between the adjacent frames have small changes. Firstly we merge the multiple estimated pose candidates in a single frame to get the representative pose candidates. Then we propagate the final candidate backward and forward to increase the number of the confident candidates based on the Bayesian theory. We apply our method to the Buffy Video dataset and obtain the competitive result to the state-of-art.

In the i-vector model, the utterance statistics are extracted from features using universal background model. The utterance is mapped to a vector in the total variability space, which is called i-vector. The total variability space provides a basis to obtain a low dimensional fixed-length representation of a speech utterance. But, the processing is complicated for the interweaving of the statistics and machine learning method. So, we considered separating them and proposed a simple way to extract i-vector by classical principal component analysis, factor analysis and independent component analysis from normalized statistics. The results on NIST 2008 telephone data show that the performance is very close to the traditional method and they can be improved obviously after score fusion.

A voice activity detector (VAD) is the prerequisite for speaker recognition in real life. Currently, we deal with the VAD problem at the frame level through short time window function. However, when tackling with the VAD problem manually, we can easily pick out the speech segments containing several words. Inspired by this, we firstly use IIR filter to get the envelope of the waveform and divide the envelope into separate sound segments. And then we extract shape features from the obtained segments and use K-means to cluster the data featured by the amplitude of the wave crest to discard the silent part. Finally, we utilize other shape features to discard the noise part. The performance of our proposed VAD method has apparently surpassed the energy-based VAD and VQVAD with a relative 20% decrease in error rate, While the computation time of the proposed VAD method is only 30% less than that of VQVAD. We also get an encouraging result utilizing our VAD method for speaker recognition with about 3% average decrease in EER.

Currently, the recording playback attack has become a major security risk for speaker verification. The text-independent or text-dependent system is being troubled by it. In this paper, we propose an effective text-prompted system to overcome this problem, in which speaker verification and speech recognition are combined together. We further adopt speaker-adaptive hidden Markov model (HMM) so as to improve the verification performance. After HMM-based speaker adaptation, this system needs not to be retrained at each verification step. Experimental results demonstrated that the proposed method had quite good performance with the equal error rate (EER) lower than 2% and was also robust for different cases.

The correlation and redundancy of the pathological voice features, which is assorted to the feature set by the random or artificial combinations of these features, always affect the detection effect of the voice. In this paper, we present a method of optimization of pathological voice feature based on KPCA and SVM. Thus, the feature parameters are processed, the correlation and redundant information eliminated, and the representable information extracted for recognition by KPCA. Our experiments based on KPCA show that the highest recognition rate of vowel /a/ is 97.47%, the average recognition rate 91.85%, while these two rates of vowel /i/ are 91.39% and 84.15% respectively. Compared with the traditional combination method, the average recognition rate has effective improvement in our experiment based on KPCA.

This paper presents a method based on two structural features for text-independent writer identification, i.e. SIFT descriptor (SD) and triangular descriptor (TD). For SD, we modify the original SIFT algorithm to make the SD possess orientation information, called modified SIFT descriptor (MSD). Acodebookis constructed by clustering the MSDs extracted from training samples. Then the bag of word technique is used to compute a MSD histogram (MSDH) as a feature vector for writer identification. For TD, it is designed to represent the unique relationship between three selected points. A TD histogram (TDH) of the TD occurrences is computed as another feature vector by tracking the contour points of a handwriting image. The distances between MSDHs and TDHs are computed and combined as the final dissimilarity measurement for the handwriting images. Experimental results on two public challenging datasets demonstrate the efficiency of the proposed method.

The design of light source is crucial for the availability of palmprint, hand shape and high quality image of Handmetric identification. After the small (small lighting area, Gao Liangdu) LED light source is added at the front end of an appropriate optical component, it may assist to redistribute the LED flux in receiving specific distant source within a specific range. It may also favor in reducing the volume of light source. Such can be done during the improvement of the uniformity of illumination. The light’s utilization of surface is also objective. This paper analyses in detail of the track’s changes before and after the light through optical components by comparing the use of flux distribution and optical component. The results show that the uniformity is improved after luminous flux component’s distribution.

This paper develops a new image feature extraction and recognition method coined bidirectional compressed nuclear-norm based 2DPCA (BN2DPCA). BN2DPCA presents a sequentially optimal image compression mechanism, making the information of the image compact into its up-left corner. BN2DPCA is tested using the Extended Yale B and the CMU PIE face databases. The experimental results show that BN2DPCA is more effective than N2DPCA, B2DPCA, LPP and LDA for face feature extraction and recognition.

Thedual-core embedded system can make the system have higher efficiency of simultaneous running in the recognition algorithm and controlling the peripheral equipments. This paper presents a study of an embedded multi-biometric recognition system based on DSP and ARM which can realize the face and iris image acquisition, recognition, datastorage and input/output control. The ARM is used as a host to communicate with peripherals. The DSP performs the multi-biometric image acquisition and recognition. The host port interface (HPI) is used to implement the communication between DSP and ARM. We design the HPI strobe signal and the hardware device driver based on the embedded Linux to realize the data exchange and communication. Experimental results show that the dual-core embedded system has greater storage capacity and higher interactive ability.

Breath analysis systembased on electronic nose(e-nose) uses gas sensors to detect biomarkers in breath. Then the health situation of peoplecanbe estimated by analyzing the responses of the sensors. As we know,Even for the same kind of gas sensor, the physical and chemical characteristics of each copy are not same. Therefore, theoutput results are usually not same when measuring the same sample by different breath analysis devices of the same model.This situation will greatly confine the application of the devices. In this paper, a self-designed breath analysis system isintroduced, then, a standardization method is proposed to counteraction the individual difference.The results show that our method is effective. It reduces the error caused by the device variance.

The available images of biometrics recognition system in real-world applications are often degraded and of low-resolution, making the acquired images contain less detail information. Therefore, biometrics recognition of the low-resolution image is a challenging problem. It has received increasing attention in recent years. In this paper, a two-step ear recognition scheme based on super-resolution is proposed, which will contribute to both human-based and machine-based recognition. Unlike most standard super-resolution methods which aim to improve the visual quality of ordinary images, the proposed super-resolution based method is designed to improve the recognition performance of low-resolution ear image, which uses LC-KSVD algorithm to learn much more discriminative atoms of the dictionary. When applied to low-resolution ear recognition problem, the proposed method achieves better recognition performance compared with the present super-resolution method.

Finger has three biometric modalities, fingerprint (FP), finger-vein (FV) and finger-knuckle-print (FKP).Taking these modalities as a whole has a natural advantage in convenience and universality for personal identification. In this paper, a new finger recognition method based on granular computing is proposed. The used granular space consists of three layers and is constructed in a bottom-up manner. For the finger recognition, a coarse-to-fine granular matching scheme is proposed. Experiments are performed on a self-built image database with three modalities to validatethe reliability and performance of the proposed method.

Finger-based biometrics is widely used in identity authentication. In this paper, viewing fingerprint (FP), finger-knuckle-print (FKP) and finger-vein (FV) as the constitutions of finger trait, a new multimodal finger-based recognition scheme is proposed based on granular computing (GrC). First, the ridge texture features of FP, FV and FKP are extracted using the feature extraction scheme of Orientation coding and Magnitude coding (OrientCode& MagCode) which combines orientation and magnitude information extracted by Gabor filtering. Combining the OrientCode and MagCode feature maps in a color-based manner respectively, we then constitute the original feature object set of a finger. To represent finger feature effectively, they are granulated at three levels of information granularity in a bottom-up manner based on GrC. Moreover, a top-down matching method is proposed to test the performance of the multilevel feature granules. Experimental results show that the proposed method achieves higher accuracy recognition rate in multimodal finger-based recognition.

To address limitations of existing biometric access control systems in “smart” living environments, we introduce the design and construction of an intelligent access control system based on multi-biometrics of the finger. We formulate our system on three aspects: hardware structures, feature extraction and matching algorithm design, and software framework. By taking advantage of the high uniqueness of fingerprints and the strong anti-counterfeiting performance of finger veins, the system has considerable improvement in security and accuracy. In addition, it enables a security solution with a combination of pyro-electric sensor, voice message, video intercom, and the burglar alarm, etc. The experimental results show that the equal error rate is 0.27% and the time consumption of authentication is about one second, which demonstrates that our system meets the requirements of an access control system.

Eye detection is a preliminary yet important step for face recognition and analysis. It is a challenging problem especially for unconstrained images. We propose a coarse-to-fine eye detection approach by using a two-level convolutional neural network which follows a biologically-inspired trainable architecture. The first level of our network roughly detects initial bounding boxes, whereas the second level judges whether the detected bounding boxes belong to eyes or not and deletes the non-eye bounding boxes. All remaining bounding boxes yielded from the two-level network are finally merged to give the accurate locations of detected eyes. Experimental results demonstrate the effectiveness of our method for eye detection under complex scenarios.

In this paper, we present a new people counting method based on universities’ surveillance videos. Firstly, we set the threshold value for the HSV space V channel pixel-based on the color of hair so as to detect the head regions. Secondly, we fit the function of the head size and space coordinate and then remove the connected regions which are too big or too small. Finally, we detect the motion using the improved frame difference method and remove the static regions. This method has solved the problems of that the feature is not obvious when students are in different positions with different sizes due to the perspective effect of cameras. There is false detection due to the interference of static entities such as bags and basketballs on tables and chairs. Experimental results show that the average correct detecting rate can reach 87.71%. By calculating, the detected classroom occupancy rate and the actual classroom occupancy rate are almost at the same.

Identification based on electrocardiogram (ECG) is an emerging hot spot in biometric identification. Feature selection is one of the key research points on it. In the paper, features are firstly calculated from fiducial points of ECG. Secondly, the initial feature set is composed of amplitude, interval, slope, area and some clinical indexes. Thirdly, a feature selection strategy is proposed. The strategy uses stepwise discriminant analysis to calculate the contribution (weight) of each feature for ECG identification. On the basis of contribution sorting, accumulative recognition rate is calculated. Furthermore, a key feature subset for ECG identification is acquired when accumulative recognition rate reaches a steady level. Fourthly, the identification procedure works on key feature subset. ECG data from both PTB and laboratory is used in experiments. Experimental results show that the identification accuracy of the two data sets is 99.7% and 94.8% respectively.

We present a novel method for shoeprint recognition in crime scenes. First, a preprocessing algorithm is introduced to remove the complicated background, and then Gabor features and Zernike features are extracted and fused to represent the textural and statistical features of shoeprint images. Lastly, a matching approach is also presented to solve the problem of identifying incomplete shoeprints which account for a large proportion in all the captured images. The samples in our database are directly collected from crime scenes. In the experiment, 104 probe shoeprints are tested on a gallery set containing 1,225 shoeprints. Results show that our method is practical and provides better performance in identifying crime scene shoeprint than other algorithms.

Human-specified appearance features are widely used for person re-identification at present, such as color and texture histograms. Often, these features are limited by the subjective appearance of pedestrians. This paper presents a new representation to re-identification that incorporates data-driven features to improve the reliability and robustness in person matching. Firstly, we utilize a deep learning network, namely PCA Network, to learn data-driven features from person images. The features mine more discriminative cues from pedestrian data and compensate the drawback of human-specified features. Then the data-driven features and common human-specified features are combined to produce a final representation of each image. The so-obtained enriched Data-driven Representation (eDR) has been validated through experiments on two person re-identification datasets, demonstrating that the proposed representation is effective for person matching. That is, the data-driven features facilitate more accurate re-identification when they are fused together with the human-specified features.

This paper addresses the group activity recognition in the still image. We formulate an alternative discriminant contextual model on feature level. On the one hand, it mines the person-joint-context feature model, which describes the interaction of a focal person and its surrounding context. In the meanwhile, the surrounding context is featured with the relative pose, relative location and the scene background. On the other hand, a similar interaction model is formed to learn the interactive correlation between a focal person and its surrounding context. An optimization criterion is proposed to learn the similar interaction model. We show that the optimization problem can be optimized efficiently. Our experimental results show that the proposed model outperforms related works, even though temporal information is not available.

For the problems of the finger root deformation caused by hands’ locations’ inaccurate and a big error of hand shape feature’s extraction. Taking advantage of its role as Finger contour high stability, a present algorithm of hand location is improved. An algorithm is presented for Hand shape recognition which is without the connection with the finger root contour. This method first locates the finger’s central axis, and then extracts finger geometric features that are none of the finger root’s contour. The final statistical characteristic difference between different fingers is adopted to recognize the shape of hands. The method in this paper is good to solve the problem of large deformation at the finger root contour and the big error of the finger root’s location. We can extract the hand’s shape’s features with high stability. The recognition rate can reach to 99.51%.

In this paper, an improved algorithm of thinning hand vein’s image is proposed in order to meet the requirements of post feature extraction and the recognition for hand vein recognition system. First, the image is enhanced and smoothed. Then, the image is segmented by using dynamic threshold segment algorithm. Finally, the segmented image are smoothed, thinned and deburred. The simulative experiments show that the proposed algorithm is effective and is able to obtain a vein skeleton with less distortion.

Several methods have been used to classify patients with Alzheimer’s disease (AD) or its prodromal stage, mild cognitive impairment (MCI) from cognitive normal (CN) based on T1-weighted MRI. In this study, we used LLE to discriminate 453 subjects form the ADNI database. We conducted six pair wise classification experiments: CN (cognitive normal) vs. sMCI (MCI who kept stability and had not converted to AD within 18 months, stable MCI — sMCI), CN vs. cMCI (MCI who had converted to AD within 18 months, converters MCI — cMCI), CN vs. AD, sMCI vs. cMCI, sMCI vs. AD, and cMCI vs. AD. Each of them was repeated for 10 times. The proposed method got the average accuracy of 0.67, 0.79, 0.85, 0.72, 0.75 and 0.65, respectively. The outcomes suggested that the LLE method is useful in the clinical diagnosis and the prediction of AD.

There are two typos in the original version of this paper.

1. In the 3

th

line of the second paragraph in Section 2 (Page 349) the notation should read as follows:

Where x, y denotes the point from the data set X and Y, respectively; A denotes a real symmetrical matrix corresponding to different distance function. (e.g., when A is the identity matrix,

d

(·) becomes a function of the Euclidean distance.)

2. The first line of the last paragraph in Section 3 (Page 352) should read as follows:

It is easy to prove that

E

 = 

Z

Ω

 + 

Z

T

and

F

 = 

Z

Ω

 −

Z

T

are both symmetrical matrix, and the solution to the above optimization problem with respect to P can be computed by the

D

c

smallest eigenvectors of the generalized eigenvalue problem

EP

 = 

λFP

.