RESULTS AND ANALYSIS

1. Fingerprint Quality Analysis: Figure 3.6 represents the histogram of NFIQ scores before and after enhancement. The average NFIQ score has improved from 4.96 to 1.91 after enhancement (smaller score means better quality) on IIITD-MOLF and 4.48 to 2.64 on IIIT-D MSLF database (see Table 3.4) which validates the improved clarity of ridges and valleys (thereby improving the quality score) in the enhanced fingerprints generated by the proposed model.

FIGURE 3.6 Evaluation of quality of fingerprint images using the NFIQ module of NBIS [2] for latent fingerprint images from (a) IIITD-MOLF and (b) II1TD-MSLF database.

TABLE 3.4

Average NFIQ Scores Before and After Enhancement by the Proposed Model on IIITD-MOLF and IIITD-MSLF Databases

• 2. Ridge Structure Preservation: In Figure 3.7, we present some sample test cases with their ground-truth binarisation and the output of the proposed algorithm. High match score and high SSIM value between the ground- truth binarised image and the output of the proposed method illustrate that the proposed algorithm preserves the ridge information of input latent fingerprint images including fingerprint class, orientation of ridges and minutiae, while enhancing them.
• 3. Matching Latent Fingerprints to Multi-Sensor Fingerprints: In

Figure 3.8, we show CMCs for matching performance achieved by Bozorth and MCC matcher on the enhanced image generated by the proposed model across two different galleries. We also compare the Rank-50 accuracy of the proposed model with the recently proposed latent fingerprint algorithm [17] (see Table 3.5). The magnitude of improvement obtained over raw images using the proposed algorithm is much more than the previous work (Rank-50 accuracy of 34.43% on DB1 and 30.50% on DB2 gallery using the proposed algorithm compared to 22.36% on DB1 and 19.50% on DB2 gallery by the previous work [17]). This demonstrates that the proposed algorithm performs better than [17] in improving ridge- valley contrast, removing background noise while preserving ridge details

FIGURE 3.7 Left side shows the enhanced fingerprint generated from the synthetic latent fingerprint, corresponding to the ground-truth binarised image (shown in the middle). Right side shows the SSIM value and the matching score (obtained using Bozorth) for each enhanced image corresponding to the ground-truth image.

FIGURE 3.8 CMC for proposed algorithm’s matching performance for the IIITD-MOLF DB1 gallery and DB2 gallery at л = 5. using (a) Bozorth (b) MCC.

due to which improved feature extraction and thereby, improved matching performance is obtained.

4. Matching Multi-Surface Latent Fingerprints to Gallery of Live-Scan Fingerprints: The Rank-50 accuracies before and after enhancement on IIITD-MSLF database using Bozorth are 11.43% and 12.80%, respectively. The CMC is shown in Figure 3.9. The accuracy obtained on IIITD-MSLF database is lesser compared to the accuracy achieved on IIITD-MOLF.

TABLE 3.5

Rank-50 Obtained on IIITD-MSLF Database Before and After Enhancement by the Proposed Model

FIGURE 3.9 CMC representing matching performance on IIITD-MSLF database, before and after enhancement by the proposed algorithm using Bozorth.

This is due to the complex background present in IIITD-MSLF database images. We observe that the intensity values of foreground and the background fingerprint regions have similar distribution in many images. This leads to spurious pattern generation by the proposed algorithm, which adversely affects the matching performance.

• 5. Significance of Latent Fingerprint Reconstruction Loss: To demonstrate the significance of the reconstruction loss in the objective function, we train the proposed model with only adversarial loss (Л = 0). We observe that the model becomes unstable and doesn’t converge. In Figure 3.10, we show the sample results obtained with only reconstruction loss. Therefore, we conclude that the reconstruction loss is essential to stabilise the proposed model.
• 6. Role of Hyper-Parameters: During the various experiments conducted in this chapter, we find that the MCC is a better fingerprint matcher for latent fingerprints (see Tables 3.6-3.9). We conclude our observations based on the results obtained using the MCC matcher. By default, the hyper-parameters used are X = 10, batch size = 2, and number of epochs = 200.

FIGURE 3.10 Sample-enhanced images obtained by the model when trained without latent fingerprint reconstruction loss.

FIGURE 3.11 Failure cases of NFIQ (lower score means better quality).

i. Weight Hyper-parameter (A): We observe that the Rank-50 achieved by the proposed model increases as the weight of enhanced fingerprint reconstruction loss is increased from 1 to 5 (as depicted in Table 3.9, Figures 3.12a and b and 3.13a and b). However, on increasing the weight further, the performance starts degrading. The best Rank-50 accuracies of 34.43% across DB1 and 30.50% across DB2 gallery are achieved for A = 5. The quality, on the other hand, improves, while A is increased from 1 to 10. On increasing A further, the quality starts degrading (see Table 3.10 and Figure 3.14a). This suggests that the model is sensitive to the choice of weight parameters, and a careful combination of adversarial loss and reconstruction loss is required to efficiently train the proposed model.

TABLE 3.6

Rank-50 Accuracy Obtained over Different Epochs on IIITD-MOLF Latent Fingerprints

TABLE 3.7

Rank-50 Accuracy Obtained on IIITD-MOLF Latent Fingerprints with and without Adding NIST-SD4 Images in Training Data

TABLE 3.8

Rank-50 Accuracy Obtained IIITD-MOLF Latent Fingerprints for Different Batch Size

ii. Number of Epochs: As shown in Table 3.6 and Figures 3.12c and d and 3.13c and d, the Rank-50 accuracy initially improves with the number of epochs till 60 epochs, after which it fluctuates and is approximately the same till 200 epochs. The performance of the model starts degrading after 200 epochs due to over-fitting. NFIQ score, on the other hand, is improved initially till 150 epochs and then it fluctuates and no clear trend is found (see Table 3.11).

iii. Batch Size: We compare the Rank-50 accuracy achieved by the model at batch size = 2, 4, and 8 (see Table 3.8, Figures 3.12e and f and

TABLE 3.9

Rank-50 Accuracy Obtained on IIITD-MOLF Latent Fingerprints over Different Values of X

3.13e and f). As the batch size is increased, the number of parameter updates per epoch reduces which leads to faster training of the model. Batch size = 2 turns out to be the best value of hyper-parameter batch size. The best performance at batch size = 2 is attributed to more parameter updates and thus better training.

Better quality score is obtained for batch size = 8 than batch size = 2 and batch size = 4, as can be seen in Table 3.12 and Figure 3.14c. This is a counter-intuitive result as the enhanced fingerprint generated at batch size = 8 has large missing regions compared to the fingerprints generated at batch size = 2. Many of the images generated at batch size = 2 have poor ridge smoothness compared to the image generated at batch size = 8 which have very small reconstructed fingerprint area but better smoothness (see Figure 3.11). Due to this, NFIQ gives a better score to images generated at batch size = 8 than the images generated at batch size = 2. Thus, we conclude that the anomaly in the quality score is due to the limitation of NFIQ. In reality, the quality of fingerprints generated at batch size = 2 and batch size = 4 is better than the quality of fingerprints generated at batch size=8 which is evident through the much higher Rank-50 accuracy achieved for batch size = 2 and batch size = 4 than batch size = 8.

7. Effect of Training the Model with Real NIST SD4 Images: The Rank-50 accuracy achieved by the images generated by the proposed model trained on both synthetic latent fingerprints and real fingerprints from NIST SD4 is better compared to the model trained on only synthetic latent fingerprints, as shown in Figures 3.15, 3.12g and h, and 3.13g and h). Similar trend is seen in the NFIQ quality scores of the enhanced fingerprints obtained using the proposed model (see Table 3.13 and Figure 3.14d). The real fingerprints have practical cases of non-linear distortion and non-uniform ridge width which are also observed in latent fingerprints. Thus, the real inked fingerprints like those of NIST SD4 database help the model to learn to be invariant to such distortions.

FIGURE 3.12 CMC for matching the proposed algorithm using Bozorth on the IIITD- MOLF DB1 and DB2 galleries across different training settings.

FIGURE 3.13 CMC for matching the proposed algorithm using MCC on the IIITD-MOLF DB1 and DB2 galleries across different training settings.

TABLE 3.10

Average NFIQ Scores of the Enhanced Fingerprints Obtained for IIITD-MOLF Database Using the Proposed Algorithm over Different Values of A

FIGURE 3.14 NFIQ score distribution of the enhanced images produced by the proposed algorithm across different training settings.