EXPERIMENTS, RESULTS, AND ANALYSES

The experimentation in this work requires the nail plate ROI extraction from the index, middle, and ring fingers of all the hand images in the database of 890 hand dorsal images. Thus, 2,670 (178 volunteers x 5 images each x 3 fingers) ROIs have been extracted in total.

The TLA, TLR, and TLD feature-sets have been extracted from the all the three fingernail plate ROI databases. It is to be noted that for any kind of processing, the images are required to be sized as per the pre-defined image input size of the respective deep learning models. In accordance with the same, the images have been resized to 227 x 227 x 3 for the TLA network and to 224 x 224 x 3 for both the TLR and the TLD networks.

Choosing Euclidean distance as the similarity measure, and the training to test ratio as 4:1, the genuine and imposter scores have been computed from the features, which are then used to check the performance of the traits in verification systems. For analysing the performance of fingernail plate in identification systems, the corresponding ranking lists for all the three nail plates are obtained by sorting the scores.

Performance of Fingernail Plates in Verification Systems

Performance of Fingernail Plates in Unimodal Verification Systems

In this section, performances of the index, middle, and ring fingernail plates in unimodal verification systems have been comparatively analysed for each of the three deep learning models considered. Figure 9.7 compares the verification performance in the form of Receiver Operating Characteristic (ROC) curves of the three fingernail plates using TLA as the feature extraction technique. Similar comparative depiction of the nail plates for the TLR and TLD feature-sets are given in Figures 9.8 and 9.9, respectively.

ROCs comparing the performance of Index, Middle, and Ring Fingernail

FIGURE 9.7 ROCs comparing the performance of Index, Middle, and Ring Fingernail

Plates in Unimodal Verification Systems using TLA as the feature extraction technique.

Figure 9.7 demonstrates that verification systems built from either of the three nail plates shall give similar and appreciable results at all levels of security, when the TLA feature-sets are being used.

Figure 9.8 portrays that the index and middle fingernail plates outperform the ring fingernail plate in terms of verification performance when TLR feature-sets are being used.

Figure 9.9 shows that the verification systems comprising either the index or the ring fingernail plate shall give very good results when the TLD feature-sets are being used. It is evident from the figure that the middle fingernail plate also gives appreciable results, but lags behind the other two nail plates.

ROCs comparing the performance of Index, Middle, and Ring Fingernail Plates in Unimodal Verification Systems using TLR as the feature extraction technique

FIGURE 9.8 ROCs comparing the performance of Index, Middle, and Ring Fingernail Plates in Unimodal Verification Systems using TLR as the feature extraction technique.

ROCs comparing the performance of Index. Middle, and Ring Fingernail

FIGURE 9.9 ROCs comparing the performance of Index. Middle, and Ring Fingernail

Plates in Unimodal Verification Systems using TLD as the feature extraction technique.

TABLE 9.2

GAR (in %) at FAR = 0.012% for Various Unimodal Verification Biometric Systems built using Index, Middle, and Ring Fingernail Plates*

Model

Trait

Index

Middle

Ring

TLA

56.74

53.37

56.18

TLR

74.16

74.16

62.36

TLD

68.53

60.11

67.41

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

The values of Genuine Acceptance Rate (GAR) at False Acceptance Rate (FAR) = 0.012%, obtained from all the three nail plates from all the three deep learning models are tabulated in Table 9.2.

Analyses of the particular results tabulated in Table 9.2 show that amongst all the three nail plates considered, the best verification performance is provided by the index fingernail plate, for each of the three feature-sets. Table 9.2 also shows that TLR gives the best performance followed by TLD and TLA in the mentioned sequence at FAR = 0.012%.

Performance of Fingernail Plates in Multimodal Verification Systems

In this section, fusion of the scores obtained from all the three fingernail plates has been carried out using the four fusion rules - sum, product, min, and max - as detailed in the Section 9.5.1.

At the very outset, the fusion of TLA scores obtained from all three nail plates has been implemented. Figure 9.10 depicts the verification performance of the same

ROCs after Score-Level Fusion of scores from Index, Middle, and Ring Fingernail Plates using TLA feature-sets and four different score-level fusion rules

FIGURE 9.10 ROCs after Score-Level Fusion of scores from Index, Middle, and Ring Fingernail Plates using TLA feature-sets and four different score-level fusion rules.

ROCs after Score-Level Fusion of scores from Index, Middle, and Ring Fingernail Plates using TLR feature-sets and four different score-level fusion rules

FIGURE 9.11 ROCs after Score-Level Fusion of scores from Index, Middle, and Ring Fingernail Plates using TLR feature-sets and four different score-level fusion rules.

ROCs after Score-Level Fusion of scores from Index, Middle, and Ring Fingernail Plates using TLD feature-sets and four different score-level fusion rules

FIGURE 9.12 ROCs after Score-Level Fusion of scores from Index, Middle, and Ring Fingernail Plates using TLD feature-sets and four different score-level fusion rules.

in the form of ROCs. Similar performances after score-level fusion of TLR and TLD based scores are given in Figures 9.11 and 9.12, respectively.

Analysing the results in Figures 9.10-9.12 and comparing them with that of unimodal verification systems reported in Section 9.6.1.1 shows that the multimodal systems outperform their unimodal counterparts with all each of the four rules used. These results also show that for the chosen traits, the ‘product rule’ fusion method performs better than the other three score-level fusion methods for almost all operating points.

TABLE 9.3

Comparison of GAR (in %) at FAR = 0.012% of Various Multimodal Verification Biometric Systems with that of Unimodal Verification Systems3

Model

Rule^ait

Sum

Product

Min

Max

Index

Middle

Ring

TLA

84.27

87.64

65.17

64.04

56.74

53.37

56.18

TLR

93.82

94.94

82.02

84.83

74.16

74.16

62.36

TLD

04.38

04.38

80.34

88.20

68.53

60.11

67.41

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

Table 9.3 makes a comparison of the values of GAR obtained at FAR = 0.012% after score-level fusion with that of the unimodal fingernail plate systems.

The results tabulated in Table 9.3 show that the best verification performance is obtained from the fusion of TLR features-based scores, followed by TLD and TLA in the mentioned sequence. This is the same sequence of performance that is observed in the unimodal verification systems. Table 9.3 also shows that the highest GAR at FAR = 0.012% is given by the product rule. However, the sum rule also performs well and even equals the performance given by the product rule in the case where the TLD match scores are fused.

Thus, it may be said that appreciable verification performance can be achieved by the combination of index, middle, and ring fingernail plates.

However, there are situations where score-level fusion is not feasible or not practiceable. Moreover, certain circumstances like criminal investigation demand for the authentication to be carried out in identification mode. Keeping the same in mind, the next section analyses the performance of fingernail plates under rank-level fusion.

 
Source
< Prev   CONTENTS   Source   Next >