Fusion by Biometrics 主講人:李佳明、陳明暘 指導教授:林維暘 1 Outline Introduction Score normalization methods Fusion methods Experiment Results Conclusion 2 Reference Paper A. K. Jain, K. Nandakumar, and A. Ross, “Score normalization in multimodal biometric systems," Pattern Recognition , 2005. 3 Why score normalization ? 1. The matching scores at the output of the individual matchers may not be homogeneous. 2. The outputs of the individual matchers need not be on the same numerical scale (range). 3. The matching scores at the output of the matchers may follow different statistical distributions. 4 Why score normalization ? Score normalization refers to changing the location and scale parameters of the matching score distributions at the output of the individual matchers, so that the matching scores of different matchers are transformed into a common domain. 5 Score normalization When the parameters used for normalization are determined using a fixed training set, it is referred to as fixed score normalization. In adaptive score normalization, the normalization parameters are estimated based on the current feature vector. 6 A good normalization scheme Robustness refers to insensitivity to the presence of outliers. Efficiency refers to the proximity of the obtained estimate to the optimal estimate when the distribution of the data is known. 7 Normalization Techniques 1. 2. 3. 4. 5. 6. Min-max Decimal scaling z-score Median and MAD Double sigmoid function tanh-estimators 8 1. Min-max normalization Min-max normalization is best suited for the case where the bounds (maximum and minimum values) of the scores produced by a matcher are known. We usually shift the minimum and maximum scores to 0 and 1. xk : the kth matching score before normalization xk’ : the kth matching score after normalization 9 1. Min-max normalization This method is not robust (i.e., the method is highly sensitive to outliers in the data used for estimation). Min-max normalization retains the original distribution of scores except for a scaling factor and transforms all the scores into a common range [0, 1]. 10 2. Decimal scaling For example, if one matcher has scores in the range [0, 1] and the other has scores in the range [0, 1000], the following normalization could be applied. The problems with this approach are lack of robustness. 11 3. z-score The most commonly used score normalization technique is the z-score that is calculated using the arithmetic mean and standard deviation of the given data. Both mean and standard deviation are sensitive to outliers and, hence, this method is not robust. Z-score normalization does not guarantee a common numerical range for the normalized scores of the different matchers. If the input scores are not Gaussian distributed, z-score normalization does not retain the input distribution at the output. 12 3. z-score 13 4. Median and MAD Robust : The median and median absolute deviation (MAD) are insensitive to outliers and the points in the extreme tails of the distribution. This normalization technique does not retain the input distribution and does not transform the scores into a common numerical range. 14 4. Median and MAD 15 5. Double sigmoid function The normalized score is given by where m is the reference operating point and s1 and s2 denote the left and right edges of the region 16 5. Double sigmoid function where the scores in the [0, 300] range are mapped to the [0, 1] range using m = 200, s1 = 20 and s2 = 30. Generally, m is chosen to be some value falling in the region of overlap between the genuine and impostor score distribution, and s1 and s2 are made equal to the extent of overlap between the two distributions toward the left and right of m, respectively. 17 5. Double sigmoid function 18 6. tanh-estimators The tanh-estimators introduced by Hampel et al. are robust and highly efficient. The normalization is given by Hampel estimators are based on the following influence ( )function: where μGH and σGH are the mean and standard deviation estimates, respectively, of the genuine score distribution as given by Hampel estimators 19 6. tanh-estimators 20 Summary of Normalization Techniques 21 Experimental Results Database of 100 users with three modalities. Each user having five biometric templates for each modality. 22 Experimental Results 23 Experimental Results 24 Experimental Results 25 Experimental Results 26 Experimental Results 27 Experimental Results 28 Experimental Results 29 Feature Level Fusion “Biometric A” feature vectors : X “Biometric B” feature vectors : Y Normalization -> X’ , Y’ Dimension reduction Combine two vector -> Z’ = { X’ , Y’ } 30 Reference Paper A. Ross and R. Govindarajan. “Feature Level Fusion Using Hand and Face Biometrics.” In Proc. SPIE Conf. on Biometric Technology for Human Identication II, volume 5779, pages 196-204,Orlando, 2005. SON, B. and LEE, Y. “The Fusion of Two User-friendly Biometric Modalities: Iris and Face”, IEICE Transactions on Information and Systems , 2006. 31 Fusion in feature and matching level Normalization method : median and MAD Dimension reduction method : PCA , LDA Matching score fusion method : sum rule Consider feature vectors {Xi , Yi} and {Xj , Yj} obtained at two different time instances i and j. Fusion in feature level -> { Zi , Zj } Let sX and sY be the normalized match (Euclidean distance) scores generated by comparing Xi with Xj and Yi with Yj , respectively. smatch = (sX + sY)/2 be the fused match score obtained using the simple sum rule. 32 Experimentation A set of 500 face images and hand images were acquired from 100 users (5 biometric samples per user per biometric) Each face image was decomposed into its component R, G, B channels. Further, the grayscale rendition of the color image - I - was also computed. The Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) were performed on these component images (i.e., R, G, B, I) in order to extract representational features. 33 Experimental Results 34 Experimental Results 35 The Fusion of Two User-friendly Biometric Modalities: Iris and Face 36 Fusion method Dimension reduction method : Wavelet Transform LDA 37 Experimentation Face Databases: IISFace : We sampled frontal face images of 100 subjects from the IIS face database. Each subject has 10 images with varying expressions. Iris Databases: Iris1 : This data set consists of 1000 iris images acquired from 100 individuals. (good quality images) Iris2 : The Iris2 database consists of 1000 iris images containing some bad quality ones acquired from 100 individuals. 38 Experimental Results 39 Experimentation Face Databases: ORLFace : The ORL data set consists of 400 frontal faces: 10 tightly cropped images of 40 subjects with variations in poses, illuminations, facial expressions and accessories. Iris Databases: Iris3 : The Iris3 database is composed of 400 good quality images sampled from the Iris1 database to combine with the ORLFace database. Iris4 : The Iris4 is composed of 400 iris images containing some bad quality ones sampled from the Iris2 database to combine with the ORLFace. 40 Experimental Results 41 Reference Paper M. Indovina, U. Uludag, R. Snelick, A. Mink and A. Jain, "Multimodal Biometric Authentication Methods: A COTS Approach", Proc. MMUA 2003, Workshop on Multimodal User Authentication, pp. 99-106, Santa Barbara, CA, December 11-12, 2003. 42 abstract We examine the performance of multimodal biometric authentication systems using Commercial Off-the-Shelf (COTS) fingerprint and face biometrics. It introduce novel methods of fusion and normalization that improve accuracy still further through population analysis. 43 Normalization methods a matcher score as s from the set S of all scores for that matcher and the corresponding normalized score as n. Min-Max maps the scores to the [0, 1] range. 44 Normalization methods Z-score transforms the scores to a distribution with mean of 0 and standard deviation of 1. Tanh robust statistical techniques. maps the scores to the (0, 1) range. 45 Normalization methods Adaptive Using an adaptive normalization procedure that aims to increase the separation of the genuine and impostor distributions. 46 Normalization methods Two-Quadrics composed of 2 quadratic segments that change concavity at c. 47 Normalization methods Logistic logistic function The general shape of the curve is similar to Two-Quadrics f(0) is equal to the constant Δ, which is selected to be a small value (0.01 in this study). 48 Normalization methods Quadric-Line-Quadric The overlapped zone, w, is left unchanged while the other regions are mapped with two quadratic function segments. 49 Fusion methods Simple Sum Min Score Scores for an individual are summed. Choose the minimum of an individual’s scores. Max Score Choose the maximum of an individual’s scores. 50 Fusion methods Matcher Weighting Matcher weighting-based fusion makes use of the Equal Error Rate (EER). the weights are inversely proportional to the corresponding errors. The fused score is calculated as 51 Fusion methods EER if the score distributions overlap, the FAR and FRR intersect at a certain point. The value of the FAR and the FRR at this point, which is of course the same for both of them, is called the Equal Error Rate (EER). 52 Fusion methods User Weighting This method applies weights to individual matchers differently for every user (individual). The calculation of these user-dependent weights make use of the wolf-lamb concept introduced by Doddington, et al. 53 Fusion methods We assume that for every (i, m) pair, the mean and standard deviation of the associated genuine and impostor distributions are known. use the d-prime metric as a measure of the separation of these two distributions. 54 Experiment Results The best EER values in individual columns are indicated with bold typeface; the best EER values in individual rows are indicated with a star (*) symbol. 55 Experiment Results Normalization: varied, fusion: Simple Sum 56 Experiment Results Normalization: varied, fusion: Min Score 57 Experiment Results Normalization: varied, fusion: Max Score 58 Experiment Results Normalization: varied, fusion: Matcher weight 59 Experiment Results Normalization: varied, fusion: User weight 60 Experiment Results Normalization: Min-Max, fusion: varied 61 Experiment Results Normalization: Z-score, fusion: varied 62 Experiment Results Normalization: Tanh, fusion: varied 63 Experiment Results Normalization: Quadric-Line-Quadric, fusion: varied 64 Experiment Results Normalization: Quadric-Line-Quadric, fusion: Simple Sum 65 Reference Paper Y. Wang, T. Tan and A. K. Jain, "Combining Face and Iris Biometrics for Identity Verification", Proc. of 4th Int'l Conf. on Audio- and Video-Based Biometric Person Authentication (AVBPA), pp. 805-813, Guildford, UK, June 9-11, 2003. 66 abstract Combining Face and Iris to improve verification performance. Fusing in scores level. We use two different strategies. The first strategy is to compute either an unweighted or weighted sum . The second strategy is to treat the matching distances of face and iris classifiers as a twodimensional feature vector and to use a classifier such as Fisher’s discriminant analysis. 67 Fusion methods Weighted Sum Rule 68 Fusion methods Fisher Discriminant Analysis we treat the face and iris matcher outputs x1 and x2 as a feature vector X = (x1 ,x2). Decision Rule 69 Experiment Results Distribution of matching distances 70 Experiment Results Total error rate 71 Reference Paper Michalis Petrakos, Jon Atli Benediktsson, and Ioannis Kanellopoulos,” The Effect of Classifier Agreement on the Accuracy of the Combined Classifier in Decision Level Fusion”, IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 39, NO. 11, NOVEMBER 2001 72 abstract Agreement among classifiers can inhibit the gains obtained regardless of the method used to combine them. In this work, the level of agreement between different classifiers used in remote sensing is assessed based on statistical measures. A study is performed in which an image is classified by several methods with different degrees of agreement between them. As stated previously, the LOP and the LOGP are widely used decision fusion approaches. 73 Fusion methods LOP (Linear Opinion Pool) LOGP (Logarithmic Opinion Pool) 74 Fusion methods Then we can combine LOP/LOGP and neural network. Their combination schemes were considered to be two stage processes: statistical classifiers in stage one, and a single neural network in stage two. 75 Fusion methods Their suggested voting schemes were: Majority Voting: When the majority of the individual agree on the classification of a sample, the sample is classified to that class. Complete Agreement: When all the individual source-specific classifiers agree on the classification of a sample, the sample is classified to that class. CONSNN-NN: A neural network and a LOGP classifier are trained separately on all the data. 76 Experiment Results Database 77 Experiment Results The minimum distance (MD) computes the Mahalanobis distance with every one of the 12 land-use classes for each sample, and assigns the sample to the class with the smallest distance. Linear discriminant analysis (LDA) assigns a sample to the class with the maximum posterior probability given that all classes have a common (pooled) covariance matrix. Quadratic discriminant analysis (QDA) assigns a sample to the class with the maximum posterior probability. In this case, each class has its own covariance matrix. A conjugate gradient neural network (CGNN) with 13 inputs, one hidden layer containing 20 neurons, and 12 outputs. 78 Experiment Results 79 Experiment Results 80 Conclusion Some fusion methods are the best in a biometric system, but they maybe not the best in other biometric system. The effects of different score normalization techniques in a multimodal biometric system have been studied. 81