CHALLENGES OBSERVED

While conducting different experiments, it has been found that the proposed algorithm improves matching performance (see Figure 3.16). However, we observe some cases where the proposed algorithm does not generate good results. Analysis of these cases is given in the following points:

TABLE 3.11

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

Epoch

NFIQ Score

30

2.07

60

2.03

90

2.00

120

1.86

150

1.82

180

1.84

200

1.83

210

1.83

240

1.81

270

1.83

TABLE 3.12

Average NFIQ Scores of the Enhanced Fingerprints Obtained for IIITD- MOLF Database Using the Proposed Algorithm over with and without Adding SD4 Images in Training Data

Batch Size

NFIQ Score

2

1.83

4

1.83

8

1.18

Sample results obtained by the model when trained with and without NIST SD4 images in the training dataset

FIGURE 3.15 Sample results obtained by the model when trained with and without NIST SD4 images in the training dataset.

TABLE 3.13

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

Training Data

NFIQ Score

Without SD4

2.33

With SD4

1.83

Samples of successful enhancement of latent fingerprints by the proposed model

FIGURE 3.16 Samples of successful enhancement of latent fingerprints by the proposed model.

  • 1. We find that many of the input latent fingerprint images have low ridge information. However, even for such images, the proposed algorithm enhances those regions of the latent fingerprint image which have some ridge information (see the left-most column of Figure 3.17). We understand that it will be difficult for any enhancement algorithm to enhance such cases while preserving the minutiae details.
  • 2. While matching latent fingerprint images, ROI is manually marked by forensic experts and the enhancement is performed only on ROI. However, the proposed algorithm automatically segments the foreground and background and then enhances the foreground fingerprint. Due to this, it sometimes misinterprets the background as foreground (see the last three columns from right in Figure 3.17) when the intensity distributions of background and foreground fingerprint are similar.
  • 3. We found that the NFIQ is not a robust fingerprint quality assessment metric (see Figure 3.11). NFIQ2 is a more effective metric than NFIQ; however, it fails to process latent fingerprints. Thus, there is a need to introduce a more robust latent fingerprint quality assessment tool in the public domain to facilitate improved research in latent fingerprint matching.
  • 4. The proposed model is observed to be highly sensitive to the choice of hyperparameters and does not perform well if the training hyper-parameters are not carefully chosen.
  • 5. The loss function is carefully designed for enhancement of latent fingerprints. Any change in the loss function can lead to unstable training of the model (as observed while training the model without enhanced reconstruction loss, as shown in Figure 3.10).
Some challenging cases for the proposed model

FIGURE 3.17 Some challenging cases for the proposed model.

CONCLUSIONS

Motivated by the successful applications of GANs in various image processing applications, we formulate latent fingerprint enhancement like an image-to- image translation problem. The proposed model is trained using an enhancer and a discriminator network in an adversarial fashion. The model is trained using both synthetic and real fingerprints due to which it is robust to distortions observed in latent fingerprints. Moreover, the proposed model does not need a real latent fingerprint database to train the network. Two latent fingerprint databases available in the public domain are used for evaluating the proposed enhancement model. A detailed analysis of performance of model over hyper-parameters such as lambda, number of epochs, batch size is performed. We also gain insights on the role of real inked prints while training the model and the significance of reconstruction loss in the objective function.

We analyse the failure cases and some cases have been encountered when the ridge information is insufficient and the proposed algorithm generates spurious features. To address these limitations, the possibility of recoverability needs to be explored such that the algorithm can decide which portions of fingerprints can be reconstructed and which ones cannot. Training with a larger database with more variations in texture and background can help to achieve even better performance on IIITD-MSLF database. The proposed algorithm can also be utilised in challenging scenarios like latent to latent fingerprint matching.

ACKNOWLEDGEMENTS

The authors thank IIT Delhi HPC facility for computational resources. The authors express their gratitude to Mayank Vatsa and Richa Singh from HIT Delhi for their inputs. The authors also thank Himanshu Gandhi and Vijay Kumar for helpful discussions.

 
Source
< Prev   CONTENTS   Source   Next >