This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.
Unimodal biometrics has several problems such as noisy data, intra class variation, interclass similarities, non universality and spoofing which cause this system less accurate and secure. To overcome these problems and to increase level of security multimodal biometrics are used which uses multiple source of information for personal authentication. Multimodal biometrics has becoming very popular now days since it is at the frontier of unimodal biometrics. This paper presents an overview of multimodal biometrics. This includes the block diagram of multimodal biometrics, modules of multimodal biometric system and different levels of fusion in multimodal biometrics and related work is also covered in this paper.
Keywords-Biometrics, multimodal biometrics, fusion level.
An accurate automatic personal authentication is becoming more and more important for the operation of our electronically interconnected information society . Several systems require authenticating a person's identity before giving access to resources. Biometrics has long been known as a robust approach for person authentication . With new advances in technologies, biometrics has becoming emerging technology for authentication of individuals. Biometrics system identifies or verifies a person based on his or her physiological characteristics such as fingerprint, face, palm print, iris or behavioral characteristics such as voice, writing style, and gait. Any human physiological or behavioral characteristic we can use to make a personal authentication but it has to satisfy features like universality, uniqueness, permanence and finally collectability.
Unlike the possession based and knowledge based personal identification schemes, the biometric identifiers can not be misplaced, forgotten, guessed or easily forged , some examples of biometric system are fingerprint recognition, face recognition, palm print recognition, voice recognition etc. Traditional personal identification systems are based on "Something that you have" e.g. Key or "Something that you know" e.g. Personal Identification Number [PIN], but biometrics relies on "Something that you are". Biometric systems used in real world applications are unimodal . They rely on the evidence of a single source of information for authentication. Though this unimodal biometrics has many advantages, it has to face with variety problems like:
Noisy data: - Susceptibility of biometric sensors to noise leads to inaccurate matching, as noisy data may lead to false rejection.
Intraclass Variation: - The biometric data acquired during verification will not be identical to the data used for generating template during enrollment for an individual. This is known as intra-class variation. Large intra-class variations increase the False Reject rate (FRR) of a biometric system.
Interclass Similarities: - Inter-class similarity refers to the overlap of feature spaces corresponding to multiple individuals. Large Inter-class similarities increase the False Acceptance Rate (FAR) of a biometric system.
Non universality: -Some persons cannot provide the required standalone biometric, owing to illness or disabilities .
Spoofing: - Unimodal biometrics is vulnerable to spoofing where the data can be imitated or forged.
Noisy data, Intraclass Variation, Interclass Similarities, Non universality, Spoofing etc problems are imposed by unimodal biometric systems which tend to increase False Acceptance Rate [FAR] and False Rejection Rate [FRR], ultimately reflecting to poor performance. Some of the limitations imposed by unimodal biometrics can be overcome by including multiple source of information for establishing identity of person . Multimodal biometrics refers to the use of a combination of two or more biometric modalities in a Verification or Identification system. They address the problem of non- universality, since multiple traits ensure sufficient population coverage . Multimodal biometrics also address the problem of spoofing as it concern with multiple traits or modalities, it would be very difficult for an imposter to spoof or attack multiple traits of genuine user simultaneously.
Multimodal biometric system can be applied in many civilian applications like banking security, information system security etc. A multimodal biometric system gives decision as either a "genuine" or an "imposter" individual. In principle, Genuine Acceptance Rate [GAR], False Rejection Rate [FRR], False Acceptance Rate [FAR] and Equal Error Rate [ERR] is used to measure the accuracy of system. Generally multimodal biometrics operates in two modes i.e. Enrollment mode and Authentication mode which are described as follows:
Enrollment mode: - In enrollment mode, biometric traits of a user are captured and these are stored in the system database as a template for that user and which is further used for authentication mode.
Authentication mode: - In authentication mode, once again traits of a user captured and system uses this to either identify or verify a person. Identification is one to many matching which involves comparing captured data with templates corresponding to all users in database while verification is one to one matching which involves comparing captured data with template of claimed identity only .
modules of multimodal biometrics
Multimodal biometric system has four modules - sensor module, feature extraction module, matching module and decision making module respectively.
Sensor module: - At sensor module biometric modalities are captured and these modalities are given as inputs for feature extraction module.
Feature extraction module: - At feature extraction module features are extracted from different modalities after preprocessing. These features yields a compact representation of these traits or modalities and these extracted features are then further given to the matching module for comparison.
Matching module: - In matching module extracted features are compared against the template(s) which is stored in database.
Decision making module: - In this module user is either accepted or rejected based on the matching in the matching module.
Multimodal biometric system can operate in serial mode or parallel mode. In serial mode of operation, the output of one modality is used to narrow down the number of possible identities before the next modality is used . Serial mode of operation reduces the overall recognition time. In parallel mode of operation, information from different modalities is used simultaneously. In case of multimodal biometric system decision can be made at various levels of fusion like Feature level fusion, Matching score level fusion and Decision level fusion. The general block diagram for multimodal biometric system is as shown in Figure 1.
General block daigram of multimodal biometrics
fusion levels in multimodal biometrics
According to Jain and Ross , there are three fusion levels in multimodal biometrics: feature level fusion, matching score level fusion and decision level fusion. It is generally believed that a combination scheme applied as early as possible in the recognition system is more effective [8, 9]. These three levels of fusion are described as follows:
Feature level fusion: - In the feature level fusion, signals coming from different biometric traits are first processed and feature vectors are extracted separately from the each biometric trait. After that these feature vectors are combined to form a composite feature vector which is further used for classification. In case of feature level fusion some reduction technique must be used in order to select only useful features. Some of the researchers have applied fusion at feature level. Since features contain richer information of biometric trait than matching score or decision of matcher, it has observed that when features of different modalities are compatible with each other then fusion at feature level achieves more accuracy and gives better recognition rate. Figure 2. Shows Feature level fusion.
Fusion at feature level
Matching score level fusion: - In this level, rather than combining feature vectors, they are processed separately and individual matching score is found and finally these matching scores are combined to make classification. Various techniques such as logistic regression, highest rank, borda count and weighted sum, weighted product, bayes rule, mean fusion, Linear Discriminent Analysis [LDA] fusion, k-nearest neighbor [KNN] fusion, and hidden Markov model [HMM] etc may be used to combine match scores. One important aspect has to be addressed in the matching score level is the normalization of scores obtained from multiple modalities . Z-score, min-max, double-sigmoid, median-MAD, tan-h, and piecewise linear these various normalization techniques can be used for normalization of the match scores. Matching score level is the most widely used fusion level due its less complexity. Many of the researchers have applied fusion at Matching score level [10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 22, 23, 24, 25, 26, 27]. Figure 2. Shows matching score level fusion.
Fusion at matching score level
Decision level fusion: - In this level of fusion first of all each modality is pre classified independent to each other i.e. each biometric trait is captured then features are extracted from that captured trait, based on that extracted features these traits are classified like accept or reject. The final classification is based on fusion of the outputs of different modalities. Figure 3. Shows matching score level fusion. In 2010, A. Cheraghian et al. have applied fusion at decision level .
Fusion at decision level
Till date many multimodal systems have been implemented and deployed for commercial use, these systems are implemented according to different fusion levels and different algorithms, some of them are wavelet based algorithm [9, 10, 11, 18, 24]. Hariprasath. S et al. has given an approach of multimodal system with iris and palmprint using Wavelet Packet Transform [WPT] and score level fusion. It has found that WTP gives high accuracy . A. Kumar et al. proposed a multimodal framework based on face and ear modalities, using Haar wavelet and Scale Invariant Feature Transform [SIFT] features are extracted. Finally integration of their ranks has been done with modified Borda count and Logistic regression method. According to A. Kumar et al. logistic regression is better than borda count method . S. Jahanbin et al. proposed a novel multimodal framework for (2D+ 3D) face recognition using Gabor Wavelet. Gabor coefficients are first computed and finally decision has made with fusion at matching score level .
T. Murakami et al. proposed multimodal biometric system based on face, fingerprint and iris modalities with Bayes decision rule- score level fusion technique and Permutation based indexing technique for identification of these modalities . Y. Zheng et al. have proposed system using multispectral face images with four different fusion methods i.e. mean fusion, Linear Discriminent Analysis [LDA] fusion, k-nearest neighbor [KNN] fusion, and hidden Markov model [HMM] and according to Y. Zheng et al. HMM fusion is the most reliable score fusion method . N. Gargouri Ben Ayed et al. have developed system using fingerprint and face using Gabor wavelet and Local Binary patterns [LBP]. Finally fusion has done at match- score level with weighted sum method and found to be excellent method giving higher performance . Mahesh P.K. et al. proposed the multimodal biometrics system using two traits speech and palm print, Wavelet-Based Kernel PCA and Mel Frequency Cepstral Coefficients (MFCC) is used to extract features of signal. Finally decision has made by fusion at matching score level with weighted sum .
A. Kumar et al. have proposed an adaptive multimodal biometrics using various modalities. Particle swarm optimization at matching score level fusion is used. According to A. Kumar system outperforms at matching score level than decision level . A. P. Yazdanpanah et al. has proposed system using face, ear and gait with the use of gabor wavelet and fusion at matching score level with weighted sum and weighted product approaches . F. A. Fernandez et al. proposed quality-based conditional processing in multi biometrics with fusion at rank level using linear logistic regression approach . A. Cheraghian et al. has proposed system based on 2D and 3D face image by applying gabor wavelet transformation with fusion at decision level fusion . A. Bhattacharjee et al. have proposed multimodal approach with iris and speech modalities with Daubechies wavelet for feature extraction and decision has made by fusion at feature level . D. R. Kisku et al. has addressed multimodal biometric system using face and palm print modalities. According to D. R. Kisku et al. fusion of biometric images at low level makes the system more robust .
Md. M. Monwar et al. has given new approach for multimodal biometric system using face, ear and signature modalities, Principle Component Analysis [PCA] has been used with fusion at rank level with logistic regression, highest rank and borda count approaches and logistic regression . P. Kartik et al. have proposed system using face, speech and signature features. Principal Component Analysis [PCA], Linear Discriminant Analysis [LDA] and Mel Frequency Cepstral Coefficients [MFCC] used for feature extraction. Finally decision has made at matching score level with sum rule . F. Yang et al. have proposed ultimodal biometric systems based on fingerprint, palmprint and hand geometry modalities using Support Vector machine [SVM] and wavelet transform with fusion at matching score itself . Xiao-Na Xu et al. have presented a novel method of feature-level fusion based on Kernel Fisher Discriminant Analysis (KFDA) with weighted-sum rule, average rule and product rule applied to multimodal biometrics based on fusion of ear and profile face biometrics .
S. Ribaric et al. evaluated different techniques of matching score fusion level like Bayes-based normalization, min-max, zscore, median-MAD, double-sigmoid, tanh, and piecewiselinear on different multimodal biometric systems with fingerprint, face and palmprint modalities . A. Jain et al. have developed multimodal biometrics using face, fingerprint and hand geometry modalities with matching score level fusion. As in case of matching score level fusion normalization of match score is important aspect, A. Jain et al. have done normalization of match scores of these modalities with different such as min-max, z-score and tanh technique .
We have observed that multimodal biometrics is frontier to the unimodal biometrics as it overcomes the problems related with unimodal biometrics like noisy data, interclass similarities, intra class variation, non universality and spoofing. There are many multimodal biometrics in existence, but choice of fusion technique plays important role to make these systems more accurate and efficient.