Performance of Fingernail Plates in Identification Systems

For identification systems, the performance may be evaluated using the parameter True Positive Identification Rate (TPIR). TPIR is the proportion of times the identity determined by the system is actually the true identity of the person who is providing the biometric sample [37]. If the biometric system provides the identities of the top x matches, the Rank-a TPIR, Rx,is the proportion of times the true identity of the individual is contained in the top x matching identities. For performance analysis of an identification system, the TPIR at various ranks may be depicted in the Cumulative Match Characteristics (CMC) curve where TPIR, Rx, are plotted against Rank x = 1,2 ,...N where N is the total number of people enrolled in the database.

Performance of Fingernail Plates in Unimodal Identification Systems

In this section, performances of the three fingernail plates in unimodal identification systems have been analysed for each of the three deep learning models - TLA, TLR, and TLD. Figure 9.13 makes a comparison of the performance of the index, middle, and ring fingernail plate using the TLA feature-set. Similar comparative performance portrayal of the three fingernail plates are depicted in Figures 9.14 and 9.15 for the TLR and TLD feature-sets, respectively.

It is observed from Figures 9.13-9.15 that identification systems built from either the index or the middle fingernail plate shall provide good performance. The ring fingernail plate also fares well, but lags in performance behind the other two nail plates. Analysing the results depicted in these three figures also shows that for all the three nail plates, the best identification accuracy is given by TLD, followed by TLR and TLA, in that sequence. This, of course, is because of the depth of the respective models. Also, for all the three nail plates, TLD provides a high identification accuracy of above 93%. The other two models, TLA and TLD, also fare reasonably well.

CMCs comparing the Performance of Index, Middle, and Ring Fingernail Plates in Unimodal Identification Systems using TLA as the feature extraction technique

FIGURE 9.13 CMCs comparing the Performance of Index, Middle, and Ring Fingernail Plates in Unimodal Identification Systems using TLA as the feature extraction technique.

CMCs comparing the Performance of Index, Middle, and Ring Fingernail

FIGURE 9.14 CMCs comparing the Performance of Index, Middle, and Ring Fingernail

Plates in Unimodal Identification Systems using TLR as the feature extraction technique.

CMCs comparing the Performance of Index, Middle, and Ring Fingernail Plates in Unimodal Identification Systems using TLD as the feature extraction technique

FIGURE 9.15 CMCs comparing the Performance of Index, Middle, and Ring Fingernail Plates in Unimodal Identification Systems using TLD as the feature extraction technique.

Table 9.4 tabulates the values of Rank-l TPIRs obtained from all the three nail plates from the three deep learning models considered.

The lowest Rank-1 TPIR is obtained from the index nail plate when the TLA feature-set is used; and that is as high as 85.39%. Table 9.4 also demonstrates that the best Rank-1 identification performance is provided by the middle fingernail plate for TLA, by the index and the middle fingernail plates for TLR, and by index fingernail plate for TLD.

Э.6.2.2 Performance of Fingernail Plates in Multimodal Identification Systems

In this section, the index, middle, and ring fingernail plates have been subjected to different frameworks of rank-level fusion. For all the experiments conducted, the performance of rank-level fusion of these three nail plates has been analysed for all the three deep learning models: TLA, TLR, and TLD.

TABLE 9.4

Rank-1 TPIR (in %) for Various Unimodal Identification Biometric Systems built using Index, Middle, and Ring Fingernail Plates-1

Model

Trait

Index

Middle

Ring

TLA

85.39

90.45

85.95

TLR

91.01

91.01

90.45

TLD

93.82

93.26

93.26

a Numbers in bold and italics signify the best performance across one or the other parameter.

9.6.2.2.1 Experiment A

For the first set of experiments under this section, the three ranking lists of all the three fingernail plates have been fused separately for the three models: TLA, TLR, and TLD. In this experiment, fusion has been performed using two different linear, weighted fusion methods, namely the Logistic Regression and the Mixed Group Rank. Here, the weights have been chosen via empirical computation. Figure 9.16 compares the performance of the nail plates after rank-level fusion through Logistic Regression and Mixed Group Rank, where the TLA feature-set is used. Similar comparative performances of rank-level fusion for the TLR and TLD models are shown in Figures 9.17 and 9.18, respectively.

CMCs after Rank-Level Fusion of all three Fingernail Plates for TLA feature-set using Logistic Regression and Mixed Group Rank, where all weights have been computed empirically

FIGURE 9.16 CMCs after Rank-Level Fusion of all three Fingernail Plates for TLA feature-set using Logistic Regression and Mixed Group Rank, where all weights have been computed empirically.

CMCs after Rank-Level Fusion of all three Fingernail Plates for TLR feature- set using Logistic Regression and Mixed Group Rank, where all weights have been computed empirically

FIGURE 9.17 CMCs after Rank-Level Fusion of all three Fingernail Plates for TLR feature- set using Logistic Regression and Mixed Group Rank, where all weights have been computed empirically.

CMCs after Rank-Level Fusion of all three Fingernail Plates for TLD feature-set using Logistic Regression and Mixed Group Rank, where all weights have been computed empirically

FIGURE 9.18 CMCs after Rank-Level Fusion of all three Fingernail Plates for TLD feature-set using Logistic Regression and Mixed Group Rank, where all weights have been computed empirically.

Figures 9.16-9.18 establish that an identification system built from the index, middle, and ring fingernail plates gives good performance accuracy, where the TLD- based results are better than those based on TLR and TLA. Analyses of the results depicted in Figures 9.16-9.18 also bring out an interesting point. The corresponding Rank-1 TPIR values are the same for all three models, when Logistic Regression and Mixed Group Rank are the chosen fusion rules. However, with increasing ranks, the TPIR values increase in two different trends for the two fusion methods. While for all three deep learning models, 100% TPIR is achieved at Rank-12 for the Mixed Group Rank rule, the same is achieved by using the Logistic Regression method only at the final rank, i.e., Rank 178.

A tabular depiction of the TPIR values at a few' selected ranks is given in Table 9.5 to illustrate the aforementioned trend of results.

9.6.2.2.2 Experiment В

The weights for the three different fingernail plates in Experiment A have been assigned through empirical computation. With an aim to further improve accuracy, the exact experiments under Experiment A have been repeated by computing weights using PSO. This has been done to ensure optimal weight assignment to the three nail plates, and thus to provide better results. In Figure 9.19 the performance of the nail plates is depicted when their corresponding TLA feature-sets are subjected to rank- level fusion using Logistic Regression and Mixed Group Rank methods. Similar performances are depicted in Figures 9.20 and 9.21 w'hen TLR and TLD feature- sets are used. All the three aforementioned figures show that w'hen PSO is used for the computation of weights, the results improve for each of the cases, wuth the best Rank-1 accuracy being provided by the TLD based ranking lists for the Mixed Group Rank rule.

TABLE 9.5

True Positive Identification Rates (in %) of the Multimodal Systems where Nail Plates of All Three Fingers are fused using Two Linear Weighted Fusion Methods (Weights Computed Empirically)

Logistic Regression Method

Rank Model Ф

1

2

3

8

178

TLA

94.38

96.63

98.88

99.44

100

TLR

96.07

97.75

99.44

99.44

100

TLD

97.19

98.32

99.44

99.44

100

Mixed Group Rank Method

Rank Model Ф

1

2

3

8

12

TLA

94.38

95.51

96.63

98.88

100

TLR

96.07

97.75

98.88

100

100

TLD

97.19

98.32

98.88

100

100

CMCs after Rank-Level Fusion of all three Fingernail Plates for TLA feature-set using Logistic Regression and Mixed Group Rank, where all weights have been computed using PSO

FIGURE 9.19 CMCs after Rank-Level Fusion of all three Fingernail Plates for TLA feature-set using Logistic Regression and Mixed Group Rank, where all weights have been computed using PSO.

Figures 9.19-9.21 show' that the three fingernail plates can be combined to build a reliable identification system. Systems where the fusion is performed using Logistic Regression give appreciable performance. However, the systems w'hich employ Mixed Group Rank outdo the former.

Table 9.6 portrays the improvement in Rank-1 TPIR values obtained in this set of experiments when the weights are optimised using PSO.

CMCs after Rank-Level Fusion of all three Fingernail Plates for TLR feature-set using Logistic Regression and Mixed Group Rank, where all weights have been computed using PSO

FIGURE 9.20 CMCs after Rank-Level Fusion of all three Fingernail Plates for TLR feature-set using Logistic Regression and Mixed Group Rank, where all weights have been computed using PSO.

CMCs after Rank-Level Fusion of all three Fingernail Plates for TLD feature-set using Logistic Regression and Mixed Group Rank, where all «'eights have been computed using PSO

FIGURE 9.21 CMCs after Rank-Level Fusion of all three Fingernail Plates for TLD feature-set using Logistic Regression and Mixed Group Rank, where all «'eights have been computed using PSO.

9.6.2.2.3 Experiment C

To check the efficacy of the proposed system further, the next experiment has implemented rank-level fusion of the three nail plates using another fusion rule: the Inverse Rank Position method. Figure 9.22 shows the results obtained after carrying out this experiment. It is seen that fusion of the TLR feature-based scores through this method gives the highest Rank-1 identification accuracy (98.88%). Table 9.7 illustrates the Rank-1 TPIRs for the three models.

TABLE 9.6

Rank-1 Identification Rates (in %) of the Multimodal Systems where Nail Plates of All Three Fingers are fused using Two Different Linear Weighted Fusion Methodsa

Model

Fusion Method

Logistic Regression

Mixed Group Rank

Weights computed Empirically

TLA

94.38

94.38

TLR

96.07

96.07

TLD

97.19

97.19

Weights determined using PSO

TLA

96.07

96.63

TLR

97.75

98.32

TLD

98.32

98.88

a Number in bold and italics signify the best performance obtained.

9.6.2.2.4 Experiment D

Appreciable identification accuracy obtained from experiments A-C performed under the current section motivated authors to explore the fingernail plate multimodal identification system further. With the same intent, multimodal systems have been designed where all three fingernail plates have been fused using two different nonlinear weighted fusion rules.

CMCs after Rank-Level Fusion of all three Fingernail Plates for TLA. TLR, and TLR using Inverse Rank method

FIGURE 9.22 CMCs after Rank-Level Fusion of all three Fingernail Plates for TLA. TLR, and TLR using Inverse Rank method.

TABLE 9.7

Rank-1 Identification Rates (in %) of the Multimodal Systems where Nail Plates of All Three Fingers are fused using the Inverse Rank Position methoda

TLA

TLR

TLD

97.75

98.88

98.32

a Number in bold and italics signify the best performance obtained.

The experiment carried out under this section builds a multimodal system where the index, middle, and ring fingernail plates have been fused using Weighted Exponential and Hyperbolic Tangent: two different nonlinear, weighted fusion rules. Figure 9.23 gives the comparative depiction of the performance of the index, middle, and ring fingernail plates when their corresponding TLA based scores are fused at the rank- level using Weighted Exponential and Hyperbolic Tangent, while the same finding for TLR feature-set is shown in Figure 9.24, and that for TLD feature-set is given in Figure 9.25. For this experiment, all weights have been determined via empirical computation.

Figures 9.23-9.25 demonstrate that while the identification performance might be considered to be satisfactory, the results obtained in this experiment, especially the Rank-1 TPIRs are lower than that obtained through the Inverse Rank Position method, or those obtained through the Linear Weighted methods even when respective weights are computed empirically. Also, the Rank-1 TPIRs obtained using TLD model is less than that obtained using both TLR and TLA when Hyperbolic Tangent is used as the fusion rule. This is highly unlikely as TLD is much deeper than TLA, and this may have been caused because of possible inappropriate weight attribution to the three nail plates considered. Table 9.8 enlists the Rank-1 TPIRs obtained under this experiment.

CMCs after Rank-Level Fusion of all three Fingernail Plates for TLA feature-set using Weighted Exponential and Hyperbolic Tangent, where all weights have been computed empirically

FIGURE 9.23 CMCs after Rank-Level Fusion of all three Fingernail Plates for TLA feature-set using Weighted Exponential and Hyperbolic Tangent, where all weights have been computed empirically.

CMCs after Rank-Level Fusion of all three Fingernail Plates for TLR feature-set using Weighted Exponential and Hyperbolic Tangent, where all weights have been computed empirically

FIGURE 9.24 CMCs after Rank-Level Fusion of all three Fingernail Plates for TLR feature-set using Weighted Exponential and Hyperbolic Tangent, where all weights have been computed empirically.

CMCs after Rank-Level Fusion of all three Fingernail Plates for TLD feature-set using Weighted Exponential and Hyperbolic Tangent, where all weights have been computed empirically

FIGURE 9.25 CMCs after Rank-Level Fusion of all three Fingernail Plates for TLD feature-set using Weighted Exponential and Hyperbolic Tangent, where all weights have been computed empirically.

9.6.2.2.5 Experiment E

With an aim to achieve better identification accuracy, and also to check the sort of unlikely results obtained under the previous set of experiments, the same have been repeated after computing weights using PSO. Also, it is important to note that results obtained from Experiment В confirms that optimisation of weights improves identification accuracy considerably. Thus under this experiment, the three nail

TABLE 9.8

Rank-1 Identification Rates (in %) of the Multimodal Systems where Nail Plates of All Three Fingers are fused using Two Different Nonlinear Weighted Fusion methods (Weights computed Empirically)

Model

Fusion Method

Weighted Exponential

Hyperbolic Tangent

TLA

93.26

97.19

TLR

96.07

98.32

TLD

96.07

94.38

plates have been fused using Weighted Exponential and Hyperbolic Tangent rules, where weights have been calculated using PSO.

Figure 9.26 gives the comparative depiction of the performance of fused nail plates using the TLA feature-set when the fusion is performed using Weighted Exponential and Hyperbolic Tangent. The same findings for TLR and TLD feature-sets are given in Figures 9.27 and 9.28, respectively.

Analysing the CMCs in Figures 9.26-9.28 reveals that in this experiment, the highest Rank-1 TPIR of 99.44% has been provided by both TLR and TLD, when fusion is performed through the Hyperbolic Tangent rule. Comparing these results with those obtained in the previous experiment shows that the performance accuracy has improved significantly when the weight attribution has been performed using PSO.

CMCs after Rank-Level Fusion of all three Fingernail Plates for TLA feature-set using Weighted Exponential and Hyperbolic Tangent, where all weights have been computed using PSO

FIGURE 9.26 CMCs after Rank-Level Fusion of all three Fingernail Plates for TLA feature-set using Weighted Exponential and Hyperbolic Tangent, where all weights have been computed using PSO.

CMCs after Rank-Level Fusion of all three Fingernail Plates for TLR feature-set using Weighted Exponential and Hyperbolic Tangent, where all «'eights have been computed using PSO

FIGURE 9.27 CMCs after Rank-Level Fusion of all three Fingernail Plates for TLR feature-set using Weighted Exponential and Hyperbolic Tangent, where all «'eights have been computed using PSO.

CMCs after Rank-Level Fusion of all three Fingernail Plates for TLD feature-set using Weighted Exponential and Hyperbolic Tangent, where all weights have been computed using PSO

FIGURE 9.28 CMCs after Rank-Level Fusion of all three Fingernail Plates for TLD feature-set using Weighted Exponential and Hyperbolic Tangent, where all weights have been computed using PSO.

The Rank-1 TPIRs of the different systems in this experiment have been given in Table 9.9 to alleviate their comparative representation.

To make a comparison of the methodologies adopted and results obtained in the current work with that of some of the previously reported works which investigated the fingernail, a tabular depiction has been made in Table 9.10.

TABLE 9.9

Rank-1 Identification Rates (in %) of the Multimodal Systems where Nail Plates of Three Fingers are fused using Nonlinear Weighted Fusion Methods (Weights computed using PSO)-1

Model

Fusion Method

Weighted Exponential

Hyperbolic Tangent

TLA

94.94

98.32

TLR

97.75

99.44

TLD

98.88

99.44

a Numbers in bold and italics signify the best performance obtained.

TABLE 9.10

Comparison of Proposed Work with Significant Existing Works on Fingernail1

Ref.

No.

Part of Finger Explored

Feature

Extraction

Technique

Database

Results of Unimodal System

Results of Fusion

[15]

Index

Fingernail

Bed

None

Not reported

Binary

representation of the relative positions of capillary loops is obtained, which is unique to every individual.

Not explored

[14]

Nail-

Ridges

None

Not reported

Bands of colour obtained. Each represents a single ridge or valley of nail surface.

Not explored

[16]

Nail

Surface of Index, Middle, and Ring Fingers

Hand-crafted Approach: Haar Wavelet

Database of 5 images/180 users = 900 images per modality

Highest accuracy reported: Verification: GAR = 50% (at FAR = 0.01%) Identification not investigated.

Fusion of three nail surfaces done. Verification results as high as GAR = 72% (at FAR = 0.01%) for Product Rule.

Identification not investigated.

(Continued)

TABLE 9.10 (Continued)

Comparison of Proposed Work with Significant Existing Works on Fingernail-1

Ref.

No.

Part of Finger Explored

Feature

Extraction

Technique

Database

Results of Unimodal System

Results of Fusion

[17]

Nail Plate of Index, Middle, and Ring Fingers

Hand-crafted Approaches: Haar Wavelet and ICA

Database of 5 images/180 users = 900 images per modality

Highest accuracy reported: Verification: GAR = 55% (at FAR = 0.01%) Identification: 81% Rank-1 TP1R

Fusion of three nail plates done.

Verification and Identification accuracy as high as GAR = 85% (at FAR = 0.01%) and 96.5% Rank-1 TPIR respectively reported

[8]

Nail Plate of Index, Middle, and Ring Fingers

Hand-crafted Approaches: Haar Wavelet and ICA

Database of 5 images/180 users = 900 images per modality

Highest accuracy reported: Verification: GAR = 60%

(at FAR = 0.01%) Identification: 89% Rank-1 TPIR

Fusion of three nail plates done.

Verification and Identification accuracy as high as GAR = 80% (at FAR = 0.01%) and 96.5% Rank-1 TPIR respectively reported

[5]

Knuckle and Nail Plate of Index, Middle and Ring Fingers

Deep

Learning

Approach:

AlexNet

Database of 5

images/178 users =

890 images per

modality

Highest accuracy reported from Nail Plates: Verification: GAR = 56.74% (at FAR = 0.01%) Identification: 90.45% Rank-1 TPIR

Fusion of a) three nail plates, b) three knuckles, c) each nail plate with corresponding knuckle done.

For fusion of nail plate and knuckle. Verification and Identification accuracy as high as GAR = 96.63% (at FAR = 0.01%) and 98.31% Rank-1 TPIR, respectively, reported

For fusion of three nail plates, Verification and Identification accuracy as high as GAR = 87.64% (at FAR = 0.01%) and 98.31% Rank-1 TPIR, respectively, reported

(Continued)

TABLE 9.10 (Continued)

Comparison of Proposed Work with Significant Existing Works on Fingernail1

Ref.

No.

Part of Finger Explored

Feature

Extraction

Technique

Database

Results of Unimodal System

Results of Fusion

This

Work

Nail Plate of Index, Middle, and Ring Fingers

Deep Learning Approaches: AlexNet, ResNet and DenseNet

Same

database as in Ref. [5]

Highest accuracy Verification: GAR = 74.16% (at FAR = 0.01%) Identification: 93.82% Rank-1 TP1R

Verification and Identification accuracy as high as GAR = 94.94% (at FAR = 0.01%) and 99.44% Rank-1 TPIR respectively

■ Numbers in bold and italics signify the best performance obtained in the current work.

 
Source
< Prev   CONTENTS   Source   Next >