Open World Face Recognition with Credibility and Confidence Measures

Similar documents
Cross-Over Analysis Using T-Tests

Detecting Possibly Fraudulent or Error-Prone Survey Data Using Benford s Law

Data Center Power System Reliability Beyond the 9 s: A Practical Approach

A New Evaluation Measure for Information Retrieval Systems

10.2 Systems of Linear Equations: Matrices

How To Segmentate An Insurance Customer In An Insurance Business

A Generalization of Sauer s Lemma to Classes of Large-Margin Functions

Stock Market Value Prediction Using Neural Networks

On Adaboost and Optimal Betting Strategies

View Synthesis by Image Mapping and Interpolation

Ch 10. Arithmetic Average Options and Asian Opitons

MSc. Econ: MATHEMATICAL STATISTICS, 1995 MAXIMUM-LIKELIHOOD ESTIMATION

The one-year non-life insurance risk

Optimal Energy Commitments with Storage and Intermittent Supply

Calibration of the broad band UV Radiometer

State of Louisiana Office of Information Technology. Change Management Plan

Modelling and Resolving Software Dependencies

A Data Placement Strategy in Scientific Cloud Workflows

A New Pricing Model for Competitive Telecommunications Services Using Congestion Discounts

Firewall Design: Consistency, Completeness, and Compactness

JON HOLTAN. if P&C Insurance Ltd., Oslo, Norway ABSTRACT

A Comparison of Performance Measures for Online Algorithms

Minimizing Makespan in Flow Shop Scheduling Using a Network Approach

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.436J/15.085J Fall 2008 Lecture 14 10/27/2008 MOMENT GENERATING FUNCTIONS

A SPATIAL UNIT LEVEL MODEL FOR SMALL AREA ESTIMATION

Forecasting and Staffing Call Centers with Multiple Interdependent Uncertain Arrival Streams

INFLUENCE OF GPS TECHNOLOGY ON COST CONTROL AND MAINTENANCE OF VEHICLES

HOST SELECTION METHODOLOGY IN CLOUD COMPUTING ENVIRONMENT

An intertemporal model of the real exchange rate, stock market, and international debt dynamics: policy simulations

! # % & ( ) +,,),. / % ( 345 6, & & & &&3 6

Improving Direct Marketing Profitability with Neural Networks

Optimal Control Policy of a Production and Inventory System for multi-product in Segmented Market

Professional Level Options Module, Paper P4(SGP)

CALCULATION INSTRUCTIONS

ThroughputScheduler: Learning to Schedule on Heterogeneous Hadoop Clusters

Supporting Adaptive Workflows in Advanced Application Environments

How To Find Out How To Calculate Volume Of A Sphere

Achieving quality audio testing for mobile phones

Measures of distance between samples: Euclidean

S&P Systematic Global Macro Index (S&P SGMI) Methodology

Feedback linearization control of a two-link robot using a multi-crossover genetic algorithm

MODELLING OF TWO STRATEGIES IN INVENTORY CONTROL SYSTEM WITH RANDOM LEAD TIME AND DEMAND

Unsteady Flow Visualization by Animating Evenly-Spaced Streamlines

USING SIMPLIFIED DISCRETE-EVENT SIMULATION MODELS FOR HEALTH CARE APPLICATIONS

Hull, Chapter 11 + Sections 17.1 and 17.2 Additional reference: John Cox and Mark Rubinstein, Options Markets, Chapter 5

Cost Efficient Datacenter Selection for Cloud Services

Risk Adjustment for Poker Players

A Theory of Exchange Rates and the Term Structure of Interest Rates

Performance And Analysis Of Risk Assessment Methodologies In Information Security

A secure face tracking system

Modelling football match results and the efficiency of fixed-odds betting

Class-specific Sparse Coding for Learning of Object Representations

TO DETERMINE THE SHELF LIFE OF IBUPROFEN SOLUTION

BOSCH. CAN Specification. Version , Robert Bosch GmbH, Postfach , D Stuttgart

Modeling and Predicting Popularity Dynamics via Reinforced Poisson Processes

An Alternative Approach of Operating a Passive RFID Device Embedded on Metallic Implants

Minimum-Energy Broadcast in All-Wireless Networks: NP-Completeness and Distribution Issues

Hybrid Model Predictive Control Applied to Production-Inventory Systems

Math , Fall 2012: HW 1 Solutions

How To Understand The Structure Of A Can (Can)

Definition of the spin current: The angular spin current and its physical consequences

A Blame-Based Approach to Generating Proposals for Handling Inconsistency in Software Requirements

Heat-And-Mass Transfer Relationship to Determine Shear Stress in Tubular Membrane Systems Ratkovich, Nicolas Rios; Nopens, Ingmar

RUNESTONE, an International Student Collaboration Project

Adaptive Face Recognition System from Myanmar NRC Card

Unbalanced Power Flow Analysis in a Micro Grid

An introduction to the Red Cross Red Crescent s Learning platform and how to adopt it

Automatic Long-Term Loudness and Dynamics Matching

FAST JOINING AND REPAIRING OF SANDWICH MATERIALS WITH DETACHABLE MECHANICAL CONNECTION TECHNOLOGY

Using research evidence in mental health: user-rating and focus group study of clinicians preferences for a new clinical question-answering service

Option Pricing for Inventory Management and Control

Product Differentiation for Software-as-a-Service Providers

Seeing the Unseen: Revealing Mobile Malware Hidden Communications via Energy Consumption and Artificial Intelligence

Dynamic Network Security Deployment Under Partial Information

ISSN: ISO 9001:2008 Certified International Journal of Engineering and Innovative Technology (IJEIT) Volume 3, Issue 12, June 2014

A Monte Carlo Simulation of Multivariate General

Face Recognition in Low-resolution Images by Using Local Zernike Moments

Learning using Large Datasets

GPRS performance estimation in GSM circuit switched services and GPRS shared resource systems *

Mathematics Review for Economists

Gender Differences in Educational Attainment: The Case of University Students in England and Wales

Transcription:

Open Worl Face Recognition with Creibility an Confience Measures Fayin Li an Harry Wechsler Department of Computer Science George Mason University Fairfax, VA 22030 {fli, wechsler}@cs.gmu.eu Abstract. his paper escribes a novel framework for the Open Worl face recognition problem, where one has to provie for the Reject option. Base upon algorithmic ranomness an transuction, a particular form of inuction, we escribe the CM-kNN (ransuction Confience Machine knearest Neighbor) algorithm for Open Worl face recognition. he algorithm propose performs much better than PCA an is comparable with Fisherfaces. In aition to recognition an rejection, the algorithm can assign creibility ( likelihoo ) an confience ( lack of ambiguity ) measures with the ientification ecisions taken.. Introuction he choices facing face recognition systems shoul inclue: ACCEP, REJEC ( is not here ), an AMBIGUI ( nee more information ). he inclusion of the REJEC option, which correspons to an open worl of (face recognition) hypotheses, as complexity to the whole process an makes face recognition much harer compare to the more traitional close worl biometric systems available toay. In aition to seeking how similar or close some probe face image is to each subject in the face gallery set, one nees some measure of confience when making any ientification ecision. his paper escribes a novel methoology for hanling an open worl of hypotheses, incluing the REJEC option, an provies the means to associate creibility an confience measures with each of the ecisions mae regaring HumanID. he propose methoology, base upon ranomness concepts an transuctive learning, is formally valiate on challenging (varying illumination) an large overlapping FERE ata sets. 2. Ranomness an p-values Confience measures can be base upon universal tests for ranomness, or their approximation. A Martin-Lof ranomness eficiency (Li an Vitanyi, 997) base on such tests is a universal version of the stanar statistical notion of p-values. Universal tests for

ranomness are not computable an hence one has to approximate the p-values using nonuniversal tests. We use the p-value construction in Proerou et al. (200) to efine the quality of information. he assumption use is that ata items are inepenent an are prouce by the same stochastic mechanism. Given a sequence of proximities (istances) between the given training (gallery) set an an unknown sample (test) probe, one quantifies to what extent the (classification) ecision taken is reliable, i.e., non-ranom. owars that en one efines the strangeness of the unknown sample probe i with putative label y in relation to the rest of the training set exemplars as: D k j k y ¹ j D ij D y ij ¹ he strangeness measure is the ratio of the sum of the k nearest istances D from the same class (y) to the sum of the k nearest istances from all other classes ( y). he strangeness of an exemplar increases when the istance from the exemplars of the same class becomes larger an when the istance from the other classes becomes smaller. A vali ranomness test (Nouretinov et al., 200) efines then the p-value measure of a test exemplar with a possible classification assigne to it as p f ( D ) f ( D 2) ( m f ( ) f ( D D new m) ) f ( Dnew ) where f is some monotonic non-ecreasing function with f(0) = 0, e.g., f(d) = D, m is the number of training examples, an Dnew is the strangeness measure of a new potential test probe exemplar c new. An alternative efinition available for the p-value is p( c ) #{ i : D t D }/( m new i new ). Using the p-value one can now preict the class membership as the one that yiels the largest p-value, which is efine as the creibility of the assignment mae. he associate confience measure, which is one minus the 2n largest p-value, inicates how close the first two assignments are. he confience value inicates how improbable the classifications other than the preicte classification are an the creibility value shows how suitable the training set is for the classification of that testing example. One can compare the top ranke assignments, rather than only the first two assignments, an efine aitional confience criteria. Both the creibility an confience measures allow the face recognition moule to aapt to existing conitions an act accoringly. 3. ransuction Confience Machine (CM)- knn Another form of learning, beyon inuction, is transuction. Given an unlabele valiation test, in aition to the training set, the task now is to estimate the class for each unlabele

pattern in orer to construct the best classifier rule for both the training an valiation sets. CM-kNN Algorithm for i = to m y y Fin an store D i an D i Calculate the alpha strangeness values for all the training exemplars Calculate the similarity ist vector as the istances of the new exemplar from all the training exemplars for j = to c o for every training exemplar t classifie as j o if D > ist(t), i = }k, recalculate the alpha value of exemplar t j ti for every training exemplar t classifie as non-j o if D ti j > ist(t), i = }k, recalculate the alpha value of exemplar t Calculate alpha value for the new exemplar classifie as j Calculate p-value for the new exemplar classifie as j Preict the class with the largest p-value Output as confience one minus the 2n largest p-value Output as creibility the largest p-value he constraints on the (geometric) layout of the learning space an the search for improve classification margins are aresse in this paper using algorithmic ranomness (Vovk et al., 999), universal measures of confience ranomness (Vovk et al., 999), an transuctive confience (learning) machines (CM) (Proerou et al., 200). he experimental ata presente later on that valiates our approach, is base on CM-kNN which is an augmente CM using locality-base evience, e.g., the k-nearest Neighbors (knn) concept. he similarity istances ist (in script) use are shown next. Given two n-imensional n vectors, ƒ, the istance measures use are efine as follows: n 2 L (, ) i i L2 (, ) ( ) ( ) i cos (, ) Dice 2 (, ) 2 2 2

Jaccar Mah L (, ) 2 2 2(, ) ( ) 6 ( ) Mah 6 cos (, ) where 6 is the scatter matrix of the training ata. For PCA, 6 is iagonal an the iagonal elements are the (eigenvalues) variances of the corresponing components. he Mahalanobis + L istance efine only for PCA is n i i Mah L (, ) 4. Data Collection i i Figure. Face Images Our ata set is rawn from the FERE atabase, which has become a e facto stanar for evaluating face recognition technologies (Phillips et al., 998). he ata set consists of 600 FERE frontal face images corresponing to 200 subjects, which were acquire uner variable illumination an facial expressions. Each subject has three images of size 256x384 with 256 gray scale levels. Face image normalization is carrie out as follows: first, the centers of the eyes of an image are manually etecte, then rotation an scaling transformations align the centers of the eyes to preefine locations, an finally, the face image is croppe to the size of 28x28 to extract the facial region. he extracte facial region is further normalize to zero mean an unit variance. Fig. shows some exemplar images use in our experiments that are alreay croppe to the size of 28x28. Each

column in Fig. correspons to one subject. Note that for each subject, two images are ranomly chosen for training, while the remaining image (unseen uring training) is use for testing. he normalize face images are processe to yiel 400 PCA coefficients, accoring to eqs. 7 9 from Liu an Wechsler (2002), an 200 Fisherfaces using FLD (Fisher Linear Discriminant), accoring to eqs. 0 2 from Liu an Wechsler (2002) on a reuce 200 imensional space PCA space. 5. Open Worl Face Recognition Algorithms Open Worl CM-kNN Algorithm Calculate the alpha values for all the training exemplars for i = to c o for every training exemplar t classifie as i o for j = to c an j!= i o Assume t is classifie as j, which shoul be rejecte Recalculate the alpha value for all the training exemplars classifie as non-j Calculate alpha value for the exemplar t classifie as j Calculate p-value for the exemplar t classifie as j Calculate the P max, P mean an P stev (stanar eviation) for the p-value of exemplar t Calculate the PSR value for exemplar t: PSR = (P max P mean )/P stev Calculate the mean, stev (stanar eviation) for all the PSR values Calculate the mean + 3*stev as threshol for rejection Calculate the istances of the probe exemplar from all the training exemplars for i = to c o Calculate alpha value for the probe exemplar classifie as i Calculate p-value for the probe exemplar classifie as i Calculate the largest p-value max for the probe exemplar Calculate the mean an stev for the probe p-value without max Calculate the PSR value for the probe exemplar: PSR = (max mean)/ stev Reject the probe exemplar if its PSR is less than or equal to the threshol. Otherwise preict the class with the largest p-value

Open Worl {PCA, Fisherfaces} Algorithm for i = to m Fin the maximum intra-within-istance an minimum inter-between-istance Calculate the mean an stanar eviation for all maximum intra-istances an minimum inter-istances: mean intra, mean inter, stev intra an stev inter Calculate mean intra + 3* stev intra as the lower boun of the threshol Calculate mean inter - 3* stev inter as the upper boun of the threshol Choose the threshol base on the lower an upper boun Calculate the istances of the probe exemplar from all the training exemplars Fin the minimum istance ist min of the probe exemplar If ist min >= threshol, then reject the probe exemplar Else preict the class with the minimum istance ist min 6. Experimental Results We foun that the best similarity istances for PCA an Fisherfaces are {Mahalanobis + (L, L 2 or cos)} an {cosine, Dice, Jaccar, (Mahalonobis + cos)}, respectively. hose istances are use in our experiments. he experiments reporte were carrie out on the ata escribe in the previous section. Both the gallery an the probe sets consist of 00 subjects, an the overlap portion between the two sets on the average 50 subjects. he recognition rate is the percentage of the subjects whose probe is correctly recognize or rejecte. 0.8 0.75 0.9 0.7 0.8 Recognition Rate 0.65 0.6 0.55 Recognition Rate 0.7 0.6 0.5 0.5 0.45 0.5 0.6 0.7 0.8 0.9..2.3.4 hreshol 0.4-0.9-0.85-0.8-0.75-0.7-0.65-0.6-0.55-0.5-0.45 hreshol Figure 2. he Recognition Rate vs hreshol: PCA (Left) an Fisherfaces (Right) Open Worl (PCA an Fisherfaces) Face Recognition he ata was ranomly chosen, the same experiment was run 00 times, an Fig. 2 shows the mean recognition rates for ifferent threshols. he istance measurements for PCA an Fisherfaces, which yiel the best results, are Mahalanobis + L 2 an cosine, respectively. Fig. 2 shows that the best recognition rate for PCA is 77% if the threshol

can be chosen correctly, while for Fisherfaces is 9% if the threshol is chosen as its upper boun. he stanar eviation for the best recognition rate for PCA an Fisherfaces are 3.7% an 2.4%, respectively. 0.65 0.6 correct rejection correct recognition false recognition 0.08 correct rejection correct recognition 0.55 0.07 wrong recognition 0.5 0.06 p-value 0.45 0.4 p-value 0.05 0.04 0.35 0.03 0.3 0.25 0.02 0.2 0 0 20 30 40 50 60 70 80 90 00 classes 0.0 0 0 20 30 40 50 60 70 80 90 00 classes Figure 3. est p-value istribution of rejection, correct an false recognition using PCA with (Mahalanobis + L2) istance (Left) an Fisherfaces with cosine istance (Right) PCA PSR histogram with Mahalanobis + L istance Fisherface PSR histogram with cos istance 20 90 80 00 70 80 60 50 60 40 40 30 20 20 0 0 2 3 4 5 6 7 8 9 0 0 2 4 6 8 0 2 4 Figure 4. he PSR value histogram: PCA (Left) an Fisherfaces (Right) CM-kNN he ata use are either the (400) PCA or (200) Fisherface components, an k =. he threshol is compute accoring to the algorithm escribe in Sect. 5 base on the training exemplars. he p-value istributions shown in Fig.3 inicate that the test PSR values are useful for rejection an recognition. Recognition is riven by large PSR values. he best recognition rate using PCA components is 87.87% using the Mahalanobis + L istance, an its stanar eviation is 3.0%. he threshol is 6.57 compute from the PSR histogram shown in Fig. 4 (left). he best recognition rate using Fisherface components is 90% using the cosine istance, an its stanar eviation is 2.7%. he threshol is 9.20 compute from the PSR histogram shown in Fig. 4 (right). CM-kNN provies aitional information regaring the creibility an confience in the recognition ecision taken. he corresponing 2D istribution for correct an false

recognition is shown in Fig. 5, where one can see that false recognition, for both the PCA an Fisherfaces components, shows up at low values. correct recognition false recognition 0.35 correct recognition false recognition 0.9 0.3 0.25 0.8 Creibility 0.7 Creibility 0.2 0.5 0.6 0. 0.5 0.05 0.4 0.6 0.65 0.7 0.75 0.8 0.85 Confience 0 0.96 0.965 0.97 0.975 0.98 0.985 0.99 0.995 Confience Figure 5: Distribution of confience an Creibility: PCA (left) an Fisherfaces (right). 7. Conclusions We introuce in this paper a new face recognition algorithm suitable for open worl face recognition. he feasibility an usefulness of the algorithm has been shown on varying illumination an facial expression images rawn from FERE. Furthermore, both creibility an confience measures are provie for both the recognition an rejection ecisions. We plan to use those measures for optimal training of the face recognition system, such that the composition an size of the training set are etermine using active rather than ranom selection 8. References. A. Gammerman, V. Vovk, an V. Vapnik (998), Learning by ransuction. In Uncertainty in Artificial Intelligence, 48 55. 2. M. Li an P. Vitanyi (997), An Introuction to Kolmogorov Complexity an Its Applications, 2e., Springer-Verlag. 3. C. Liu an H. Wechsler (2002), Gabor Feature Base Classification Using the Enhnace Fisher Linera Discriminant Moel for Face Recognition, IEEE rans. on Image Processing, Vol., No. 4, 467 476. 4. I. Nouretinov,. Melluish, an V. Vovk (200), Rige Regression Confience Machine, Proc. 7th Int. Conf. on Machine Learning. 5. P. J. Phillips, H. Wechsler, J. Huang, an P. Rauss (998), he FERE Database an Evaluation Proceure for Face Recognition Algorithms, Image an Vision Computing, Vol.6, No.5, 295-306. 6. K. Proerou, I. Nouretinov, V. Vovk an A. Gammerman (200), ransuctive Confience Machines for Pattern Recognition, R CLRC-R-0-02, Royal Holloway University of Lonon. 7. V. Vovk, A. Gammerman, an C. Sauners (999), Machine Leraning Applications of Algorithmic Ranomness, Proc. 6th Int. Conf. on Machine Learning.