Optical fingerprint recognition based on local minutiae structure coding

Yao Yi; Liangcai Cao; Wei Guo; Yaping Luo; Jianjiang Feng; Qingsheng He; Guofan Jin

doi:10.1364/OE.21.017108

1. Introduction

Automatic fingerprint identification system has been widely used in forensic and civilian applications. The recognition algorithm itself and how the algorithm should be implemented are the two critical issues attracting most of the research interests. While many algorithms and corresponding digital implementation systems are currently available, the time consumption and I/O traffic are becoming the bottlenecks of large scale digital systems [1]. Optical recognition has the merit of parallel processing and high data transfer rate. Specially, associative retrieval of volume holographic optical correlator offers a unique approach for fingerprint recognition because of its high storage density, integration of storing and computing, and multi-channel parallel filtering ability [2,3].

Previous studies have proposed several optical fingerprint recognition systems, including systems based on joint transform correlator (JTC) and volume holographic correlator (VHC). Fielding et al. built an optical fingerprint identification system using a binary JTC [4]. Grycewicz et al. utilized Fourier plane binarization and output peak intensity normalization techniques to improve the performance of a JTC system for fingerprint recognition [5]. Bal et al. used a dynamic neural-network based supervised filtering technique to enhance the fingerprint and the fringe-adjusted JTC algorithm for the identification process [6]. Yan et al. and Lee et al. demonstrated multi-channel matched correlators using photorefractive material that can output multiple correlation results with single input [7,8]. Watanabe et al. proposed an ultrahigh-speed compact optical correlation system using volume holographic disc that can be used for fingerprint recognition [9].

The systems and methods described above show promising prospects of optical fingerprint recognition in the aspects of parallel processing and easy gallery data accessing. They perform well for pre-aligned, undeformed fingerprint images. However, the performances degrade greatly when fingerprint translation, rotation and nonlinear skin distortion occur, which is very common in practice. The advanced correlation filters combining multiple training images proposed by Kumar et al. [10] alleviated this problem, but it still needs multiple training templates and fingerprint alignment. The reason of performance degradation of optical systems mainly lies on the fact that most of these systems treat fingerprint images as general gray scale or binary pictures when constructing fingerprint templates.

The robustness problem of fingerprint recognition was addressed in digital algorithms by using local minutiae structure matching in the last decade [11–16]. Minutiae are the ridge ends and bifurcations of the fingerprint as shown in Fig. 2(a). Local minutiae structures encode the relationship between a minutia and its neighboring minutiae in terms of distances and angles. These attributes are invariant with respect to translation and rotation, and need no global alignment. Besides, since nonlinear distortions are trivial in local areas of a fingerprint, this method is also robust to global nonlinear skin distortion. However, the local structure encoding require for a very large memory size of fingerprint template [16], which drastically increases the time consumption and the demand for high data transfer rate during matching process.

This paper integrates a multi-page local minutiae structure coding method for fingerprint template construction in a volume holographic optical fingerprint recognition system, aiming at combining the advantages of both optical processing (parallel processing and high data transfer rate) and local minutiae structure matching (robustness to deformation).

The rest of this paper is organized as follows: Section 2 presents the algorithm selection criteria for optical recognition and the local minutiae structure coding method. Section 3 focuses on the optical considerations including optimizing the coding parameters, improving the filtering recording time schedule and compensating channel nonuniformity by post-filtering calibration. In Section 4, experiments conducted on two fingerprint databases are reported. Section 5 draws the conclusion.

2. Template data page construction with local minutiae structure coding

How the template data pages for a fingerprint are constructed is crucial for an optical fingerprint recognition system. Although many local minutiae structure coding and matching algorithms [11–16] have been proposed to cope with fingerprint deformation, not all of them are suitable for optical implementation.

2.1 Algorithm selection criteria for optical recognition

In a 90-degree geometry VHC system which is shown in Fig. 1, if M data pages $f_{i} (x_{0}, y_{0}) (i = 1, 2, \dots M)$ are stored as filters in the photorefractive crystal by angular multiplexing, the output intensity corresponding to the ith data page on the focal plane of L1 can be expressed as:

I (f, f_{i}) = {[g (f, f_{i})]}^{2} \propto {[\iint d x_{0} d y_{0} f (x_{0}, y_{0}) f_{i}^{*} (x_{0}, y_{0})]}^{2}

when a query template (also called search argument)

f (x_{0}, y_{0})

is loaded to the spatial light modulator (SLM) in retrieval [17]. The correlation peak values between the search argument and all the stored gallery data pages are calculated in parallel. The intensity (grayscale) of each output correlation spot detected by the CCD stands for the square of the 2-D inner product.

Fig. 1 Demonstration of the recording and retrieval processes in an angular multiplexing volume holographic correlator. SLM: spatial light modulator; FT Lens: Fourier Transform Lens.

Download Full Size | PDF

The VHC architecture and working mode of optical correlation recognition exert several restrictions on the algorithms. The template encoding method and the corresponding matching algorithm should meet the following criteria: (1) The encoded template data page is a fixed-length representation which can be formatted into a 2-D image and modulated by the SLM. (2) The optical matching process has no further operation on the stored gallery templates such as translation and rotation once they are stored in the VHC. (3) The similarity between a gallery template and a query template can be evaluated based on the correlation between them.

2.2 Template data page construction with Minutiae Cylinder-Code (MCC)

After compared with many local minutiae matching algorithms [11–16], the MCC representation [16] could satisfy the previously described requirements. The basic idea of the MCC is to generate a minutia structure template for each minutia according to the spatial and directional relationships between the minutia and its fixed-radius (local) neighboring minutiae. Thus a fingerprint image is represented by $N_{m}$ minutiae template data pages, where $N_{m}$ is the minutiae number of that fingerprint. The fingerprint encoding method is named as multi-page local structure coding.

As is shown in Fig. 2, the neighborhood of minutia M is represented by a cylinder with radius $R$ and height $2 π$ . The base of the cylinder is centered on the minutia location $(x_{m}, y_{m})$ , and aligned to the minutia direction $θ_{m}$ . The cuboid enclosing the cylinder is discretized into $N_{C} = N_{A} \times N_{A} \times N_{H}$ cells. Thus each cell $(i, j, k)$ has a location $p_{i, j}$ and an angle $φ_{k}$ . A numerical value is calculated according to the following sigmoid equation:

ψ (i, j, k) = ψ (v) = Z (v, μ, τ) = \frac{1}{1 + \exp [τ (v - μ)]},

where

μ

and

τ

are two parameters that limit the contribution of dense minutiae clusters, and

v

is calculated by accumulating distance and angle contributions from all the neighboring minutiae around the cell (

m_{t} \in N_{p_{i, j}}

). The accumulating contribution is defined as follows:

v (i, j, k) = \sum_{m_{t} \in N_{p_{i, j}}} (C_{M}^{S} (m_{t}, p_{i, j}) \cdot C_{M}^{D} (m_{t}, φ_{k})) .

_{$C_{M}^{S} (m_{t}, p_{i, j})$} and

C_{M}^{D} (m_{t}, φ_{k})

are the distance and angle contribution minutia

m_{t}

gives to cell

(i, j, k)

:

C_{M}^{S} (m_{t}, p_{i, j}) = \frac{\exp (- d_{S}^{2} / 2 σ_{S}^{2})}{σ_{S} \sqrt{2 π}},

C_{M}^{D} (m_{t}, φ_{k}) = G_{D} (d ϕ (φ_{k}, d ϕ (θ_{M}, θ_{m_{t}}))),

where

d_{S}

is the Euclidean distance between

m_{t}

and

p_{i, j}

,

d ϕ (α_{1}, α_{2})

is the difference between two angles

α_{1}

and

α_{2}

, and

G_{D} (x)

is the integration of a Gaussian function.

Fig. 2 A graphic representation of the local minutiae structure associated to a given minutia M: (a) the skeletonized fingerprint with its minutiae labeled; (b) the discretized cylinder of minutia M. The cylinder is rotated horizontally so that axis i is aligned to the corresponding minutia direction.

Download Full Size | PDF

It can be inferred from the definition of $C_{M}^{S}$ , $C_{M}^{D}$ and $ψ$ that the value $ψ (i, j, k)$ attached to the cell $(i, j, k)$ represents the likelihood of finding minutiae near the cell within a certain directional difference interval. The calculation of $ψ (i, j, k)$ only involves the relative distance and angle difference, both of which are irrelevant to fingerprint translation as well as rotation. Thus there is no need for fingerprint alignment and registration when using the MCC coding method. Besides, since each minutia cylinder is calculated locally (limited to the disk with radius R around that minutia), the MCC representation is also robust to global deformation brought by the elasticity of skin. Figure 3 demonstrates the robustness of MCC representation. Although large variation occurs between the two fingerprints, the local structures and corresponding templates of minutia M are quite stable.

Fig. 3 Demonstration of MCC representation robust to fingerprint translation, rotation and elastic deformation. (a) and (b) are two impressions from the same finger. (c)-(f) are the gray-scale and binarized MCC templates for the corresponding minutia M labeled in the center of the circle in (a) and (b), respectively.

Download Full Size | PDF

After each minutia cylinder is calculated, a threshold can be chosen to binarize $ψ (i, j, k)$ , which will be discussed in Section 3.2. The fingerprint image is now totally represented by its multiple binarized minutiae templates (cylinders). These templates are further selected and reshaped to form data pages to adapt to the two-dimensional optical data channel. The characteristics of the VHC system is carefully considered during these procedures.

3. Optical considerations

When it comes from the algorithm to optical implementation, several challenges have to be overcome besides the criteria described in section 2.1. These challenges include the relationship between the inner product and the similarity, large variation of the template data page filling ratio, and channel nonuniformity. This is because unlike traditional “hit-or-miss” method, we have to know not only which data page (representing a minutia) in the gallery is the most similar to the query data page, but also how similar they are.

3.1 Similarity between two minutiae using inner product

Since the intensities of correlation spots, which stand for the square of inner products, are employed to identify matching pairs of minutiae, it is important to establish the relationship between the inner product and template similarity. Traditional optical recognition uses inner product directly to represent similarity between the gallery and probe templates. However, this similarity computing method causes a problem that a probe template with higher filling ratio, which is defined as the number of pixels ON divided by the total number of pixels, tends to get higher inner products with all the gallery templates. The output cannot reflect the real similarity when the filling factor is not uniform for all the data pages. Here we propose the following approach for similarity calculation.

The similarity of two vectors $v_{1}$ and $v_{2}$ with the same length can be defined as

γ (v_{1}, v_{2}) = 1 - \frac{{‖ v_{1} - v_{2} ‖}_{p}}{{‖ v_{1} ‖}_{p} + {‖ v_{2} ‖}_{p}},

where

{‖ v ‖}_{p}

denotes the p-norm of vector

v

. It can easily be inferred that

γ

is always in the range [0, 1]: One represents maximum similarity while zero denotes minimum similarity. If

v_{1}

and

v_{2}

are binarized vectors, and

p = 1

, Eq. (6) can be simplified as

γ (v_{1}, v_{2}) = \frac{2 \times \sum_{i = 1}^{N} [v_{1} (i) \cdot v_{2} (i)]}{N \times [f i l l (v_{1}) + f i l l (v_{2})]},

where

N

is the length of vector

v_{1}

and

v_{2}

,

f i l l (v_{1})

and

f i l l (v_{2})

are the filling ratios,

v_{1} (i)

and

v_{2} (i)

are the ith element of

v_{1}

and

v_{2}

, respectively. This expression can easily be extended to 2-D case if we substitute the two vectors with two 2-D minutia template data pages:

γ (T, T_{S}) = \frac{2 \sum_{i = 1}^{R o w} \sum_{j = 1}^{C o l} [T (i, j) \cdot T_{S} (i, j)]}{(R o w \cdot C o l) \cdot [f i l l (T) + f i l l (T_{S})]} = \frac{2 \cdot 〈 T, T_{S} 〉}{(R o w \cdot C o l) \cdot [f i l l (T) + f i l l (T_{S})]},

where

R o w

and

C o l

represent the numbers of rows and columns of a data page, and

〈 T, T_{S} 〉

is the inner product between

T

and

T_{S}

,

f i l l (T)

and

f i l l (T_{S})

are the filling ratios.

In Eq. (8) the filling ratio $f i l l (T_{S})$ of the search argument $T_{S}$ affects the numerator and denominator simultaneously. The change of inner product caused by the change of filling ratio is counteracted by the denominator, thus $γ (T, T_{S})$ can reflect the real similarity even when the filling ratios of different search arguments are not uniform.

3.2 Template data page binarization and filling ratio

Binarizing the template data pages provides better cross-correlation discrimination in the optical recognition system [4]. While choosing an appropriate filling ratio range, two factors are under consideration: optical energy passing through the SLM, and the discriminating power of minutiae templates. Low filling ratio may cause low signal-to-noise response both in the recording and retrieval process because of low optical energy pass. Meanwhile, if the filling ratio is too high, the differences between minutiae templates may be decreased because the inner products of both the matched and unmatched templates could be very high. Thus, the minutiae could be difficult to be discriminated and there is a trade-off between the two factors.

A number of experiments are conducted to determine the threshold. The results show that a threshold value of 0.1 could make most of the tested minutiae templates (from 800 different fingerprint impressions) fall into the interval [0.1, 0.5], which is suitable both for storing and retrieval according to experiments in the VHC system. The filling ratio histogram of all the tested minutiae templates with binarization threshold $T h r = 0.1$ is shown in Fig. 4.

Fig. 4 The filling ratio histogram of all the tested minutiae templates with binarization threshold Thr = 0.1.

Download Full Size | PDF

After binarization, the minutiae templates with filling ratio less than 7.81% (corresponding to 120 cells out of a total of 1536 cells in a minutia cylinder) are labeled invalid and not used for recognition. The invalid templates are approximately 6.7% of all the minutiae templates (1798 / 27073). The discard of these invalid templates would not degrade the recognition rate, yet it helps to avoid the long recording time and the low intensity response problems. The last step is to interleave each template [18] and reshape it to a 640 × 480 pixels image so as to match the pixel numbers of SLM. Four typical coded minutia data pages are shown in Fig. 5.

Fig. 5 Typical coded minutia data pages adapting to the SLM, with a size of 640 × 480 pixels.

Download Full Size | PDF

3.3 Channel nonuniformity and calibration

Now the challenge of the VHC system is how to capture diffraction spots that can truly reflect the inner products. The intensity of each spot on the image captured by CCD should be proportional to the inner product between the search argument and the gallery minutia data page in the corresponding channel. The main challenges come from large variation in filling ratio and channel-to-channel variations in diffraction efficiency. We adopt two approaches to solve these problems: improving filter recording time schedule in the recording process and post-filtering calibration in the retrieval process.

Recording time schedule is very important for the channel uniformity in multi-channel VHC. In order to achieve uniform diffraction efficiency for all channels, sequential exposure schedule was proposed [19–21]. The exposure time for each data page is:

t_{M - k} = - τ_{r} \ln [\frac{(k + 1) T_{M} - k}{k T_{M} - (k - 1)}], (k = 0, 1, 2, \dots, M - 1)

where

τ_{r}

is the time constant of the crystal and

M

is the total number of stored data pages. As shown in Fig. 4, although the threshold of binarization is delicately selected, the filling ratios of the minutia data pages still vary from 7.81% to 59%. However, Eq. (9) ignores the influence of filling ratio variation and it can be regarded as the proto exposure time schedule based on the average filling ratio of all the data pages. Then each exposure time is modified by multiplying it with a factor, which is related to the ratio between the filling ratio of the corresponding data page and the average filling ratio of all the data pages. For a proto time schedule

T_{0}^{}

, the modified time schedule

T_{0}^{'}

is experimentally determined as:

T_{0}^{'} (k) = {\begin{cases} t_{k} \cdot \frac{A v e_f i l l}{f i l l (k)} if f i l l (k) \leq A v e_f i l l \\ \max {t_{\min}, t_{k} \cdot \sqrt{\frac{A v e_f i l l}{f i l l (k)}}} if f i l l (k) > A v e_f i l l \end{cases},

where

A v e_f i l l

is the average filling ratio of all the data pages,

f i l l (k)

is the filling ratio of the kth data page. This modification in recording exposure time schedule is intended to record correlation channels that are relatively uniform rather than rigorously uniform. It assures that all the data pages with a filling ratio between 7.81% and 59% could achieve a channel diffraction efficiency around that of a data page with the average filling ratio.

The residual differences are further compensated by post-filtering calibration in the retrieval process. The role of the calibration scheme is to ensure that all the residual nonuniformity of optical channels is compensated. This can be achieved by uploading a white image (all pixels ON) to obtain a base intensity, $I_{B} (k)$ , of the response of each channel. Meanwhile, since the filling ratio of the white image is 100%, the expected response of each channel, $\bar{I_{B}} (k)$ , is proportional to the square of the filling ratio of the stored minutia data page corresponding to that channel. This relationship is described as

\bar{I_{B}} (k) = A_{B} {(k)}^{2} = β \cdot {〈 T_{k}, W 〉}^{2} = β \cdot {[(R o w \cdot C o l) \cdot f i l l (T_{k})]}^{2},

where

A_{B} (k)

is the amplitude of the diffraction beam in the kth channel,

〈 T_{k}, W 〉

is the inner product of the kth minutia data page and the white page,

f i l l (T_{k})

is the filling ratio of the kth minutia data page, and

β

is the expected constant relation between inner product and diffraction intensity, representing the expected uniform diffraction efficiency of the VHC system. Define

ξ_{c a l i b r a t i o n} (k)

as

ξ_{c a l i b r a t i o n} (k) = \frac{\bar{I_{B}} (k)}{I_{B} (k)} = β \cdot {(R o w \cdot C o l)}^{2} \cdot \frac{f i l l {(T_{k})}^{2}}{I_{B} (k)},

then

ξ_{c a l i b r a t i o n} (k)

can be taken as the intensity calibrating factor for the kth channel in the post-filtering to compensate the variations of channel diffraction efficiency. When a search argument

T_{S}

is input to the system, the calibrated inner product with the kth minutia data page can be calculated as

{〈 T_{S}, T_{k} 〉}_{c a l i b r a t i o n} = \sqrt{I_{S} (k) \cdot ξ_{c a l i b r a t i o n} (k) / β} = (R o w \cdot C o l) \cdot f i l l (T_{k}) \cdot \sqrt{\frac{I_{S} (k)}{I_{B} (k)}},

where

I_{S} (k)

is the intensity of the kth correlation spot. Substitute Eq. (13) into Eq. (8),

γ (T_{k}, T_{S}) = \frac{2 \cdot f i l l (T_{k})}{f i l l (T_{k}) + f i l l (T_{S})} \cdot \sqrt{\frac{I_{S} (k)}{I_{B} (k)}} .

Thus, the calibrated expression of data page similarity is obtained from the intensity of correlation spot.

f i l l (T_{S})

requires only single calculation in the retrieval process.

I_{B} (k)

and

f i l l (T_{k})

can be stored as a table which is easy to look up during post-processing. Figure 6 illustrates the effect of calibration. The calibrated experimental similarities are more consistent with the theoretical similarities than the experimental similarities without calibration.

Fig. 6 Illustration of the similarities between one minutia in a fingerprint and all the 37 minutiae obtained both theoretically and experimentally.

Download Full Size | PDF

After minutiae similarities are calculated between all minutiae in two fingerprints, a single value is obtained to denote the overall matching score between the two fingerprints from their minutiae similarities. A simple technique called Local Similarity Sort (LSS) [14] is utilized for this purpose. The LSS technique sorts all the minutiae similarities between two fingerprints and selects the top $n_{P}$ . The overall matching score is the average of the $n_{P}$ similarities. The value of $n_{P}$ is adaptively calculated based on the numbers of minutiae, $n_{A}$ and $n_{B}$ , of two fingerprints:

n_{P} = n_{P \min} + R {Z (\min (n_{A}, n_{B}), μ_{P}, τ_{P}) \cdot (n_{P \max} - n_{P \min})},

where

Z

is defined in Eq. (2),

R (\cdot)

denotes the rounding operation,

\min (\cdot)

denotes selecting the minimum, and the parameters are experimentally set as

μ_{P} = 20

,

τ_{P} = 2 / 5

,

n_{P \min} = 4

and

n_{P \max} = 12

.

4. Experimental demonstration

In this section, experiments aimed at evaluating the robustness and feasibility of the proposed VHC based optical recognition system are carried out and the matching results are reported. The robustness to fingerprint translation, rotation and nonlinear distortion is demonstrated on a whole set of public testing database, and the feasibility for large scale database recognition is demonstrated on 270,216 fingerprints collected from a forensic department.

4.1 Optical setup and system working procedure

The experimental setup of the VHC system for testing is shown in Fig. 7. The light source is a 200mW diode-pumped solid-state laser (λ = 532nm). The SLM used to modulate the amplitude of the object beam is a transmissive twisted nematic liquid crystal display (TN-LCD), with a resolution of 640 × 480 and pixel size of 9μm × 9μm. A diffuser of 0.2° scattering angle is placed behind the SLM to suppress the side-lobes and cross talk [22]. The binarized working mode simplifies the amplitude linearization calibrating of the SLM. The expanded beam is split by the PBS into reference beam and object beam during the recording process, and an Fe:LiNbO₃ crystal is utilized for storing the template images. Then shutter ST1 is shut down during the retrieval process, and a CCD of 768 × 576 pixels, with pixel size of 14μm × 14μm and 8-bit depth,is employed for detecting the correlation spots array.

Fig. 7 Experimental setup of the volume holographic optical correlator system. PBS, polarizing beam splitter; SLM, spatial light modulator; ST1 and ST2, shutters; L1~L3, lens; R1 and R2, reflectors; λ/2, half-wave plate.

Download Full Size | PDF

The working procedure of the recognition system comprises the following four main steps (Fig. 8):

Fig. 8 Flow chart of the VHC fingerprint recognition system.

Download Full Size | PDF

• Data page construction. The minutiae of each fingerprint are digitally extracted using a commercial fingerprint recognition software, VeriFinger [23]. For each extracted minutia, an MCC data page is constructed and filtered using the modified coding method described in Section 2.2.
• Data page storing. The encoded data pages of the database are sequentially stored into the Fe:LiNbO₃ crystal by using 2-D angular multiplexing. The recorded holograms are filters for the fingerprint recognition. Hologram fixing is not used in the current demonstration since the stored data pages work well in the short-time experiments (less than 24 hours). In long-time practical applications, thermal fixing can be used to prevent the loss of fingerprint information [24].
• Fingerprint retrieval. All the minutia data pages of the query fingerprint are sequentially uploaded to the SLM as search arguments. For each search argument, an array of correlation spots can be obtained in parallel.
• Post-filtering. Minutiae similarities of all the minutiae pairs are digitally obtained from the spots array. Fingerprint matching scores are further calculated with the LSS technique described in Section 3.3.

Step 1 and step 4 require a computer and can be done in less than a second for a single search task. The database storing process of Step 2 is a time-consuming task. Fortunately it is pre-finished off-line and only needs to be done once. Compared to the digital fingerprint recognition, the real merits of the optical fingerprint recognition system is reflected in step 3 (the fingerprint retrieval step). This is because data transfer from the memorizer to the calculator is not needed and the inner products are optically calculated in parallel, which would otherwise be the main causes of time consumption and I/O traffic in a digital system if millions of fingerprint identification should be performed.

4.2 Robustness test

The system robustness to fingerprint deformation is tested on a public collection of fingerprint images proposed in FVC2002 [25]. The benchmark data is labeled DB1, which comprises 800 flat fingerprint images captured at a resolution of 500dpi, with an optical sensor, from 100 different fingers (eight impressions per finger). Remarkable deformations including translation (as large as 150 pixels), rotation (as large as ± 40 degrees) and nonlinear distortion exist in these impressions. Figure 9 shows eight impressions of a sample finger in the database.

Fig. 9 Eight different impressions of the same sample finger (#1) in FVC2002 DB1, remarkable translation, rotation and nonlinear distortion exist.

Download Full Size | PDF

Figure 10 demonstrates the minutiae matching results between two impressions of the same finger. A pair of minutiae whose similarity is above a predefined threshold is determined as a matching pair and linked. Although remarkable translation, rotation and nonlinear distortion exist and no pre-alignment is conducted before data page construction, only one pair of minutiae is wrongly linked (marked with blue line). A correct “Match” decision is successfully achieved with the matching score calculated from the matched minutiae.

Fig. 10 Minutiae matching results of two impressions of the same finger. (a) Optical correlation results between minutia data pages of fingerprint #1_1 (37 minutiae) and #1_3 (26 minutiae); (b) corresponding minutiae pairs linked according to optical correlation results.

Download Full Size | PDF

The 800 fingerprints are divided into 8 subsets. Each subset has 100 different fingerprint impressions. A total of 3250 minutia template data pages of subset #1 are stored in a common volume of the crystal as gallery templates by angular multiplexing. Then all the 800 fingerprints are searched over the whole gallery to test the system robustness. Figure 11 demonstrates part of the optical recognition results. The matching scores of the fingerprints with no deformation, small deformation and large deformation are all above the chosen threshold and genuine fingerprint pairs are correctly matched. It verifies the robustness of the proposed optical recognition system to fingerprint deformation.

Fig. 11 Part of the matching results of the fingerprints with no deformation, small deformation and large deformation when searched over the database stored in the optical system.

Download Full Size | PDF

4.3 Feasibility test

The most concerned challenge for the optical recognition system to be used in large scale database (>100,000 fingerprints) recognition task is the system feasibility. Can the system discriminate the genuine fingerprint from its similar fake fingerprints in such a large database? The system feasibility requires both the matching algorithm and the optical system can discriminate two similar fingerprints. In this work, it is tested on a database of 270,216 rolled fingerprints collected from a forensic department both digitally and optically.

The testing protocol is as follows. A sample fingerprint is randomly chosen from the database to be the query fingerprint. The similarities between it and all the fingerprints in the database are digitally calculated. The 20 fingerprints with the highest matching scores are selected as the sub-database to represent the whole database in the feasibility test. Obviously, the query fingerprint is included in the sub-database and the other 19 fingerprints are the most similar ones with the query fingerprint. The sub-database represents the fingerprints which are the most difficult ones to discriminate from the query fingerprint in the whole database. It can be inferred that the genuine sample fingerprint can be recognized from the whole database if it can be recognized from the sub-database. Then the sub-database (their minutiae templates) are encoded and stored in the crystal as filter bank and the query fingerprint is optically retrieved.

Figure 12(a) demonstrates the digitally calculated matching scores between the randomly selected 120,091th fingerprint and the whole database. The fingerprints with the highest scores are marked with triangular and selected as the sub-database. The matching score difference between the matched and unmatched fingerprints verifies the feasibility of the algorithm. Figure 12(b) shows the digital and optical matching scores of one sample fingerprint with its sub-database. The matched fingerprint can be discriminated based on the optical matching results. A total of 100 sample fingerprints are randomly selected to finish the above-described test. The recognition rate of these tests is 100%, which indicates the optical recognition system can discriminate genuine from fake fingerprints even when the database is quite large.

Fig. 12 (a) Matching scores of the 120,091th fingerprint with all the 270,216 fingerprints in the database. (b) Comparison of digitally and optically calculated matching scores between the sample fingerprint and its corresponding most similar 20 fingerprints.

Download Full Size | PDF

The discriminating power comes from the different template data pages and different multi-page combinations. Without fingerprint encoding, the real discriminative information may be submerged by the redundant correlation noise if all the information is retained, especially when fingerprint distortion occurs. The feature extraction and binarization operation cause some information lost compared with using the fingerprint image directly. However, these operations can retain most of the discriminative information while removing the redundant correlation noise between fingerprints, provided that the encoding parameters such as the MCC fixed-radius R and the binarization threshold are optimized. The experiments show that the recognition accuracy of the optical system after the multi-page fingerprint encoding is improved on the whole.

The proposed method could be more efficiently implemented in a holographic disc based configuration with shift multiplexing. Up to ten million data pages can be stored in a regular size photopolymer holographic disc with a multiplexing displacement of approximately 20μm. The stored database can be quickly retrieved with the high-speed rotation of the holographic disc.

5. Conclusions

We integrate a local minutiae structure coding method into the VHC based fingerprint recognition system in order to promote the robustness of the optical system against translation, rotation and nonlinear distortion. The hybrid system decomposes each fingerprint into multiple data pages by using a modified local minutiae structure coding method adapted for the optical data channel . The robustness of the system comes from the invariant properties of local structure coding and no pre-alignment of fingerprint is needed. The modified similarity calculation method, binarization and post-filtering calibration are combined to suppress the similarity calculating error. Experiments conducted on a public database FVC2002 DB1 including 800 deformed flat fingerprints and a forensic database including 270,216 rolled fingerprints verify the robustness and feasibility of the optical system.

Acknowledgments

This work is supported by National Natural Science Foundation of China (61177001, 61275013), Tsinghua University Initiative Scientific Research Program, and the Ministry of Public Security of China Fellow Fund.

References and links

1. D. Maltoni, D. Maio, A. K. Jain, and S. Prabhakar, Handbook of Fingerprint Recognition (Springer, 2009), Chap. 4, pp. 167–233.

2. P. A. Mitkas and G. W. Burr, “Volume holographic optical correlators,” in Holographic Data Storage, H. J. Coufal, D. Psaltis, and G. T. Sicebox, eds. (Springer-Verlag, 2000), pp. 429–446.

3. B. V. K. Vijaya Kumar, A. Mahalanobis, and R. D. Juday, Correlation Pattern Recognition (Cambridge University, 2005), Chap. 8, pp. 295–356.

4. K. H. Fielding, J. L. Horner, and C. K. Makekau, “Optical fingerprint identification by binary joint transform correlation,” Opt. Eng. 30(12), 1958–1961 (1991). [CrossRef]

5. T. J. Grycewicz, “Techniques to improve binary joint transform correlator performance for fingerprint recognition,” Opt. Eng. 38(1), 114–119 (1999). [CrossRef]

6. A. Bal, A. M. El-Saba, and M. S. Alam, “Improved fingerprint identification with supervised filtering enhancement,” Appl. Opt. 44(5), 647–654 (2005). [CrossRef] [PubMed]

7. Y. Yan, G. Huang, W. Feng, G. Jin, and M. Wu, “Multichannel wavelet correlators for fingerprint identification by the use of associative storage in a photorefractive material,” Proc. SPIE 3458, 259–266 (1998). [CrossRef]

8. S. H. Lee, S. Y. Yi, and E. S. Kim, “Fingerprint identification by use of a volume holographic optical correlator,” Proc. SPIE 3715, 321–325 (1999). [CrossRef]

9. E. Watanabe, A. Naito, and K. Kodate, “Ultrahigh-speed compact optical correlation system using holographic disc,” Proc. SPIE 7442, 74420X (2009).

10. B. V. K. Vijaya Kumar, M. Savvides, C. Xie, K. Venkataramani, J. Thornton, and A. Mahalanobis, “Biometric verification with correlation filters,” Appl. Opt. 43(2), 391–402 (2004). [CrossRef] [PubMed]

11. X. Jiang and W. Y. Yau, “Fingerprint minutiae matching based on the local and global structures,” in Proceedings of the 15th International Conference on Pattern Recognition2 (Institute of Electrical and Electronics Engineers, New York, 2000), pp. 1038–1041.

12. N. K. Ratha, R. M. Bolle, V. D. Pandit, and V. Vaish, “Robust fingerprint authentication using local structural similarity,” in Proceedings of the 5th IEEE Workshop on Applications of Computer Vision (Institute of Electrical and Electronics Engineers, New York, 2000), pp. 29–34. [CrossRef]

13. J. Feng, “Combining minutiae descriptors for fingerprint matching,” Pattern Recognit. 41(1), 342–352 (2008). [CrossRef]

14. A. A. Paulino, J. Feng, and A. K. Jain, “Latent fingerprint matching using descriptor-based hough transform,” IEEE Trans. Inf. Foren. Sec. 8(1), 31–45 (2013). [CrossRef]

15. J. Dai, J. Feng, and J. Zhou, “Robust and efficient ridge-based palmprint matching,” IEEE Trans. Pattern Anal. Mach. Intell. 34(8), 1618–1632 (2012). [CrossRef] [PubMed]

16. R. Cappelli, M. Ferrara, and D. Maltoni, “Minutia Cylinder-Code: a new representation and matching technique for fingerprint recognition,” IEEE Trans. Pattern Anal. Mach. Intell. 32(12), 2128–2141 (2010). [CrossRef] [PubMed]

17. G. W. Burr, S. Kobras, H. Hanssen, and H. Coufal, “Content-addressable data storage by use of volume holograms,” Appl. Opt. 38(32), 6779–6784 (1999). [CrossRef] [PubMed]

18. K. Ni, Z. Qu, L. Cao, P. Su, Q. He, and G. Jin, “Improving accuracy of multichannel volume holographic correlators by using a two-dimensional interleaving method,” Opt. Lett. 32(20), 2972–2974 (2007). [CrossRef] [PubMed]

19. F. H. Mok, M. C. Tackitt, and H. M. Stoll, “Storage of 500 high-resolution holograms in a LiNbO₃ crystal,” Opt. Lett. 16(8), 605–607 (1991). [CrossRef] [PubMed]

20. A. Adibi, K. Buse, and D. Psaltis, “Multiplexing holograms in LiNbO₃:Fe:Mn crystals,” Opt. Lett. 24(10), 652–654 (1999). [CrossRef] [PubMed]

21. K. Curtis, K. Anderson, and M. R. Ayres, “M/# requirements for holographic data storage,” in Proceedings of the Optical Data Storage Topical Meeting, IEEE, 9–11 (2006). [CrossRef]

22. L. Cao, Q. He, C. Ouyang, Y. Liao, and G. Jin, “Improvement to human-face recognition in a volume holographic correlator by use of speckle modulation,” Appl. Opt. 44(4), 538–545 (2005). [CrossRef] [PubMed]

23. http://www.neurotechnology.com/vf_sdk.html

24. X. An, D. Psaltis, and G. W. Burr, “Thermal fixing of 10,000 holograms in LiNbO₃:Fe,” Appl. Opt. 38(2), 386–393 (1999). [CrossRef] [PubMed]

25. FVC2002, http://bias.csr.unibo.it/fvc2002/default.asp

Optical fingerprint recognition based on local minutiae structure coding

Abstract

1. Introduction

2. Template data page construction with local minutiae structure coding

2.1 Algorithm selection criteria for optical recognition

2.2 Template data page construction with Minutiae Cylinder-Code (MCC)

3. Optical considerations

3.1 Similarity between two minutiae using inner product

3.2 Template data page binarization and filling ratio

3.3 Channel nonuniformity and calibration

4. Experimental demonstration

4.1 Optical setup and system working procedure

4.2 Robustness test

4.3 Feasibility test

5. Conclusions

Acknowledgments

References and links

Cited By

Figures (12)

Equations (15)

Optics Express