Abstract
M.Ing. (Electrical and Electronic Engineering Technology)
Automated Fingerprint Recognition Systems (AFRSs) have not been very effective so far
in matching poor quality fingerprints because the challenges involved in low quality fingerprint matching are quite different from high quality fingerprint matching. The difficulty is
due to three main reasons: (i) poor quality of fingerprints in terms of the clarity of ridge
information due to harsh working conditions, diseases and aging, (ii) failure to acquire
adequate minutiae points after segmentation and (iii) large non-linear distortion due to
pressure variations which causes ridges to deform. Thus, low quality fingerprint recognition
is a difficult problem which still needs more attention. This is because the accuracy of a
fingerprint matching module heavily depends on the quality of the fingerprint probe image.
Poor quality fingerprints lead to maximization of False Acceptance Rate (FAR) instead
of True Acceptance Rate (TAR). As a result, researchers have suggested that extended
features must be incorporated to improve accuracy. These features have been successfully
used by Latent Print Experts (LPEs) for crime investigation purposes to increase matching
accuracy for fingerprints collected from crime scenes with those stored in the national or
international databases. There are three categories of fingerprint features: (i) level 1 (e.g.
delta), (ii) level 2 (e.g. minutiae) and (iii) level 3 or extended features (e.g. pores). In this
work, improvements have been made through fusion of minutiae and extended feature scores together with the fingerprint image quality. However, fusion algorithms designed so far are not adaptive, i.e. they assume that the effect of the quality of the image on the matching score is the same for different matchers based on different features. To test this assumption, this work adopted an algorithm from the literature that first assigns quality score to different regions of a fingerprint. Quality scores assigned to each region of the segmented
fingerprint was mapped to extracted minutiae and extended features (pores). The overall
quality rating of each of these were calculated as the sum of all quality scores assigned to
regions. This procedure helped the designed fusion algorithm to assign more weight on
highly reliable features and less weight on unreliable features. Two experiments conducted
for rating minutiae and pore features that are based on this procedure, showed that quality
scores for features under study do not stay constant. An adaptive weighted sum fusion
algorithm was designed, implemented, tested and compared to non-adaptive algorithms,
namely, simple sum and weighted sum fusion. The proposed adaptive weighted sum differs
from traditional weighted sum fusion algorithm in that it uses weights assigned to each
feature based on the quality map of each region of the fingerprint as opposed to the whole
image. The performance of the system was tested using PlyU High Resolution Fingerprint
(HRF) Database. Two performance measures were used to rate the proposed algorithm in
comparison with simple sum and traditional weighted sum, namely, Area Under the Curve
(AUC) and Equal Error Rate (EER). Both these performance measures showed that the
algorithm proposed in this work outperforms both simple sum and traditional weighted sum
fusion approaches. The proposed algorithm yields an improvement of 8% and 13.33% in
EER and AUC, respectively for weighted sum fusion and 2% and 4.8% in EER and AUC,
respectively for simple sum fusion.