Fig-1.
International Journal of Scientific & Engineering Research, Volume 3, Issue 11, November-2012
ISSN 2229-5518
A Case Study on Multi-instance Finger Knuckle
Print Score and Decision Level Fusions
Harbi AlMahafzah, H.S. Sheshadri, and Mohammad Imran
ABSTRACT
Abstract—This paper proposed the use of multi-instance as a means to improve the performance of Finger Knuckle Print (FKP) verification. A log- Gabor filter has been used to extract the image local orientation information, and represent the FKP features. Experiments are performed using the DZhang FKP database, which consists of 7,920 images. The influence on biometric performance using Decision and Score Level Fusions has been demonstrated in this paper. Results indicate that the Multi-instance verification approach at Score Level (Max Rule) and Decision Level (OR Rule) outperforms higher performance than using any single instance. Whereas at Score Level (Min Rule) and Decision Level (AND Rule) does not have performance improvement.
Index Terms—: Decision Level Fusion, Multi-Biometric; Multi-instance; Log-Gabor; Score Level Fusion.
— — — — — — — — — — — — — — — — — — — —
he need for reliable user authentication techniques has increased in the wake of heightened concerns about secu- rity and rapid advancements in networking, communica-
tion, and mobility. A wide variety of applications require reli- able verification schemes to confirm the identity of an indi- vidual requesting their service. Examples of such applications include; secure access to buildings, computer systems, laptops, cellular phones, and ATMs. In the absence of robust verifica- tion schemes, these systems are vulnerable to the wiles of an impostor [1],[2],[10]. Traditionally, passwords (knowledge- based security) and ID cards (token-based security) have been used to restrict access to applications. However, in these ap- plications security can be easily breached when a password is divulged to an unauthorized user or a badge is stolen by an impostor. The emergence of biometrics has addressed the problems that plague the traditional verification methods.
The term biometric comes from the Greek words bios (life)
and metrikos (measure) [10]. Biometric refers to the automatic
recognition of individuals based on their physiological and
behavioural characteristics. Biometrics systems are commonly
classified into two categories: physiological biometrics and
behavioral biometrics. Physiological biometrics (fingerprint, iris, retina, hand geometry, face, etc) use measurements from
the human body. Behavioural biometrics (signature, keys- trokes, voice, etc) use dynamics measurements based on hu- man actions [1],[3].
————————————————
P.E.T Foundation Research, University of Mysore, Mandya, India. E-mail: hmahafzah@hotmail.com
These systems are based on pattern recognition methodolo- gy, which follows the acquisition of the biometric data by building a biometric feature set, and comparing versus a pre- stored template pattern. These are unimodal which rely on the evidence of a single source of information for authentication, which have to contend with a variety of problems such as (noise in sensed data, intra-class variations, and inter-class similarities, etc).
Possible solutions to compensate for the false classification problem due to intra-class variability and inter-class similarity can be found in the fusion of biometric systems or experts [8]. Fusion based on multi-biometrics can also be used to improve performance in a practical system. The fusions of biometric refer as multibiometrics. The term multibiometrics denotes the fusion of different types of information (e.g., fingerprint and face of the same person, or fingerprints from two different fingers of a person). Thus this paper evaluates the perfor- mance of multi-instance approach by fusing the data at match score and decision levels using different rules. The rest of the paper is organized as follows: Section 2 presents related works, proposed method is given in section 3, experimental results are given in Section 4, and conclusion is mentioned in Section 5.
Lin Zhang et al.[4] proposed an effective FKP recognition scheme by extracting and assembling local and global features of FKP images. Specifically, the orientation information ex- tracted by the Gabor filters is coded as the local feature. The proposed scheme exploits both local and global information for the FKP verification. The authors experimental results conducted on FKP database indicate that the proposed scheme could achieve much better performance in terms of EER and the decidability index than the other state-of- the-art competi- tors.
T.C. Faltemier et al. [6] proposed a multi-instance enroll- ment for face recognition as a means to improve the perfor- mance of 3D face recognition. The authors show that using
IJSER © 2012
International Journal of Scientific & Engineering Research Volume 3, Issue 11, November -2012
ISSN 2229-5518
multiple images to enroll a person in a gallery can improve the overall performance of a biometric system. The authors dem- onstrated that when using multiple images to enroll a person, sampling from different expressions improves performance over sampling only the same expression.
M. Vatsa, et al. [7] proposed a generalized biometric match score fusion framework using belief function theory. Multi- instance iris verification and multi-unit iris verification are used as the two case. Experimental results show that the pro- posed fusion framework with PCR rule can effectively fuse the match scores even when the individual biometric classifiers provide highly conflicting match scores
Tobias Scheidat, et al. [8] proposed a fusion of two in- stances of the same semantic, where semantics are alternative handwritten contents such as numbers or sentences, in addi- tion to commonly used signature. The fusion is carried out by the combination of the matching scores of two instances of one handwritten semantic. The authors demonstrated that when using three semantics and fusion strategies, improvements can be observed in comparison to the best individual results.
The term multibiometrics denotes the fusion of different types of information. Multibiometric systems can offer substantial improvement in the matching accuracy of a biometric system depending upon the information being combined and the fu- sion methodology adopted [1]. Multi-sensor: Multiple sensors can be used to collect the same biometric. Multi-modal: Mul- tiple biometric modalities can be collected from the same indi- vidual, e.g. fingerprint and face, which requires different sen- sors. Multi-sample: Multiple readings of the same biometric are collected during the enrolment and/or recognition phases, e.g. a number of fingerprint readings are taken from the same finger. Multi-algorithms: Multiple algorithms for feature ex- traction and matching are used on the same biometric trait. Multi-instance: Multi-instance biometrics means the use of the same type of raw biometric sample and processing on multiple instances of similar body parts, (such as two fingers, or two irises) also been referred to as multi-unit systems in the literature [9]. These systems can be cost-effective since a single sensor is used to acquire the multi-unit data in a sequential fashion, and these systems generally do not necessitate the introduction of new sensors nor do they entail the develop- ment of new feature extraction and matching algorithms. Mul- ti-instance systems are especially beneficial to users whose biometric traits cannot be reliably captured due to inherent problems. Multi-instance systems are often necessary in appli- cations where the size of the system database (i.e., the number of enrolled individuals) is very large (FBI’s database currently has 50 million ten-print images, and multiple fingers provide additional discriminatory information) [9]. Combination of multi-instances can improve the performance of the biometric system.
Based on the structure of a biometric authentication, there are
four levels within the process to fuse biometric systems, they are [11],[12]: Sensor Level Fusion; entails the consolidation of evidence presented by multiple sources of raw data before they are subjected to feature extraction. Sensor level fusion can benefit multi-sample systems which capture multiple snap- shots of the same biometric. Feature Level Fusion; consolidat- ing the feature sets obtained from multiple biometric algo- rithms into a single feature set (vector), after normalization, transformation and reduction schemes.
Decision Level Fusion; falls under a broader area known as distributed detection systems and is the process of selecting one hypothesis from multiple M hypotheses given the deci- sions of multiple N sensors in the presence of noise and inter- ference. In biometrics, decision level fusion creates a single decision from typically two hypotheses, imposter or genuine user, from multiple biometric sensor decisions, which may or may not be identical sensors. Often, decision level fusion is implemented to save communication bandwidth as well as improve decision accuracy. A statistical performance model for each biometric sensor is needed a prior to support the sys- tem wide optimization in terms of two error rates: false accep- tance rate (FAR), admitting an imposter, and false rejection rate (FRR), rejecting the genuine user.
In all the experiment, the data have been fused at score and decision level fusions using different rules (decision “OR and AND” rules, score “Max and Min” rules) for two and three instances combination out of four fingers.
In this paper a hand-based biometric technique, finger- knuckle-print (FKP) have been used, FKP refers to the image pattern of the outer surface around the phalangeal joint of one’s finger, which is formed by bending slightly the finger knuckle [4]. The experiments are developed for personal au- thentication using DZhang FKP data base. FKP images were collected from 165 volunteers, including 125 males and 40 fe- males. The database contains FKPs from four different fingers, right index, right middle, left index, and left middle fingers. The DZhang database is available at the website of Biometrics Research Centre, the Hong Kong Polytechnic University.
This section describe the Region of Interest (ROI) extraction, the process involved to extract ROI for each instance is as fol- lows. It is necessary and critical to align FKP images by adap- tively constructing a local coordinate system for each image. With such a coordinate system, ROI can be cropped from the
IJSER © 2012
International Journal of Scientific & Engineering Research Volume 3, Issue 11, November -2012
ISSN 2229-5518
original image using the steps suggested in [4], as shown in
Fig-1.
Fig-1: a)Image acquisition device is being used to collect FKP samples b)Sample FKP image c)ROI coordinate system, where the rectangle indicates the ROI area d)Extracted ROI
Log-Gabor has been used as feature extraction algorithm; it eliminates the limitations in Gabor filters. Log-Gabor func- tions, by definition, always have no DC component, and, the transfer function of the log Gabor function has an extended tail at the high frequency end. Log-Gabor function proposed by Field [5], Field suggests that natural images are better coded by filters that have Gaussian transfer functions when viewed on the logarithmic frequency scale. On the linear fre- quency scale log-Gabor has a transform function of the form
G(w) e( log(w / w0 ) ) /(2(log(k / w0 ) )
where w0 is the filter's centre frequency. To obtain constant shape ratio filters the term k/w0 must be held constant for varying w0.
This section deals with the investigation consequences com- parison of combining many biometrics at score and decision level fusion to measure the performance of multi-instance sys- tem. In all the experiments, performance is measured in terms of False Acceptance Rate (FAR in %) and corresponding Ge- nuine Acceptance Rate (GAR in %). First the performance of a single instance biometric system is measured; later the results for multi-instance biometric system are evaluated. The results obtained from single instance biometric system are tabulated in Table-1, and depicted as Receiver Operating Characteristic (ROC) curve in Figure-2.
TABLE 1
SINGLE INSTANCE PERFORMANCE
From Table-1 it can be observed that the right middle finger has more score performance than the other fingers.
Fig-2: The ROC Curve Performance of Single Instance.
Go further the experimental results of the fusion of two and three instances, under Decision and Score Level Fusions, are shown. Table-2 shows the fusion result of two instances and Table-3 shows the fusion result of three instances.
TABLE 2
THE FUSION RESULT OF TWO INSTANCES
FAR (%) | GAR (%) Decision Level (OR Rule) | |||||
FAR (%) | RI+RM | RI+LI | RI+LM | RM+LI | RM+LM | LI+LM |
0.01 | 88.67 | 72.00 | 76.67 | 77.67 | 81.67 | 81.00 |
0.10 | 91.00 | 80.00 | 85.33 | 85.33 | 87.00 | 86.00 |
1.00 | 94.67 | 95.00 | 94.00 | 93.67 | 95.00 | 94.00 |
GAR (%) Decision Level (AND Rule) | ||||||
0.01 | 72.33 | 66.00 | 67.00 | 67.67 | 67.67 | 64.00 |
0.10 | 79.67 | 78.00 | 78.00 | 78.33 | 76.67 | 74.33 |
1.00 | 86.00 | 86.67 | 85.67 | 88.33 | 86.67 | 85.33 |
GAR (%) Score Level (Max Rule) | ||||||
0.01 | 86.33 | 72.00 | 76.67 | 77.67 | 81.67 | 81.00 |
0.10 | 91.00 | 81.33 | 85.00 | 85.33 | 87.00 | 86.33 |
1.00 | 94.67 | 95.00 | 94.00 | 94.67 | 95.00 | 94.00 |
GAR (%) Score Level (Min Rule) | ||||||
0.01 | 72.33 | 66.00 | 67.00 | 73.67 | 67.67 | 64.67 |
0.10 | 79.67 | 78.67 | 78.33 | 78.33 | 76.67 | 74.67 |
1.00 | 86.00 | 86.67 | 85.67 | 90.33 | 86.67 | 85.33 |
From Table-2 it can be observed that the fusion of two in- stances finger at score level (Max rule) and decision level (OR rule) has a significant improve score over the single instance. Whereas at score level (Min rule) decision level (AND rule) does not have performance improvement over a single in- stance, since both matchers has to agree for the decision to be true (in case of AND). For this reason the performance will be like the best matcher. Fig-3 shows the ROC curve for the per- formance fusion of two instances.
From Table-3 it can be observed that the fusion of three in- stances at score level (Max rule) and decision level (OR rule) has more improvement over the fusion of two instances, with some better performance of level (Max rule) over decision lev- el (OR rule). Whereas at score level (Min rule) and at decision level (AND rule) does not have performance improvement
IJSER © 2012
International Journal of Scientific & Engineering Research Volume 3, Issue 11, November -2012
ISSN 2229-5518
even over a single instance for the same reason shown above. Fig-4 shows the ROC curve for the performance of the fusion of three instances.
TABLE 3
THE FUSION RESULT OF THREE INSTANCES
Analyzing the performance of using FKP images as a biome- tric has been done. From the analysis of experimental results and observations it can be concluded that a multi-instance biometric fusion is given better performance at score level
FAR
GAR (%)Decision Level (OR Rule)
(Max rule) and decision level (OR rule) than single instance.
Whereas at score level (Min rule) and at decision level (AND rule) does not have performance improvement. This shows that using multiple instance of biometric which collected using single sensor; can have the security level. However, from the experimental results and observations the degree of improve- ment in accuracy by fusing multiple instances is marginal. Since different instance of the same trait produces the same redundant features.
Fig-3 The ROC curve Performance at Score and Decision Level Fusions combination of two Instances a) Score Level (Max Rule)
b) Score Level (Min Rule) c) Decision Level (OR Rule) d) Decision Level (AND Rule).
IJSER © 2012
International Journal of Scientific & Engineering Research Volume 3, Issue 11, November -2012
ISSN 2229-5518
Fig-4: The ROC curve Performance at Score and Decision Level Fusions combination of three Instances a) Score Level (Max Rule)
b) Score Level (Min Rule) c) Decision Level (OR Rule) d) Decision Level (AND Rule).
[1] Andrew Teoh, S. A. Samad and A. Hussain, “Nearest Neighbour- hood Classifiers in a Bimodal Biometric Verification System Fusion Decision”, Journal of Research and Practice in Information Technolo- gy, Vol. 36, No. 1, February 2004.
[2] George Chellin Chandra. J and Rajesh. R.S., “Performance Analysis
of Multimodal Biometric System Authentication”, IJCSNS 290 Inter- national Journal of Computer Science and Network Security, VOL.9
No.3, March 2009.
[3] A.Ross, K.Nandakumar, and A.K. Jain, “Handbook of Multibiome-
trics”, Springer-Verlag edition, 2006.khj
[4] Lin Zhang, Lei Zhang, David Zhang, and Hailong Zhu “Ensemble of local and global information for finger–knuckle-print recognition”. Elseveir / Pattern Recognition 44 (2011) 1990–1998
[5] D.J. Field. “Relation between the statistics of natural images and the
response properties of cortical cells”. J. Opt. Soc. Am. A,
4(12):2379_2394, 1987.
[6] T.C. Faltemier, K.W. Bowyer, and P.J. Flynn “Using multi-instance enrollment to improve performance of 3D face recognition”. Elseveir
/ Computer Vision and Image Understanding 112 (2008) 114–125.
[7] M. Vatsa, R. Singh, A. Noore, and S.K. Singh “Belief Function Theory based Biometric Match Score Fusion: Case Studies in Multi-instance and Multi-unit Iris Verification” Seventh International Conference on Advances in Pattern Recognition IEEE DOI 10.1109/ICAPR.2009.98
[8] Tobias Scheidat, Claus Vielhauer, Jana Dittmann “Single-Semantic Multi-Instance Fusion of Handwriting Based Biometric Authentica- tion Systems” IEEE ICIP 2007
[9] David Zhang, Fengxi Song, Yong Xu, and Zhizhen Liang “Advanced
Pattern Recognition Technologies with Applications to Biometrics”.
[10] Kresimir Delac and Mislav Grgic, “A Survey of Biometric Recogni- tion Methods” 46th International Symposium Electronics in Marine, ELMAR-2004, 16-18 June 2004, Zadar, Croatia
[11] Arun Ross and Rohin Govindarajanb, “Feature Level Fusion Using Hand and Face Biometrics”, SPIE Conference on Biometric Technol- ogy for Human Identification II, Volume, 5779, pp.196-204(Orlando, USA) March 2005.
[12] Stan Z. Li, Anil K. Jain “Encyclopedia of Biometrics” Springer.
IJSER © 2012