16.1 Introduction

Biometrics is the asset of a person comprising of physical appearance and behavioral characteristics by which he/she is identified. In recent years biometric systems are successfully installed in many real-world applications like security, authorization, forensic science, etc. Biometrics efficiency is based on the location of data gathering, changes in the environment, and quality of individual interaction with the biometric system. Due to the effective learning capabilities, 3D biometric systems are gaining popularity in various real-world applications and scientific research. This chapter describes face, fingerprint, and iris recognition of 3D images which are proposed in the literature during the year 2009 to 2019. The chapter gives a generalized overview of face, fingerprint, and iris recognition of 3D biometrics and presents challenges along with their trends and prospects.

Some of the evolutionary algorithms like genetic algorithm, particle swarm optimization, and principal component analysis are used in the recognition of 3D biometric traits. Spoofing is a dangerous activity which causes serious effect to biometric systems and leads to abnormal detection. In this chapter, we discuss some of the anti-spoofing methods proposed for 3D biometric traits (face, iris, and fingerprint). Finally, a list of some popular open-source softwares for biometric identification is presented.

The reminder of the chapter is arranged as follows: a detailed advancement on 3D biometric systems are explained in Sect. 16.2; recent anti-spoofing methods are discussed in Sect. 16.3. A brief review of open-source softwares is given in Sect. 16.4, and the conclusions are given in Sect. 16.5.

16.2 Developments in 3D Biometric Systems

Biometric authentication use physical or behavioral characteristics of an individual to identify him. Biometric systems work on two modes – enrolment (acquiring, processing, and storage of biometric samples into a database) and authentication (the test sample system is compared with the database samples to decide its genuineness). We discuss some of the recent major advancement techniques proposed for face, fingerprint, and iris recognition.

16.2.1 Face Recognition

Face recognition gain a significant attention in image analysis and understanding [1, 2]. Automatic recognition of facial expressions and moments are essential topics in computer vision. Facial image recognition attracts a lot of interest in today’s world, because of more usage of digital image processing and computer graphics. There are some well-known challenges existing in the recognition of face:

  • In real world, building a face variation model is a challenging task.

  • Developing novel face detection techniques that are independent of:

    • Facial expression

    • Image condition

    • Pose of the face relative to camera face

    • Absence or presence of facial components such as mustaches, glasses, beards, etc.

  • Developing hybrid face recognition system is a big challenging issue. It is the combination of both local features and holistic approach (enables quick recognition but not suitable for handling very large datasets).

A 3D face recognition technique is proposed to achieve robustness against facial expressions [3]. It uses elementary geometric descriptor, and superiority of the method is evaluated based on recognition rate and cross-validation recognition rate using GavabDB face and Notre Dame FRGC 3D databases [3]. A recognition technique is developed using the surface of the face and principal component analysis (PCA) [4]. It uses surface of the image, average curvatures, and Gaussian as input to PCA. A multilevel approximation technique using B-splines is used for facial surface normalization. Performance of the method is evaluated using ZJU3DFED database and achieves a recognition rate of 94.5% as compared to PCA [4]. In another study, a framework is introduced using simulated annealing (SA) approach for image recognition[5]. The Surface Interpenetration Measure (SIM) is used to match two images and gauge the similarity and get the authentication score combining SIM value and four different face regions (such as forehead, circular, full face region, and elliptical areas around the nose). It uses FRGC v2 database for experimentation which is composed of 3D face images (4,007 images) with various facial expressions and achieves 98.4% classification accuracy [5].

In recent years, face recognition using fuzzy logic is gaining popularity. A fuzzy rule-based matching technique is performed to recognize 3D face [6]. It uses Hausdorff distance to compute similarities among intra-class members. The superiority of the method is evaluated for both synthesized and original face images of Frav3D and GavabDB databases. It shows more than 7% improvement in classification accuracy as compared to original images [6]. The identification of scanned 3D face shape is performed by normalization [7]. This helps to detect the facial landmark and analyze the face shape based on the position of the 3D image. The method comprises of three important phases. Firstly, it converts 3D scanned image to 2D image and then extracts facial landmark features based on CNN and finally converts 2D image into 3D image. The classification accuracy is competent to other methods [7]. A 3D fuzzy GIST feature extraction is used for the analysis of EEG signals. It uses Support Vector Machine (SVM) classifier for classification and considers L*C*H color and information of orientation, based on the movie clip [8].

Feature selection is an important technique to improve the recognition rate. Recently an entropy-based technique is introduced to select the features to improve the facial expression classification rate [9]. It uses two-level SVM to avoid the confusions between the expressions. Experiments are conducted on U-3DFE database, and an average recognition rate of 88.28% is achieved, which is 8% higher than standard technique [9]. Feature selection is a best optimization factor in detection of human beings. A modified multi-objective method is introduced using genetic algorithm to improve the recognition rate in multimodal biometric system [10]. It uses incremental principal component analysis to take out the features and take support of genetic algorithm to achieve multiple goals by optimizing search space. It uses k-nearest neighbor classifier for classification and shows superior performance with respect to false acceptance ratio [10].

The evolutionary algorithms are well-known optimization algorithms which are widely used in various real-world applications. In designing biometric systems, some of the popular algorithms such as particle swarm optimization (PSO) [11] and Ant colony optimization techniques [12] are used to improve the recognition rate.

Fuzzy decision tree is used to classify the knuckle with training by Gaussian and trapezoidal membership functions and measurement of fuzzy information gain and Gini index. The optimal fusion parameters are chosen using Ant colony optimization technique with respect to the level of security [12]. More recently, a hybrid technique based on an evolutionary single Gabor kernel is proposed to detect the face [13]. It uses both particle swarm optimization and gravitational search algorithm to optimize the parameters in a single Gabor filter. It incorporates eigenvalue classifier to estimate the significance of the proposed technique as compared to other techniques like PCA and LDA [13].

16.2.2 Fingerprint Recognition

Fingerprint is a unique and highly reliable feature in human authentication. Traditionally, it is very popular in criminal investigation, and recently it is also used in applications like financial security and access control, etc. It suffers from some bottlenecks to achieve high recognition rate such as:

  • Lack of novel feature extraction techniques.

  • Reliable similarity measurement methods between fingerprints.

  • Proper alignment of fingerprints.

  • Identification of incomplete fingerprints.

  • Lack of effective fingerprints matching method.

A partial or incomplete fingerprint identification is a challenging task. A region-based fingerprint method is proposed to overcome this challenge [14]. In this method pixel-wise technique is used to match the features with the help of correlation coefficient. It has 3 main important steps – alignment, extraction of common region, and computing the degree of similarity. The common regions are obtained by dividing the image into multiple smaller regions. The Fisher Transform is used to compute the local similarities and decrease the effect of distorted regions and use mean to find final degree of similarity between fingerprints [14]. More recently, minutiae points are used to detect partial fingerprints. It uses both bifurcation and termination minutiae points and a crossing number method (9,10, and 11) to scan pixels neighborhood [15].

A unique 3D fingerprint software is described using fringe projection [16] and uses patterns of color sinusoidal fringe. Experiments are carried out using three fringe numbers. The superiority of the method is evaluated based on standard deviation and absolute error [16]. The curvature features, like skeleton of the curve and overall maximum curvatures, are used for classification of human gender [17]. It uses 541 fingerprint database and shows promising equal error rate and also exhibits sectional maximum curvatures to do human gender classification [17]. There are various ways to construct 3D fingerprint. Recently, 3D fingerprint reconstruction is obtained using multi-view 2D images [18]. It reconstruct based on the correspondence with 2D images and use ridge feature, minutiae, and scale invariant feature transformation to establish correspondences. It undergoes hierarchical matching approach and proves that the binary quadratic function is reliable for shape of the finger [18].

In recent years, contactless 3D fingerprint technique attracts many researchers due to its ability in ubiquitous personal identification and accurate recognition [19]. It uses convolutional neural network (CNN) learning model with three siamese network and fully connected layers and take support of multi-view contactless dataset to understand the performance along with contactless 3D fingerprint database. ROC curve is drawn to indicate the novelty of the method and P-value for statistical significance [19]. A hand-based contactless biometric system is proposed using infrared imagery. During image collection, it captures images in infrared and RGB format and uses fuzzy-weighted technique at fusion level to access the image quality for multimodal biometric authentication [20].

16.2.3 Iris Recognition

In this era, authentication and security become two important and mandatory issues in the biometric systems. Iris image-based biometric systems have varied applications such as lend access to premises, tracing wanted and missing human beings, maintain render authentication in ATM, and obtain attendance report for very large-scale corporate systems. Although various iris biometric systems are available in the market, there are some open challenges existing as given [21]:

  • Establishing benchmarks to make consistent in iris recognition without considering age and related issues with eye and its diseases.

  • Develop novel techniques which counter balance the spoofing.

  • Improper capture of the iris image with respect to symbol perspective.

  • To explore iris recognition through deep learning techniques and artificial intelligence.

  • Develop superior techniques to transform 2D iris images to 3D iris images.

  • Techniques to detect fake iris image.

There are various methods to construct 3D iris model such as microplenoptic camera and Python Photogrammetry Toolbox. Age, dropping of eyelids movement, and segmentation of eyelash are the major problems in the detection of iris. A novel method is proposed to obtain 3D images of detached retina [22]. In this method, 12 partial retinal images are taken in clockwise direction and then cut into 12 sectors and resized to near relative sectors. The color information of 2D image is extracted from sphere mapping algorithm, and visualization tool kit is used to create 3D image [22]. A 3D iris image technique is proposed for the detection of iris [23]. Due to the poor lighting and eyelids and eyelashes movement, good quality iris images are not possible to obtain. A 3D iris image is constructed based on the salient fiducial points of two 2D images with the help of triangulation and random sampling consensus algorithm to map corner points [23]. A 3D eyeball tracking method is proposed with low response time and error. It estimates the eyeball movement and eye features during page scrolling [24].

A fuzzy-based classification is performed to detect both palm and iris [25]. The valley detection and neighbor pixel value algorithms are used to detect region of interest in palm and iris, respectively. It also uses statistical-local binary pattern technique to extract local features and bring the feature vector in the same range using Max normalization. Palm features are enhanced using both histogram of oriented gradient (HOG) and discrete cosine transform (DCT), and features of iris are extracted using Gabor-Zernike moment [25]. The Gabor filter is used to extract palmprint and iris features in four orientations with two-level wavelet decomposition [26].

In networking environment, biometric identification faces several challenges, such as leakage of users privacy and storing of users biometric template. Fuzzy vault and fuzzy commitment methods are used to store short keys and biometric templates. A crypto biometric scheme is used to retrieve a secret key based on iris template with the help of fuzzy extractor [27]. Fuzzy extractor is a biometric tool to authenticate the user based on his/her own template. It has two phases: one is enrollment phase where iris template masks the secret key, and another phase is called verification, where secret key is returned based on the similarity of both reference and query template. The Hamming distance is used to understand the variability of both intra- and inter-users. The efficiency of the method is evaluated on CASIA iris database [27]. In another study, fuzzy commitment technique is used to assign the secret key [28]. This technique randomly assigns a secret key with respect to the subject along with the binary features. Iris fuzzy commitment system is developed to improve both security and privacy. The Reed-Solomon and Hadamard codes are used to assess both security and privacy, and Markov chain model is used to describe the iris distribution. Experiments on CASIA iris database show the superiority of the method [28].

Some researchers have used evolutionary techniques to achieve better recognition rate for iris and obtain minimum number of features. Particle swarm optimization technique is used to get the features from iris and control it by using fuzzy rules. Haralick method is used to extract the crucial features from the iris [29]. In another study, iris features are extracted from Daubechies wavelets, and a subset of informative features are obtained from genetic algorithm. The WVU, UBIRIS Version 2, and ICE 2005 datasets are used to indicate its superiority. SVM is used to showcase the recognition rate [30].

In the next section, we discuss some of the recent anti-spoofing methods designed to overcome intrusions in 3D biometric systems.

16.3 Anti-spoofing

Spoofing is a technique to obstruct the normal operation of a biometric system in order to gain unauthenticated access. Anti-spoofing is a counter measure for spoofing and implemented using following approaches [31].

  • Installation of additional hardware: It is costlier and can be invalidated using biometric traits.

  • Accumulation of extra data (information) to distinguish the features.

  • Use of authenticated biometric data which is captured from biometric systems.

The intrusion can happen at sensor level or feature level [32]. At sensor level, intruders pose as clients by presentation or direct attacks like, video, mask, and photo attacks. At feature level, the intruders try to modify the captured biometric data. The sensor-level spoofing is overcome by devising sensors or algorithms that distinguish fake and real faces. The feature-level spoofing is overcome by providing protection to biometric templates in the form of encryption techniques. In this study, we focus on sensor-level attacks.

Two levels of anti-spoofing methods are performed:

  1. (i)

    Hardware: It uses specialized devices along with the sensor to detect characteristics of a live trait like blood pressure, facial thermogram, fingerprint sweat, reflection properties of the eye, etc.

  2. (ii)

    Software: These are applied on the sensor data to extract features that distinguish a live and fake face. Further, software-based methods are classified as static and dynamic methods. The static methods are applied on still images like photographs, while dynamic methods are applied on video sequences. The photo and video attacks are categorized as 2D face recognition system, while mask attacks are categorized as both 2D and 3D face recognition systems. We discuss some of the methods proposed in the literature to overcome photo, video, and mask attacks.

In next section, we describe anti-spoofing methods for face, fingerprint, and iris recognition.

16.3.1 Face Anti-spoofing

The deep learning methods like convolutional neural network are incorporated for efficient anti-spoofing in face recognition system. A novel approach to anti-spoofing using noise modeling and denoising algorithms is proposed [33]. This method handles anti-spoofing for paper attacks and replay attacks. It performs face de-spoofing by decomposing it into noise pattern and live face. The measurement of the noise pattern is done using convolutional neural network. The degradation of the live face occurs in the given steps: color distortion, display artifacts, presenting artifacts, and imaging artifacts. This architecture consists of three parts: (i) the De-Spoof Net (DS Net) which estimates the noise pattern of the image and reconstructs the live face by subtracting input image from estimated noise (ii) the Discriminative Quality Net (DQ Net) and (iii) Visual Quality Net (VQ Net) which are used to control the visual appearance and liveliness of the reconstructed image. This method is tested on three face anti-spoofing datasets, Oulu-NPU, CASIA-MFSD, and Replay-Attack [34]. It uses metrics like Attack and Bona Fide Presentation Classification Error Rates and Half Total Error Rate to indicate the superiority.

A convolutional neural network (CNN) is used to learn the features of the face [35]. Anti-spoofing of face is performed using temporal, color, and patch-based features. Temporal features are converted into gray images and fed into CNN. For color features, the RGB image is transformed into HSV and YCbCr color spaces, since they are found to have more discriminative features and then fed to CNN to build the model. It creates equal size patches to extract the local information and used to train the CNN formed by 18-layer residual network. Each CNN outputs the probability whether a face image is live or fake. Finally, a SVM combines all the probabilities and classifies the given face image as live or fake image. Three databases are used for experimentation, CASIA-FASD [36], OULU-NPU, and REPLAY-MOBILE. This method is measured using equal error rate by incorporating three different face features.

A novel face anti-spoofing is introduced based on textural features and depth information of the face [37]. A CNN is used to train the texture features of face region and background. The depth images are captured from Kinect which is used along with the webcamera. The face regions of live and fake depth images are captured, and then the depth features are extracted using LBP. The video sequences are captured using camera and kinect. The final decision is based on decisions from CNN based texture features and Kinect-based depth features. The input image is classified as live if both the decisions classify it as live else it is classified as fake. A dataset containing depth information of 20 persons is generated using Kinect and RGB camera. This method exhibits lower Half Total Error Rate by combining both texture and depth information.

Long short-term memory (LSTM) units with CNN is implemented to deal with face spoof attacks [38]. The spatial features of video frames are extracted through deep neural network. Temporal features are fed to LSTM units for classification. The input to deep residual network, ResNet-50, is color image of size 3 × 224 × 224. A 2048 feature vector is extracted from the CNN and fed to 256 LSTM units. Finally, classification is performed using softmax as the decision function. It uses CASIA-FASD [36] and Replay-Attack [34] databases to clarify the novelty of the method and achieve lesser error rates as compared to static and dynamic feature-based methods.

Some methods are proposed based on extracting color texture features and depth of the information. Color texture Markov feature extraction and redundant feature elimination using SVM are introduced to recognize the face [39]. Initially, adjacent pixels of face image are analyzed, and Markov process trains the model to classify real and fake images for each color channel. To capture the differences between adjacent pixels, directional difference filtering is employed. It is found that the consistency between adjacent pixels is deteriorated for a fake image in comparison to live image. Further, color channels are explored using mutual texture information. Finally, SVM-recursive feature elimination method is adopted to select distinct features based on weight magnitude, which is a ranking criteria. The SVM classifier gives the final decision for the input face image as a fake or live image. Experiments are conducted by using Oulu-NPU, CASIA-FASD [36], MSU-MSFD, and Replay-Attack databases [34] and show the superiority of the method over PCA-, LDA-, and LBP-based methods.

A new feature extraction method is introduced based on analysis of linear discriminant and legendre moments to extract the face features [40]. The maximum likelihood classifier calculates the Gaussian probability of the feature vector. This classifier is efficient when the variance around mean is narrow and the overlap between various classes is small. The likelihood of a real face belonging to a class is high in comparison to 3D mask face. The experiments conducted on 3DMAD database [41] achieve a recognition rate of 97.6% and spoof FAR of 0.83%. The proposed face recognition system also includes the task of verification, thus avoiding a separate verification stage.

A superior anti-spoofing technique is proposed for face recognition based on gradient texture information and weighted gradient-oriented feature vector from the depth map. Texture properties of the image enable to identify whether face is real or fake. Extensive experiments are performed to find the efficiency of the introduced method using Replay-Attack, NUAA imposter, and CASIA datasets and use detection rate as a metric [42]. Facial recognition systems are vulnerable to mask attacks. A 2D recognition system is proposed to deal with 3D mask attacks. The angular radial transformation (ART) method is used for feature extraction wherein images are projected orthogonally on a radial basis. It includes the features with both imaginary and real part for each circular moments. Further, feature reduction is achieved though LDA which enhances between class variations. For classification, nearest neighbor and maximum likelihood (ML) methods are used. The method (ART +  ML) is tested on 3D Mask Attack database and is found to perform better than LBP +  LDA method. Among the classifiers, the ML classifier exhibits lesser Half Total Error Rate [43].

A multimodal approach of feature-level fusion of different color space features is proposed for face spoof detection [44]. The RGB color space does not show any difference in terms of luminance and chrominance information and use HSV and YCbCr to extract color spaces. The face features in these color spaces are extracted using Enhanced Discrete Gaussian-Hermite moment-based Speeded-Up Robust Feature descriptor. Different band images are fused using Oppositional Gray Wolf Optimization algorithm by assigning optimal weight scores and used K-SVM classifier for classification. The classifier is a combination of k-NN classifier and multiple k-SVM classifiers connected serially to classify the image by using CASIA-FASD [36], Replay-Attack [34], and MSU-MSFD databases. In comparison to other classifiers like CNN +  Backpropagation and CNN +  Levenberg-Marquardt, the proposed classifier achieves higher recognition rates and lesser error rates.

A method to generate 3D face spoof data using virtual synthesis is proposed [45]. Since deep learning-based methods require extensive training samples, this method provides a solution by generating virtual data. A printed photo is transferred to 3D object, and its appearance is manipulated in 3D space. The 3D face object is meshed using Delaunay algorithm. After meshing, transformation operations like rotating and bending are applied on the 3D meshed face. To overcome the imbalance between spoof samples and live samples, (a) ratio of sampled live and spoof instances is fixed during training, or (b) external live samples are imported. This method is tested on CASIA-MFSD, CASIA-RFS, and Replay-Attack databases [34] and is found to perform better in terms of Attack and Bona Fide Presentation Classification Error Rate, Average Classification Error Rate (ACER), and Top-1 accuracy.

16.3.2 Fingerprint Anti-spoofing

A novel local descriptor is proposed for fingerprint liveness detection [46]. The Weber local binary descriptor computes the pixel variations in a local image patch by considering the background intensity also. In addition, the proposed descriptor extracts gradient orientation from center-symmetric pixel pairs. The feature vector is of size 944 with 8 neighborhood pixels. The Mahalanobis distance is used to compare the multivariable distributions and compute the recognition rate using SVM. The experiments are performed using LivDet2011DB, LivDet2013DB, and LivDet2015DB databases and exhibit lesser error rates. An approach utilizing multiple features (gradient and textural) for fingerprint liveness detection is proposed [47]. The low-level gradient features are extracted using Speeded-Up Robust Features (SURF) which is invariant to illumination, scale, and rotation. To overcome the variations due to geometric transformations, the local shape information is extracted using histogram of orientation gradient (PHOG). Also, texture features are captured using Gabor filters. For dimensionality reduction of SURF+PHOG and Gabor features, PCA is applied. The classification is performed using SVM and Random Forest, and dynamic score level combines the results. The experiments conducted on LivDet 2013 fingerprint database reveal that SVM performs better for SURF +  PHOG features and Random Forest performs better for Gabor features.

A convolutional neural network approach is proposed for fingerprint liveness detection [48]. A comparison of four models is performed. The first model uses a convolutional neural network with random weights (CNN-Random) for feature extraction, followed by dimensionality reduction using PCA. Recognition rate is obtained by building model using SVM with RBF. The second and third models are CNN-AlexNet [49] and CNN-VGG [50], respectively, which are pre-trained for natural images. The fourth model extracts features using binary patterns which are present locally. The histogram image of LBP is further reduced by PCA and classified using SVM classifier. The experimentation is carried out on LivDet 2009, 2011, and 2013 databases. The CNN-VGG model exhibits the least error rate in comparison to other models.

16.3.3 Iris Anti-spoofing

Iris liveness recognition is based on quality assessment parameters [51]. It uses 22 features pertaining to focus, motion, occlusion, contrast, and dilation properties. Pixel intensity, angle information using directional filters, etc. are gathered from different sources. To overcome the issue of large dimensionality, Pudil’s sequential floating feature selection algorithm is used for feature selection. Finally, iris image is classified either fake or real using standard quadratic classifier. From the experiments conducted on BioSec baseline database, it is found that individually the occlusion features exhibit least classification error rate, while the combination of all the features (occlusion, dilation, contrast, and others) reaches a zero error rate. A novel work based on Laplacian decomposition for iris images is proposed to overcome presentation attacks in visible spectrum and near-infrared iris systems [52]. It decomposes each image into multiple scales of Laplacian pyramids. At each scale, short-time Fourier transform (STFT) is applied to obtain responses in four directions (0, 45, 90, and 135 degrees). The histogram is formed for each response and generates a feature vector. The final vector is the concatenation of all the feature vectors of each scale in four directions. Finally, classification is performed using SVM with polynomial kernel. It uses presentation Attack Video Iris Database obtained from iPhone 5S and Nokia Lumia 1020 and exhibits a classification error rate of 0.64%. The proposed system is also efficient in achieving an error rate of 1.37% for LivDet iris database comprising of near-infrared images.

A methodology based on pupil dynamics for eye liveness detection is proposed [53]. The pupil dynamics is expressed in terms of change in its size and shape which is considered as a circular approximation. Hough transform is used to localize pupil in each frame, and a directional image representing the image gradient and direction is generated. Each iris image is converted into a time series of pupil radii. The gradient values above a certain threshold are only considered, and if not even a single gradient is above the threshold, the pupil is not detected. The dilation and constriction of pupil in the presence of variation in light intensity are modeled by Kohn and Clynes and transformed into a seven-dimensional feature space. Finally, SVM is used for classification using linear, radial, and polynomial kernels. A self-generated dataset is used to compute the performance, and acceptable error rates are obtained. However, this method has drawbacks, like the time required for capturing the pupil dynamics, variations in pupil with age, and psychological conditions.

16.4 Open-Source Softwares

In this section, we introduce some open sources which are proposed in the literature. An OpenBR Collaboratory is an open-source project for the development of biometric research having an introduction to 4SF face recognition algorithm [54]. The ImageWare Systems maintains an open-source project called Open Biometrics Initiative (OBI) [55] having two APIs – OpenEBTS API based on Electronic Biometric Transmission Specification standards and OpenM1 API based on the INCITS and ISO standards.

BioSecure NOE [56, 57] has developed many open-source systems using publicly available datasets. It has modalities for iris, fingerprint, hand, signature, speech, and talking face. The signature modality is based on Hidden Markov Model and Levenshtein distance and uses geometry of fingers to recognize hand modality with support of MCYT-100 benchmark dataset [58]. The BioSecure reference system developed two algorithms, closet iterative points and thin plate spline warping for 3D face modality using 3D RMA database. It also developed modality for iris using CBS dataset [59]. The BioAPI is open source for biometric technology to provide single sign-on web authentication system [60].

16.5 Conclusions

The challenges of 3D biometric systems with respect to face, fingerprint, and iris are presented. The recent advancements in these systems are discussed. A detailed explanation about various anti-spoofing methods are discussed to overcome intrusions from impostors. Finally, an overview of existing open-source softwares is mentioned.