Fingerprint Recognition System using Scale Invariant Feature Transform
Abbas H. Hassin Alasadi*1 and Roqaya Hamad Jaffar 2
Department of Computer Science, College of Information Technology, Basra University, Iraq
Email2: rhjaafar@yahoo.com
*Corresponding author
Abstract:
The fingerprint is a powerfully applicable tool for person authentication in commercial business, civil, and forensic usage. Minutiae points and ridge patterns consider the main source of features that mostly used in recent fingerprint identification systems. In this paper, Scale Invariant Feature Transform (SIFT) is extensively used for representation and extraction of features. After that, RANSAC algorithm is used for determining the matching area exactly. Database of International Fingerprint Verification Competition (FVC2000 and FVC2002) has been used in experimental results. Finally, the final results were compared with other works. Empirically, the results of the proposed algorithm were acceptable and better.
Keywords: Fingerprint matching, Feature extraction, SIFT, RANSAC.
1 Introduction
Nowadays, many methods, such as iris, face, fingerprints, veins, hand geometry, voice, retina, handwriting recognition have been implemented as biometrics authenticate identity system.
Because of its uniqueness, acceptability, and low cost, the fingerprint is one of the widespread biometrics that increasingly utilizes [1]. The fingerprint is one of the tools that are used often for personal determining in both forensic application and civilian.
Because of the 'ngerprints database is huge size, the recognition process encounters a time problem for searching the person's distinctiveness inside the database. Consequently, the reducing the time of processing are main goal and challenges of any fingerprint systems. a reduction of processing of identi'cation time can be performed by managing two factors, that is, ' the process time of a single matching system' and 'the number of matching system'.
Methods of fingerprint recognition can be mainly classified as texture based or minutiae based which all depend on one of the features; i) minutiae, ii) correlation and iii) ridge features. [2,3,4,5, 6,7,8,9].
Consequently, using additional feature may be required for accuracy matching in a fingerprint image. More than1400 pores may be found in one image of a fingerprint. According to [10], only 20 to 40 pores are required for fingerprint identification.
This paper proposes an approach for fingerprint matching using Scale Invariant Feature Transform (SIFT) describing fingerprint image features. In processing images, Lowe provided the SIFT algorithm [11] as a method for extraction of distinguished invariant features. It has been profitably applied to various problems of depended on feature matching, which involves processes such as recognizing objects, estimating poses, retrieving the image.
The paper's rest was organized as the following. Section 2 discusses SIFT algorithm structure. Section 3 describes the proposed fingerprint identification approach using SIFT. Section 4 explores experimental results, and A conclusion was given in Section 5.
2 Database
In this paper, database of International Fingerprint Verification Competition (FVC2000 and FVC2002) has used in experimental results. Images are collected by the Biometric System Lab (University of Bologna), the Pattern Recognition and Image Processing Laboratory of Michigan State University and the U.S. National Biometric Test Center (San Jose State University). The database consists of four groups of images. Every group has 80 images, 8 images for every person. Figure (1) exhibits a sample of FVC2000 database. For more specific distinctive of these two databases are brief in Table (1) and Table (2).
Group Image1 Image2 Image3 Image4 Image5 Image6 Image7 Image8
DB-1
DB-2
DB-3
DB-4
Figure (1) : Sample of FVC2000 database.
Table (2): Description of FVC 2002 DB1 and DB2 databases [12].
Sensor Type Image Size Number of images Resolution
DB1 Optical Sensor 388×374 (142K pixels) 10×8 500 dpi
DB2 Optical Sensor 296×560 (162K pixels) 10×8 569 dpi
Table (1): explanation of FVC 2000 DB1, DB2, DB3, and DB4 databases [12].
Sensor Type Image Size Number of images Resolution
DB1 Low cost- Optical Sensor 300 x300 10×8 500 dpi
DB2 Low Cost-Capacitive Sensor 256×364 10×8 500 dpi
DB3 Optical Sensor 448×478 10×8 500 dpi
DB4 Synthetic Generator 240×320 10×8 About 500dpi
3 Fingerprint Matching System (FPMS)
There are two stages of the phase proposed structure of fingerprint matching approach: the testing and training. Each stage has specific functions, so all functions have explained in detail as the following subsections. The testing phase and the training phase are the same, but, in the testing phase, the features do not put in the database just entered to matching system. Figure (2) describes the block diagrams of the training and testing stages of FPMS respectively.
3.1 Preprocessing
In this paper, the data sets from FVC2000 and FVC2002 have collected. The database contain 740 images are captured in the TIF format, with different resolution and different size. Consequently, two steps of preprocessing are prepared. First step is converting the TIF format into JPG. Second is size normalization to 200×200 pixels. Furthermore, contrast and brightness both are corrected.
For reducing the time consuming, we merged the 8 images for each person into one image with converting it into gray color.
Figure (2): A block diagrams of the training and testing stages of FPMS.
3.2 Feature Extraction
Feature extraction is considered as a substantial step in fingerprint discrimination, whereas the preprocessor results have utilized as a guide for features extraction. However, SIFT algorithm was applied for all fingerprint images. The SIFT algorithm involves four major stages for detection and description of local features, or key points, in the image [11]
1. Scale-space extrema detection
The SIFT algorithm begins by identifying the locations of candidate keypoints as the local maxima and minima of a deference-of-Gaussian pyramid that approximates the second-order derivatives of the image's scale space.
2. Keypoint localization and filtering
After candidate keypoints are identified, their locations in scale space are interpolated to sub-unit accuracy, and interpolated keypoints with low contrast or a high edge response| computed based.
3. Orientation assignment
The keypoints that survive filtering are assigned one or more canonical orientations based on the dominant directions of the local scale-space gradients. After orientation assignment, each keypoint's descriptor can be computed relative to the keypoint's location, scale, and orientation to provide invariance to these transformations.
4. Descriptor computation
Finally, a descriptor is computed for each keypoint by partitioning the scale-space region around the keypoint into a grid, computing a histogram of local gradient directions within each grid square, and concatenating those histograms into a vector. To provide invariance to illumination change, each descriptor vector is normalized to unit length, threshold to reduce the amount of large gradient values, and then renormalized.
3.3 Background of SIFT
In this section, we describe in details the proposed method to detect duplicated and distorted regions in an image. Figure (3) illustrating the main steps of SIFT based method using the preprocessed image of a given fingerprint.
A. Finding Image Keypoints and Collecting Features
We detect duplicated regions in the illumination domain, so RGB images are first converted to grayscale images using standard color space conversion. The first step in our method is to find image keypoints and collect image features at the detected keypoints.
Keypoints are locations carrying distinct information of the image content. Each keypoint is characterized by a feature vector that consists of a set of image statistics collected at the local neighborhood of the corresponding keypoint. Worthy keypoints and features should represent distinct locations in an image, be efficient to compute and robust to local geometrical distortion, illumination variations, noise and other degradations. The SIFT keypoints are found by searching for locations that are stable local extrema in the scale space[13]. At each keypoint, a 128 dimensional feature vector is generated from the histograms of local gradients in its neighborhood.
To ensure the obtained feature vector invariant to rotation and scaling, the size of the neighborhood is determined by the dominant scale of the keypoint, and all gradients within are aligned with the keypoint's dominant orientation dominant orientation. Furthermore, the obtained histograms are normalized to unit length, which renders the feature vector invariant to local illumination changes.
As duplicated regions typically account for only small fraction of the total area of the image, we limit keypoint detection to a small range of scales. In our experiment, we construct the scale space with Gaussian kernels of initial width of 1.6 pixels up to 7 octaves.
B. Putative Keypoint Matching
The detected SIFT keypoints are then uncertainly matched based on their feature vectors using the best-bin-first algorithm. For a keypoint at location with feature , we match it with keypoint , whose corresponding feature vector is the nearest neighbor to measure with their (Euclidean) distance. Due to the smoothness of natural images, the best match of a keypoint usually lies within its close spatial adjacency. To avoid searching nearest neighbors of a keypoint from the same region, we perform the search outside a pixels window centered at the keypoint. Further, many keypoints can match with each other, but we only keep those with distinct similarities. Specifically, we require that for any other feature vector other than and , the distance between and has to be smaller than that of f and f 0 by at least a factor of , as , where is a preset threshold controlling the distinctiveness of the matching. We use a default to provide a good trade-off between matching accuracy and ratio of outliers.
Figure (3): The main steps of SIFT based method
C. Estimation of Affine Transform Between Matched Keypoints
Next, based on the putative keypoint matching, we estimate the possible geometric distortions of the duplicated regions. To generalize transforms such as rotation, scaling and shearing that are supported in most photo-editing software, we model the distortion as affine transform of pixel.
D. Robust Estimation of Affine Transform
We can use the putative matching of SIFT keypoints to estimate the affine transform parameters, but the obtained results are inaccurate due to the large number of mismatched keypoints. To prune out unreliable keypoint correspondences and obtain accurate transform parameters simultaneously, we employ a widely used robust estimation method known as the Random Sample Consensus (RANSAC) algorithm [14].
4 Experimental Results
A lot of published works used FVC2000 DB2, DB1, DB3, and DB4 databases in order to evaluate their algorithms performance; this paper depends on this databases as well. Each database composed of images of 100 different fingers with 8 impressions of 500-dpi resolution for every fingerprint.
1. Positive case : Figure(4A) computes the SIFT for the fingerprint and another image for eight fingerprints; firstly, we compute the keypoints for fingerprint which equals to 7547. Figure(4B) determines the descriptor that equals to 128*7547. Thirdly, Figure(4C) computes the second neighbor matching between two images (2nn) that matches to 107 features. Finally, Figure(4D) computes the geometric transform (geo) which matches to 43 features.
A B C D
Figure (4) : Experimental results as a positive case.
2. Negative case : Figure(5A) computes the SIFT for the fingerprint and another image for eight fingerprints; firstly, we computes the keypoints for fingerprint that equals to 8537. Secondly, Figure(5B) determines the descriptor which equals to 128*8537. Thirdly, Figure(5C) computes the second neighbor matching between two images (2nn) that matches to 2 features. Finally, Figure(5D) computes the geometric transform (geo) which matches to 0 features. That is mean no matching.
A B C D
Figure (5) : Experimental results as a negative case.
Lastly, for checking the performance of experimental result, we computes the whole fingerprint images in DB1 at once. Figure(6A) computes the SIFT for all fingerprints and another image for eight fingerprints; firstly, we computes the keypoints for fingerprint that equals to 49635. Secondly, Figure(6B) determines the descriptor which equals to 128*49635. Thirdly, Figure(6C) computes the second neighbour matching between two images (2nn) that matches to 995 features. Finally, Figure(6D) computes the geometric transform (geo) which matches to 620 features. There is a matching.
Figure (6) : Experimental results for all DB1 fingerprint images.
5 Performance Methods
Roc curve is a fundamental tool for diagnostic test evaluation. It is a graph which is used for summarizing the efficiency of the classifier over every conceivable threshold. It's plotting the false positive ratio (FPR) and true positive ratio (TPR) against the false negative ratio (FNR) and true negative ration (TNR) as user changes threshold to assign observations for a specific class [15].
In the proposed system, the efficiency can be classified into the following perspectives:
The sensitivity (TRUE POSITIVE )of a recognized test is the fraction of positive cases over the total of afflicted cases, which can be expressed by:
(1)
A test with a high value of sensitivity must have a minimal number of false negatives and is therefore useful in order to recognize the fingerprint.
The specificity (TRUE NEGATIVE ) of a test is the fraction of healthy cases over the total of un-afflicted cases, which can be expressed by:
(2)
A test with a high value of specificity must have a minimal number of false positives and is therefore useful to exclude the wrong fingerprint.
There are other criteria that include those four perspectives, such as accuracy that is the measure of the global performance of the algorithm about the correct decisions and precision which corresponds to the fraction of relevant recognitions:
(3)
(4)
Table (3) displays the performance measure which applied on FVC2000 databases. Simultaneously, Table (4) displays the FVC2002 database.
Table (3): The performance measure which applied on FVC2000 databases.
Criteria DB1 DB2 DB3 DB4 ALL
Sensitivity 0.93 0.94 0.89 0.94 0.92
Accuracy 0.88 0.90 0.86 0.90 0.89
Precision 0.83 0.86 0.82 0.86 0.84
Simplicity 0.85 0.87 0.83 0.87 0.85
Table (3): The performance measure which applied on FVC2002 databases.
Criteria DB1 DB2 ALL
Sensitivity 0.92 0.92 0.92
Accuracy 0.86 0.86 0.86
Precision 0.76 0.8 0.78
Simplicity 0.81 0.82 0.81
6 Proposed System vs. related work system
The implementation of proposed system has illustrated by two types of features and compared with other related system as in Table (4).
Table (4): Proposed System vs. related work system.
Algorithm EER (%)
Compensatory algorithm 0.378
FVC2002/ PA15 0.19
FVC2002 /PA27 0.33
Proposed Algorithm Fvc2002/Db1 0.21
Proposed Algorithm Fvc2002/Db2 0.20
7 Conclusion
In this paper, SIFT has been used for fingerprint feature extracting and verification . It is invariant according to picture scaling, rotation. Fingerprint matching is performed in two stages:
i) Feature extract for matching and
ii) Determining false fingerprint matches with estimation geometric translation.
Improving the performance of SIFT is done by reduction the image noise. Typically, pre-processing step start by converting the given image into grey level and then normalize all images into fixed sizes. As a feature extraction step, determine the key points that used in the SIFT operator. These key points are extraction by usability computing Gaussian and DoG pyramid. The key points of low contrast or are noxious placed alongside an area will below stand removed. The specialty is accomplished by way of the usage of a high dimensional vector. When the use of SIFT within fingerprint discrimination, the range regarding key points are extracted. totally concerning the attribute regarding an image of a fingerprint. After computing SIFT keypoints and determine the descriptors has computed the matching steps. Particularly we compare the same fingerprint with the different impression. When we applied SIFT algorithm should be used other algorithms because the huge computation, which means that it is better to usage another procedure such as RANSAC algorithm to determine matching area exactly.
References
[ 1] Jain, Anil, Arun A. Ross, and Karthik Nandakumar. Introduction to biometrics. Springer Science & Business Media, 2011.
[ 2] Stoianov, Alex, Colin Soutar, and Allan Graham. "High-speed fingerprint verification using an optical correlator." Optical Engineering 38.1 (1999): 99-107.
[ 3] Bolle, Rudolf Maarten, et al. "Fingerprint representation using localized texture feature." U.S. Patent No. 8,180,121. 15 May 2012.
[ 4] Jain, Anil K., Salil Prabhakar, and Shaoyun Chen. "Combining multiple matchers for a high security fingerprint verification system." Pattern Recognition Letters 20.11 (1999): 1371-1379.
[ 5] Willis, Andrew John, and L. Myers. "A cost-effective fingerprint recognition system for use with low-quality prints and damaged fingertips." Pattern recognition 34.2 (2001): 255-270.
[ 6] Jain, Anil K., et al. "Filterbank-based fingerprint matching." IEEE transactions on Image Processing 9.5 (2000): 846-859.
[ 7] Cappelli, Raffaele, Matteo Ferrara, and Davide Maltoni. "Minutiae-based fingerprint matching." Cross Disciplinary Biometric Systems. Springer Berlin Heidelberg, 2012. 117-150.
[ 8] Jain, Anil, Lin Hong, and Ruud Bolle. "On-line fingerprint verification." IEEE transactions on pattern analysis and machine intelligence 19.4 (1997): 302-314.
[ 9] Maltoni, Davide, Raffaele Cappelli, and Didier Meuwly. "Automated Fingerprint Identification Systems: From Fingerprints to Fingermarks." Handbook of Biometrics for Forensic Science. Springer International Publishing, 2017. 37-61.
[ 10] Mathur, Surbhi, et al. "Methodology for partial fingerprint enrollment and authentication on mobile devices." Biometrics (ICB), 2016 International Conference on. IEEE, 2016.
[ 11] D.G.Lowe, 'Distinctive Image Features from Scale-Invariant Keypoints,' International Journal of Computer Vision, 60(2), 91-110, 2004.
[ 12] D. Maio, D. Maltoni, J. L. Wayman, and A. K. Jain, FVC2002: Second Fingerprint Verification Competition, International Conference on Pattern Recognition, 811'814, 2002.
[ 13] Beis, Jeffrey S., and David G. Lowe. "Shape indexing using approximate nearest-neighbour search in high-dimensional spaces." Computer Vision and Pattern Recognition, 1997. Proceedings., 1997 IEEE Computer Society Conference on. IEEE, 1997.
[ 14] Fischler, Martin A., and Robert C. Bolles. "Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography." Communications of the ACM 24.6 (1981): 381-395.
[ 15] Parikh, Chirag R., and Heather Thiessen-Philbrook. Key concepts and limitations of statistical methods for evaluating biomarkers of kidney disease. Journal of the American Society of Nephrology, 2014, 25(8):1621-1629.