FACIAL SOFT BIOMETRICS - Library of Ph.D. Theses | EURASIP
FACIAL SOFT BIOMETRICS - Library of Ph.D. Theses | EURASIP FACIAL SOFT BIOMETRICS - Library of Ph.D. Theses | EURASIP
38 3. BAG OF FACIAL SOFT BIOMETRICS FOR HUMAN IDENTIFICATIONIn other words the bound in Lemma 5 implies that, a reduction of the relative-throughput fromits maximal value of n/ρ ≈ 1 to a sufficiently smaller n/ρ ≈ 1 2, for high enough ρ, results in asubstantial and exponential reduction in the probability of interference, from P(h = 1) ≈ ρ −ρɛ toP(h = 1 2 ) ≈ ρ−ρ(5 4 +ɛ 2 ) .We have up to now focused on the interference limited scenario, where errors occur only due tomore than one subject belonging to one category. In the next section 3.5.4 we consider estimationerror and a more pragmatic way to improve the overall reliability of a SBS.3.5.4 Estimation reliabilityIn the aforementioned operational setting of interest, the reliability of an SBS captures theprobability of false identification of a randomly chosen person out of a random set of n subjects.In such a setting, the reliability of an SBS is generally related to:– the number of categories that the system can identify.– the degree with which these features/categories represent the chosen set (of subjects) overwhich identification will take place– n, where a higher n corresponds to identifying a person among an increasingly large set ofpossibly similar-looking people– robustness with which these categories can be detectedWe here proceed to study the general SBS error probability, containing inevitably all abovementioned factors including the algorithmic categorization error–probabilities. With other wordswe examine, the re–identification error probability, regardless of the underlying source, which canbe both due to misclassifcation or due to interference.Given the knowledge of the population statistics and moreover the exact algorithmic reliabilities(true detection rates and additionally the confusion probabilities), we can use a maximum–likelihood (ML) optimizing rule to compute the maximal posterior probability for each category.We note here that the ML optimizing rule for the most probable category in which a chosen subjectbelongs, is given by:ˆφ = argmax φ∈Φ P(φ)·P(y/φ), (3.27)wherey is the observation vector,P(φ) is the pdf of the set of categories over the given population(note ∑ ρß=1 P(φ i) = 1), and P(y/φ) the probability that y is observed, given that the subjectbelongs in category φ.3.5.4.1 Improving SBS reliabilityIn the most common case of a training set, which provides insufficient information on all confusionfactors as of all P(y/φ), we can find heuristic rules to combat the overall error probabilityP err . Given the large amount of empty categories, see the distribution of over-all-categories inthe FERET population in Figure 3.1, and furthermore the above presented correlations betweentraits, certain misclassifications can be identified and reconciled. An example for a heuristic errorconciliation attempt can be the following.Example 6 We take into account the given large FERET population and the SBS presented insection 3.3.1. We simulate again the same identification scenario, where here we simulate an estimationerror of 10% for the color soft biometric traits (hair, skin and eye color). When fusing thetraits on decision level (hard fusion) those errors naturally add up. That is why a soft decision,taking into account the confidence levels of the extracted features and also the reliability of the
39Table 3.4: Example for a heuristic rule. SBS endowed with ρ = 4 categories and a given knownpopulation (distribution in the 4 categories). If our SBS estimates category φ 2 for a subject tobelong into, due to the 0 probability of occurrence, the system decides for the next probablecategory, namely φ 1 .Category µ 1 µ 2 Probability for occurrence of φ i given nφ 1 0 0 0.5φ 2 0 1 0φ 3 1 0 0.3φ 4 1 1 0.2underlying algorithm is used, since it will discriminate some error cases. In the classificationstep we can easily identify subjects classified into "empty" categories. Since those categories areof probability 0 to occur, due to the known population, an intelligent classification system recognizesthese cases and reclassifies those subjects in "similar" categories with higher probabilitiesof occurrence(=non–empty categories). A "similar" category hereby is a category with highestprobability for misclassification, given the wrongly detected category. We here assume that anerror caused by misclassification of one trait is more probable than the misclassification of twoor more traits. For visualization see Table 3.4: if a subject is classified into the category φ 2 , thesystem recognizes thatφ 2 is an empty category and searches for the next probable category, whichin this case would beφ 1 . This simple heuristic rule leads already to a significant reconciliation ofthe added up estimation error of the SBS, see Figure 3.7.0.80.7Interference limited errrorCompensated ErrorError due to Interference and estimation reliability0.60.5Perr0.40.30.20.12 4 6 8 10 12 14 16Subjects NFigure 3.7: Errors in a SBS system: interference limited error, estimation reliability error, compensatederror.In the following we outline an additional error, which can be associated to SBSs when used ina scenario, where the traits are re–identified based on a human description.3.5.4.2 Human machine interactionThe immense benefit of having human understandable and compliant soft biometric traits overclassical biometrics, enables a computer aided biometric search to have as an input a human descriptionof the target subject. The importance of related applications becomes evident in cases,such as a loss of a child in a mall, where the mother can just provide a description of the child andcomputer based search can be performed on available security video material. Along with the benefitof human compliance come though additional quantification and human-machine interactionerrors. Such errors can have different causes.
- Page 1: FACIAL SOFT BIOMETRICSMETHODS, APPL
- Page 5: AcknowledgementsThis thesis would n
- Page 8: 6hair, skin and clothes. The propos
- Page 11 and 12: 97 Practical implementation of soft
- Page 13 and 14: 11Notations used in this workE : st
- Page 15 and 16: 13Chapter 1IntroductionTraditional
- Page 17 and 18: 15event of collision, which is of s
- Page 19 and 20: 17ric. In Section 6.6 we employ the
- Page 21 and 22: 19Chapter 2Soft biometrics: charact
- Page 23 and 24: 21is the fusion of soft biometrics
- Page 25 and 26: 23plied on low resolution grey scal
- Page 27 and 28: 25Chapter 3Bag of facial soft biome
- Page 29 and 30: 27In this setting we clearly assign
- Page 31 and 32: 29Table 3.1: SBSs with symmetric tr
- Page 33 and 34: 31corresponding to p(n,ρ). Towards
- Page 35 and 36: the same category (all subjects in
- Page 37 and 38: 3.5.2 Analysis of interference patt
- Page 39: an SBS by increasing ρ, then what
- Page 43 and 44: 41for a given randomly chosen authe
- Page 45 and 46: 43Chapter 4Search pruning in video
- Page 47 and 48: 45Figure 4.1: System overview.SBS m
- Page 49 and 50: 472.52rate of decay of P(τ)1.510.5
- Page 51 and 52: 49to be the probability that the al
- Page 53 and 54: 51The following lemma describes the
- Page 55 and 56: 534.5.1 Typical behavior: average g
- Page 57 and 58: 55n = 50 subjects, out of which we
- Page 59 and 60: 5710.950.9pruning Gain r(vt)0.850.8
- Page 61 and 62: 59for one person, for trait t, t =
- Page 63 and 64: 61Chapter 5Frontal-to-side person r
- Page 65 and 66: 63Figure 5.1: Frontal / gallery and
- Page 67 and 68: 6510.90.80.7Skin colorHair colorShi
- Page 69 and 70: 6710.90.80.70.6Perr0.50.40.30.20.10
- Page 71 and 72: 69Chapter 6Soft biometrics for quan
- Page 73 and 74: 71raphy considerations include [BSS
- Page 75 and 76: 73Figure 6.3: Example image of the
- Page 77 and 78: 75A direct way to find a relationsh
- Page 79 and 80: 77- Pearson’s correlation coeffic
- Page 81 and 82: 79shown to have a high impact on ou
- Page 83 and 84: 81Chapter 7Practical implementation
- Page 85 and 86: 834) Eye glasses detection: Towards
- Page 87 and 88: 857.2 Eye color as a soft biometric
- Page 89 and 90: 87Table 7.5: GMM eye color results
38 3. BAG OF <strong>FACIAL</strong> <strong>SOFT</strong> <strong>BIOMETRICS</strong> FOR HUMAN IDENTIFICATIONIn other words the bound in Lemma 5 implies that, a reduction <strong>of</strong> the relative-throughput fromits maximal value <strong>of</strong> n/ρ ≈ 1 to a sufficiently smaller n/ρ ≈ 1 2, for high enough ρ, results in asubstantial and exponential reduction in the probability <strong>of</strong> interference, from P(h = 1) ≈ ρ −ρɛ toP(h = 1 2 ) ≈ ρ−ρ(5 4 +ɛ 2 ) .We have up to now focused on the interference limited scenario, where errors occur only due tomore than one subject belonging to one category. In the next section 3.5.4 we consider estimationerror and a more pragmatic way to improve the overall reliability <strong>of</strong> a SBS.3.5.4 Estimation reliabilityIn the aforementioned operational setting <strong>of</strong> interest, the reliability <strong>of</strong> an SBS captures theprobability <strong>of</strong> false identification <strong>of</strong> a randomly chosen person out <strong>of</strong> a random set <strong>of</strong> n subjects.In such a setting, the reliability <strong>of</strong> an SBS is generally related to:– the number <strong>of</strong> categories that the system can identify.– the degree with which these features/categories represent the chosen set (<strong>of</strong> subjects) overwhich identification will take place– n, where a higher n corresponds to identifying a person among an increasingly large set <strong>of</strong>possibly similar-looking people– robustness with which these categories can be detectedWe here proceed to study the general SBS error probability, containing inevitably all abovementioned factors including the algorithmic categorization error–probabilities. With other wordswe examine, the re–identification error probability, regardless <strong>of</strong> the underlying source, which canbe both due to misclassifcation or due to interference.Given the knowledge <strong>of</strong> the population statistics and moreover the exact algorithmic reliabilities(true detection rates and additionally the confusion probabilities), we can use a maximum–likelihood (ML) optimizing rule to compute the maximal posterior probability for each category.We note here that the ML optimizing rule for the most probable category in which a chosen subjectbelongs, is given by:ˆφ = argmax φ∈Φ P(φ)·P(y/φ), (3.27)wherey is the observation vector,P(φ) is the pdf <strong>of</strong> the set <strong>of</strong> categories over the given population(note ∑ ρß=1 P(φ i) = 1), and P(y/φ) the probability that y is observed, given that the subjectbelongs in category φ.3.5.4.1 Improving SBS reliabilityIn the most common case <strong>of</strong> a training set, which provides insufficient information on all confusionfactors as <strong>of</strong> all P(y/φ), we can find heuristic rules to combat the overall error probabilityP err . Given the large amount <strong>of</strong> empty categories, see the distribution <strong>of</strong> over-all-categories inthe FERET population in Figure 3.1, and furthermore the above presented correlations betweentraits, certain misclassifications can be identified and reconciled. An example for a heuristic errorconciliation attempt can be the following.Example 6 We take into account the given large FERET population and the SBS presented insection 3.3.1. We simulate again the same identification scenario, where here we simulate an estimationerror <strong>of</strong> 10% for the color s<strong>of</strong>t biometric traits (hair, skin and eye color). When fusing thetraits on decision level (hard fusion) those errors naturally add up. That is why a s<strong>of</strong>t decision,taking into account the confidence levels <strong>of</strong> the extracted features and also the reliability <strong>of</strong> the