PERFORMANCE EVALUATION

Databases and Tools Used

The proposed model is evaluated on two publicly available latent fingerprint databases:

  • 1. IIITD-MOLF Database [26]: Indraprastha Institute of Information Technology Delhi Multi-sensor Optical and Latent Fingerprint (IIITD- MOLF) is the biggest latent fingerprint database which is available in the public domain. It has latent fingerprints and live fingerprints acquired through different optical sensors. These fingerprints are collected from 100 subjects. This database has 4,400 latent fingerprints and 4,000 live fingerprints corresponding to each sensor.
  • 2. IIITD-MSLF Database [1]: Indraprastha Institute of Information Technology Delhi Multi-surface Latent Fingerprint IIITD-MSLF database has latent fingerprints extracted from eight different surfaces like transparent glass, compact disc, ceramic mug, hardbound cover, etc. It has 551 latent fingerprints of 51 subjects.

The standard latent fingerprint database provided by NIST, NIST-SD27 has now been removed from the public domain due to which we cannot evaluate the proposed model on NIST-SD27 database. The proposed model is designed for the standard sized 500 dpi fingerprint image whose spatial dimensions are 512x512 pixels. The latent fingerprints are pre-processed and zero-padded to have a fixed size of 512x512. Table 3.3 provides the list of publicly available tools used in this work.

Evaluation Criteria

Every fingerprint enhancement algorithm is designed to increase the clarity of ridges and valleys while preserving the ridge details to improve minutiae extraction and thereby improving fingerprint matching performance. We evaluate the proposed enhancement algorithm using the metrics given below:

1. Fingerprint Quality Analysis: Quality of a fingerprint image is determined as the ability of a fingerprint matcher to correctly match the image. Poor quality fingerprints often result in poor matching performance. We evaluate the fingerprint quality of latent fingerprints before and after enhancement using NIST Finger Image Quality (NFIQ) module of NBIS. NFIQ calculates quality of a fingerprint image using features such as: clarity of ridges and valleys, number of minutiae, size of the fingerprint image, etc. NFIQ scores a fingerprint image between 1 and 5 where 1 signifies the best fingerprint image quality and 5 means the worst quality. We compare the histogram of quality scores obtained by NFIQ before and after enhancement. Another publicly available tool to evaluate the quality of fingerprint images is NFIQ2 [25], which returns a score between 1 and 100. NFIQ2 is a more robust fingerprint quality assessment metric than NFIQ. However, NFIQ2 fails to process raw latent fingerprint images of

TABLE 3.3

Table Summarising the Publicly Available Tools Used

Tool

Purpose

Usage

MINDTCT module of NBIS

Minutiae extraction

During testing, to extract minutiae from enhanced image and gallery images

NFIQ module of NBIS

Evaluates fingerprint image quality

During testing, to evaluate quality of enhanced fingerprints

BOZORTH module of NBIS

To match fingerprints

During testing, to perform fingerprint matching on minutiae extracted by MINDTCT

MCC fingerprint matcher

To match fingerprints

During testing, to perform fingerprint matching on minutiae extracted by MINDTCT

NFIQ2

Evaluates fingerprint image quality

To evaluate quality of NIST SD4 images and keep good quality images for training the model

Binarisation module of NBIS

Binarise the fingerprint image

To generate the ground-truth binarisation of training images

IIITD-MOLF database. As a result, we only compare fingerprint quality score obtained using NFIQ.

2. Ridge Structure Preservation: The most crucial factor for any fingerprint enhancement is that it should retain the ridge structure while improving clarity of ridges and valleys. To showcase ridge structure preservation (including minutiae) by the proposed model, we synthetically generate some test cases by adding noises and backgrounds on good quality fingerprints. We showcase the similarity between ground-truth binarisation and the enhanced fingerprint image generated by the proposed algorithm using the following two measures: i. We calculate Structural Similarity Index Metric (SSIM) [27] between the ground-truth binarised image and the enhanced fingerprint. SSIM is a metric which computes similarity between image a and image b based on the contrast, luminance, and structure:

where /ua, /jb are the mean, aa, ah are the standard deviation, and oab is the covariance between image a and image b. ii. We also calculate match score (using Bozorth) between ground-truth binarised image and the enhanced image generated by the proposed model. High match scores demonstrate that the proposed algorithm preserves minutiae while enhancing the input latent fingerprint image.

3. Matching Performance: The ultimate success of a fingerprint enhancement algorithm is when it is able to improve the fingerprint matching performance. We extract minutiae using the MINDTCT module of NBIS and use Bozorth and Minutia Cylinder Code (MCC) [28-30] fingerprint matchers to evaluate fingerprint matching performance. We compare matching performance before and after enhancement using Rank-50 accuracy. Rank-k accuracy is defined as:

Rank-k accuracy = no. of probe fingerprint for which the matching fingerprint in gallery achieved top-к scores x 100/total no. of probe fingerprints

We also plot cumulative matching curve (CMC), which is a Rank-k accuracy plot over varying values of k. CMC is a standard summarisation technique to quantify the matching performance of a closed-set identification system. We compare the CMC before and after enhancement in Figure 3.9.

 
Source
< Prev   CONTENTS   Source   Next >