Fingerprint retrieval by spatial modelling and distorted sample generation

Abstract

In this study, the authors extend and refine the process of fingerprint retrieval, with the goal of boosting recognition rates for the first rank candidate and low penetration rates. On top of a baseline retrieval system which extracts Gabor features in multiple directions from fingerprint images, the authors propose spatial modelling techniques to generate artificial samples for training the system. Translational modelling, rotational modelling and distorted sample generation techniques are used to augment the original training set in order to boost the accuracy of fingerprint retrieval. The effectiveness of the models is evaluated using the well‐known National Institute of Standards and Technology database 4. Experimental results, with reference to some leading fingerprint retrieval rates reported in the literature, confirm that the authors’ proposed system is promising in recognition performance.

Improvement of Fingerprint Retrieval by a Statistical Classifier

“…The proposed system is built on top of a baseline system. The baseline system is implemented according to the baseline descriptions in our recent work focusing on artificial sample generation [12], unless otherwise specified in Section III-A-III-C below. The baseline system is built up from a number of existing algorithms and techniques.…”

“…This is due to the fact that many fingerprint images in the FVC databases have their registration point located close to the lower border of the frame. We do not use any weighted feature as described in our previous work [12] to attenuate features situating far away from the registration point [21], mainly because the FVC databases do not provide enough fingers for accurate estimation of the weights. Variable transformation that takes the power of each feature value is also not used [12].…”

“…We do not use any weighted feature as described in our previous work [12] to attenuate features situating far away from the registration point [21], mainly because the FVC databases do not provide enough fingers for accurate estimation of the weights. Variable transformation that takes the power of each feature value is also not used [12]. A graphical illustration for defining the blocks of interest for feature extraction is shown in Fig.…”

“…Regarding the second issue, it is realized that the original feature dimension is very high, ranging from a few hundred to a thousand [12], [16], [17], [19]- [22], [24]. We implement a variant of the Fisher's linear discriminant for dimension reduction and a variant of the quadratic discriminant function (Bayes decision function assuming Gaussian statistics) for lowering estimation errors.…”

“…We have attempted to tackle the problem of insufficient training samples by generating additional samples, emulating the genuine samples that would have been captured by the fingerprint sensing device during fingerprint enrollment [12]. In our previous work, we have tried to produce more training samples for the NIST-4 database by three kinds of spatial modeling, namely, translational modeling, rotational modeling, and distorted sample generation.…”