monash_165941.pdf (10.84 MB)
Download file

Partial fingerprint recognition through region-based approach

Download (10.84 MB)
thesis
posted on 02.03.2017, 02:13 authored by Zanganeh, Omid
Traditional person recognition methods rely on knowledge/possession-based approaches such as password and access cards. However, these kinds of approaches suffer from the threat that they might be shared or stolen. Thus, differentiating between authorized and unauthorized users is very difficult (if not impossible). Biometrics on the other hand offers a means to reliable person recognition which overcomes the previously mentioned issues of traditional methods. In modern days the rise of computer tools have made it achievable to make recognition processing fully automated. Fingerprints were one of the first biometric traits to be accepted and utilized by law enforcement agencies. Despite popular belief and years of research, automatic fingerprint recognition is not a foolproof system and is yet an open and challenging problem. One of the most prominent issues in automatic fingerprint recognition which prevent it from being foolproof is that impressions of different fingers may be more similar to each other than the impressions of the same finger (inter-class similarity). Likewise, different acquisition of the same finger may be dissimilar from each other compared to the impressions taken from different fingers (intra-class variation). Obviously, the above statement directly relates to how the similarity or dissimilarity of fingerprints is defined. One of the main factors leading to high intra-class variation is the low quality of the fingerprints. The performance of the fingerprint recognition system deeply depends on fingerprint image quality. The matching accuracy of automated fingerprint recognition system decreases significantly when the quality of the fingerprint is poor. For example, a fingerprint is considered of poor quality due to the existence of noise and scars on it which in these cases manual recognition achieves better accuracy than automated systems. Dealing with intra-class variation and inter-class similarity is even more challenging when it comes to partial fingerprint recognition. The issues of partial fingerprint recognition is outwardly similar to the issues of full fingerprint matching, however it keeps its own unique characteristics e.g. unavailability of all the features, covering small part of the finger, and being unclear and/or distorted. Although it is claimed by researchers that some regions of a fingerprint provide more distinguishing characteristics than others, uncertainty in what regions are available in a partial fingerprint increases the probability of misrecognising it. Also, different types of features can be extracted from a full fingerprint but not all of them are available in a partial fingerprint. Thus, correct recognition of partial fingerprint is dependent on proposing a method which is robust to the missing features and image quality as well as being able to extract the available distinguishing features, independently from partial fingerprint size and shape. In order to reasonably define the similarity in full and partial fingerprints, distinguishing the intra and inter cases, and to advance science in this field, the following contributions are presented: Firstly, a partial fingerprint alignment method is presented. Maltoni et al. and Pankanti et al. induced designers to look for additional distinguishing fingerprint features beyond minutiae (the widely-used fingerprint feature). The authors found that a grey-level fingerprint image contains richer, more discriminatory information than only the minutiae location. Considering the rich information provided by the grey-level fingerprint image, it can be used to align the fingerprints. Two fingerprints should be aligned properly in order to measure their similarity. The previously developed fingerprint alignment methods, including minutia and non-minutia feature based alignment methods, are not suitable for partial fingerprints. These methods are dependent on the fingerprint’s particular features such as reference points which might not be available in partial fingerprinting. Also,feature selection is a vital step in alignment (the same as for the matching process). Thus, the information/features used for alignment plays an important role in accurate alignment which is very important, specially in region-based matching (since corresponding regions will be compared). In the alignment stage, two methods are proposed which are suitable for partial and full fingerprint alignment. Here, the information obtained in the alignment step is used as the first level of fingerprint recognition. The first method is based on using the singular points and ridge structure of a fingerprint by cropping different regions with different sizes from query fingerprints and computing the similarity of these regions with registered fingerprints at different angles. The rotation angle that provides the highest correlation is considered as the rotation difference of query and registered fingerprint. However, singular points are not always available in partial fingerprints and additionally the correlation maximum score of comparing two regions might not always be the correct correlation. In some cases, it is possible that more than one peak (of approximately the same height) exists. That increases the probability of choosing the incorrect peak. Also a situation may happen whereby the correct peak is slightly lower than the false peak. To combat this in the second method, the consistency of corresponding regions located on the registered fingerprint is considered. By using the information provided from the alignment method, a Neural Network classifier is used to learn the behaviour of intra and inter comparisons in alignment. This classifier is used as the first-level of matching to reject/accept cases that do not need any further processing to be recognised. Secondly, a new similarity measurement technique that is suitable for partial and full fingerprints is presented. There is no agreed definition for the term similarity between researchers in terms of defining the similarity between fingerprints. As a matter of fact, an algorithm should be defined to assign high similarity scores to different captured fingerprints from the same finger even if they have high intra fingerprint variation. If the similarity could be defined in such a way that it couldcover high variation between fingerprints of the same finger and low variation between fingerprints of different fingers, then the system error could be maintained close to zero. In other words, the fingerprints from the same finger are supposed to be more similar to each other compared to the fingerprints from different fingers and the similarity measurement needs to be defined as such to assign high similarity to the fingerprints of the same finger regardless of the variability in the capture environment which introduces unpredictable errors in the captured image; and also assigns low similarity to the fingerprints of a different finger even if they appear to be very indistinguishable. The proposed similarity measurement method is based on the texture characteristic of the fingerprint which is claimed by researchers to provide more reliable and distinguishing information. Reliability and distinguishability of the texture based features compared with other available features is discussed. Also the different texture features are compared and investigated in order to identify the best texture/region based feature of the fingerprint. As a result the similarity is mainly computed based on the Normalized Cross Correlation (NCC) of the fingerprint (sub-regions) along with other techniques to lower the effect of high intra-class variation and low-inter class similarity and consequently, better distinguishing the intra and inter fingerprints even in low quality partial fingerprint. Also, the similarity of the two fingerprints is computed in such a way that the distorted regions are assigned low weights so they will not significantly skew the final similarity of two fingerprints. The main objective of this research relates to partial fingerprint recognition and the proposed method reasonably addresses the difficulties in partial fingerprint (as well as full fingerprint) matching. In all parts, the aim was to make it independent from any particular feature of the fingerprint to adaptively deal with the partial fingerprint size. The proposed method uses the fingerprint ridge structure to obtain the characteristic information of the partial fingerprints. As mentioned, fingerprints ridge structure provides rich and distinguishing information and atthe same time the similarity of partial fingerprint can be computed independently from region size, shape, and location in the fingerprint. Thus, unavailability of any particular feature such as reference points and minutiae will not be an issue and even if a small region is available, the matching can proceed. Also, there is always a mis-detection rate (however low) when extracting features like reference points and minutiae which might lead to falsely accepting or rejecting a query fingerprint. This is not an issue in the proposed method as it does not rely on any particular feature. As a result, the experiments conducted on partial dataset generated from the FVC2002 (Fingerprint Verification Competition) dataset shows the effectiveness and improvement achieved through the proposed method with approximately an 8.5% better Equal Error Rate (in average) while being able to process almost 3 times more of the partial fingerprints in the dataset (regardless of their size) compared to the state-of-the-art method proposed by Abraham et al.

History

Campus location

Australia

Principal supervisor

Nandita Bhattacharjee

Additional supervisor 1

Bala Srinivasan

Year of Award

2016

Department, School or Centre

Clayton School of Information Technology

Course

Doctor of Philosophy

Degree Type

DOCTORATE

Faculty

Faculty of Information Technology