Method of identification of fingerprints and device that makes use of it (Machine-translation by Goo

专利摘要:
Method of identification of fingerprints and device that makes use of it. A method and a device are described in this document that allow to generate a vector of characteristics of a human fingerprint by means of a series of processes carried out from the image of said fingerprint. With these vectors, it is possible to perform a classification, indexing or identification of fingerprints (and, therefore, of individuals), as well as protecting secrets or generating cryptographic keys from fingerprints. Most of the identification methods reported use vectors of characteristics and techniques to extract and compare them that are not suitable for electronic devices with reduced computation and storage resources, such as an fpga or an integrated circuit of specific applications. On the contrary, the method proposed in this invention is adequate, offering good benefits in terms of extraction times of vectors (placing them below the millisecond for standard footprint sizes), pairing and ordering times (negligible values of a few nanoseconds per user) and memory requirements (just over one hundred bytes per fingerprint). The main uses of the invention are located in automatic small, portable, cheap and/or secure fingerprint identification systems where the user is present and wants to identify himself. (Machine-translation by Google Translate, not legally binding)
公开号:ES2556276A1
申请号:ES201300721
申请日:2013-07-31
公开日:2016-01-14
发明作者:Maria Rosario ARJONA LÓPEZ；María Iluminada BATURONE CASTILLO
申请人:Consejo Superior de Investigaciones Cientificas CSIC；Universidad de Sevilla；
IPC主号:

专利说明:

OBJECT OF THE INVENTION
The present invention is framed in the field of biometric identification systems and methods.
The object of the invention consists of a method and a device that allow generating a vector of characteristics of a human fingerprint by means of a series of
15 processes carried out based on the image of said footprint. With these vectors, a classification, indexing or determination of fingerprint identity (and, therefore, of individuals) can be performed, as well as protecting secrets or generating cryptographic keys from fingerprints.
BACKGROUND OF THE INVENTION
The use of fingerprints as a biometric feature is widespread for applications of identification of individuals, access control, ele., Because of their high discrimination and because users normally accept the fact of introducing the fingerprint in a capture device (provides ease of use being a non-intrusive technique) It is one of the biometric characteristics that have been most successfully applied in forensic and police activity. more recently. in access control systems Automatic fingerprint identification systems (AFIS. 30 '"Automatic Fingerprint Identified Systems") require comparing an input fingerprint with the fingerprints stored in the system base. In these applications, individuals are previously registered on a basis using the characteristics of their fingerprints. Subsequently, when you want to identify an individual, the characteristics of their fingerprints are re-extracted and compared with the
35 characteristics stored in the footprint base.
The efficient realization of identification systems that are currently challenging
Application No. F. Efccllva
 10/30/2015
use bases with a high number of fingerprints and provide immediate response times. at the same time that accuracy in the identification The time necessary for the identification of an individual (tidentification) can be expressed as:
5 tidentification = textracción + temparejamiento * N + tdecisión
where textracción is the time invested in the extraction of characteristics of the footprint, temparejamiento is the time taken to compare the characteristics extracted from the entrance fingerprint with each of the N characteristics stored
10 at the base, and the decision is the time taken to decide which of the N registered individuals is the candidate chosen as the holder of the entry fingerprint, in the case of an identification application, or the time taken to generate a reduced list with the M individuals candidates to own it (with M much smaller than N), in the case of an indexing application.
The value of (extraction is much greater than that of tethering because the feature extraction process is much more complex than that of pairing. For example, the MINOTCT extraction algorithm developed by the NIST ("National Institute of Standards and Technology") it is an order of magnitude slower on the same PC platform with an Intel Core i7 processor than the NIST BOZORTH98 pairing algorithm, for example, more than 200 ms on average for the first and less than 20 ms for the second (both MINDTCT as BOZORTH98 available in NIST Biometric Image Software (NBIS), http: //WW'vV.nist.govlltlllad/ig/nbis.cfm) However, even if the (pairing is
25 minor, being multiplied by N, the base search may be too slow for real-time applications
To reduce the number of comparisons of the input footprint with respect to the footprints stored in the system base, so-called methods of
30 "exclusive classification" in which the footprints are distributed in pre-established disjoint groups, so that the entry fingerprint is classified in one of those groups and is only compared with the registered footprints of that group. The commonly used scheme follows the proposals of Galton and Henry (E.R Henry, "Classification and Uses of Finger PrintsR London: Routledge, 1900). Which distinguish five groups of fingerprints,
35 r arch ", uwhorl", Utended arch ", ~ Ieft loop ·, and" right loop "). The problem is that most of the fingerprints belong only to three groups (" right loop "," Ieft loop ·, and "whorl"), which does not result in a large reduction in the number of comparisons in a F.OEPM
11/02/2015
extensive fingerprint base (R. Cappelli. A. Lumini, D. Maio, and O Maltoni, ~ Fingerprint classification by directlonal ¡mage partitioning "; IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 21, pages 402-421, 1999 Another problem with the exclusive classification methods is that determining which class a fingerprint belongs to is in many cases an ambiguous operation, these inconveniences have led to the appearance of the so-called "continuous classification" methods, which assign a vector of numerical characteristics to each fingerprint. In this way, in the indexing phase rindexing "), an indexed table or base of fingerprints is created. In the recovery phase rretrieving "), before an entry fingerprint with a vector of
10 representative characteristics, the similarity between the input vector and those stored by a distance measure is calculated, so that a reduced list can be obtained with the most similar M candidates (with M much smaller than N). Other identification methods can then be applied to determine the correct individual among those selected
A fingerprint is an image where gray-scale ridges and valleys are distinguished. Direct comparison of grayscale prints does not offer good results, as well as being an expensive technique in terms of computation and storage. It is better to process the image to obtain distinctive and compact features. We usually talk about characteristics of 3 levels, according to the level of detail with which the fingerprint is analyzed. Level 1 characteristics are obtained from a global analysis of the footprint. An example is the directional image (field or map) (also called the image, field or orientation map), which contains the local orientations of the crests with respect to a reference axis. Another example is 25 singular points, which are points of the footprint where (orientations or cores) converge (or deltas) the orientations and around which most of the distinctive information of a footprint The characteristics of level 2 are obtained from a local analysis and in more detail of the footprint An example is the traditional minutiae, which can be terminations (places where the ridges
30 finish) and forks (places where the ridges separate into two others). The characteristics of level 3 (such as pores or incipient ridges) are obtained after a very detailed analysis of the footprint that requires an acquisition of the same with very good quality. Feature extraction is all the more distinctive and at the same time more expensive the higher the level.
So far, techniques that employ characteristics of
level 3. If several techniques have been reported that employ level 2 characteristics.
Previous step to all these techniques is to extract the minutiae. Extraction processes
of minutiae are complex because the image of the footprint has to be
prepared to locate the minimum details (the minutiae are not very robust to the
possible noise that the image of a fingerprint may present). For example, the technique of
s minutiae extraction more common requires that the image of the huell be improved,
segmented, binarized, and thinned A widely known example of
Minute detector algorithm is the MINOTCT discussed above. The
Minute-based fingerprint comparison is relatively slow and, therefore, 0 therefore not
It is suitable for indexing (for example, using algorithms such as BOZORTH98
10 commented above). Indexing techniques that use features of
Level 2 apply a post-processing on the minutiae extracted. So for example, in the
"Method for searching fingerprint database based on quantum algorithm" patent,
CN102495886 {Al, a set of minutiae is chosen and expressed in coordinates
polar with respect to some reference points that are also chosen
fifteen conveniently in the patent "Fast fingerprint identification and verification by
minutiae pair indexing ", W02008135521 (A2), indexing by pairs is used
of minutiae In the document "Methods and systems for automated fingerprint
recognition ", W02008098357 (A1), minute patterns are associated with fingerprints.
method and system proposed in the documentary "Progressive fingerprint matching system
twenty and method ", US2004062426 (A1) is based on pairing fingerprints by
minutiae. In the document "Fingerprint verificationn, US2005058325 (A1),
they sample the minutiae and are ordered in subsets. In the document "System and
method for matching (fingerprint) images an aligned string-based representation ~,
US6185318 (81), minutiae are used as reference points. At
25 "Methods and related apparatus for fingerprint indexing and searching" document,
US6181807 (81), minutiae are extracted from the fingerprints and compared in the process of
search In the document "Vector based topological fíngerprint matching",
W09532482 (A 1 "minutiae positions are used and assigned a
index number Other known techniques are those that use miniature triplets.
30 In the document "Fingerprint identification method based on triangulation and LOO
technology ~, CN101620677 (A), a triangulation technology is used to
extract vectors of global and local characteristics. Previously it has been proposed
use all possible triples that can be formed with every minute. Other authors
they propose to apply the Oelaunay triangulation of order 1 to the coordinates of the
35 minutiae to assign a unique topological structure to each footprint. In other techniques
known in the art Delaunay triangles of greater than 1 arden are used for
extract more geometric information. Another technique that uses level 2 features is
the MCC representation ('"Minutia Cylinder-Code"). which assigns to each minutia a local structure that encodes the probability of finding minutiae around it, with an orientation difference similar to a given value
5 Techniques that use level 1 features offer competitive features with much lower computational cost. In fact, directional image extraction is a necessary step in most minutiae and fingerprint comparison algorithms, so it could be said that its cost is zero.
techniques differ from each other in how they extract representative characteristics
10 and compact directional image. For example, in known techniques a fingerprint orientation model based on two-dimensional Fourier expansions is used to adapt to the intrinsic periodicity of the orientations. In other solutions a set of polar complex moments (PCMs) are used to extract characteristics of the directional image invariant to rotations of the
15 footprint
In the document ~ New fingerprint database retrieval method ", CN102368242 (A),
employ singular points, information on the relationship between singular points and
orthogonal invariant moments. In the document ~ Method for rapidly calculating
20 fingerprínt similarity ", CN1 01996318 (A), similar topological units are sought between each pair of fingerprints to be compared, expanded to obtain similar major topological units and grouped together to obtain a measure of global similarity.
Most solutions for fingerprint identification and indexing are
25 software implementations that involve algorithms of high computational cost in terms of time and resources. The cost is high, not only for the extraction of the characteristics, but even for the pairing algorithm that calculates the similarity between the characteristics of the input and stored footprint.
30 Indexing solutions are usually evaluated by showing, for a given penetration rate (an average percentage of the fingerprint base to be analyzed), the error ratio (percentage of input fingerprints whose registration is not recovered from the list with higher similarities with that penetration). This is the penetration ratio defined as the percentage of candidates considered
35 to see if among them is the true holder of the entry fingerprint (M / N). Another measure normally analyzed to assess the goodness of an indexing technique is the average rate in an incremental search scenario, which is calculated as the average rate to be carried out when you do not want to commit errors in the recovery of the holder of the entry fingerprint.The average times that are spent in the search are not usually reported, nor are the memory requirements in the works that are reported.
5 times shown are of realizations on PCs: 67 ms on an Intel Pentium 4 at 2.26 GHz and 1.6 ms, 14 ms 016 ms (according to the technique) on an Inlel Core 2 Quad at 2.66 GHz on 2000 fingerprints of the NIST base 084.
In solutions for identification, once the feature vector has been generated and
10 compared to stored vectors, an ordered list of fingerprints is not generated but a pairing threshold is established to accept or reject if an individual is who he claims to be. In this case, what is measured are reasons for false rejection (FRR, uFalse Rejection Rate ") and false acceptance (FAR," instead of penetration rates, as in the case of inde) ( ation.
15 called FNMR (MFalse Non-Match Rate ") or FMR r False Match Rate6
),
respectively.
The context of application of these solutions is usually the coroner and police officer, in which the traces of an individual have been acquired without their cooperation (for example, because it involves identifying a deceased or a criminal). It is called an "offline" context, so the captures can be of poor quality and the algorithms can be made on PCs without especially restrictive speed and memory consumption requirements. There are fingerprint bases, such as those of the "Fingerprint Verification Competition" (FVC), which are built with many
25 poor quality and poorly acquired catches to prove the goodness of complex identification and indexing techniques.
A different application context is what is called "online", in which the user of a recognition system cooperates with the system because he wants to authenticate (for example, in an access control system). In this case, the captures are of much better quality and, even, you can interact with the user to enter your fingerprint well. In this line, a solution for estimating the quality of a fingerprint based on an algorithm for the extraction of singular points that satisfies the restrictions in terms of resources, response time and recognition results imposed for an acquisition application is known Smart in a recessed hardware device. In the document "Melhod for authenticating an individual by use of fingerprint data-, US7136514 (81),
it is taken into account that the individual who enters his fingerprint through a sensor
Scanning can sweep your finger in different directions relative to the axis of the sensor. At
document ~ Fingerprint matching ", GB2320352 (A), quality indices are used in
the extraction of the feature vector and then use them in the calculation of
5 pairing between tracks.
The speed requirements if they are restrictive in an application context in
line, because the operation must be in real time. the ease of use of the system and
Its price can also be important requirements in this context. The user
10 you can comfortably use a small, light and cheap electronic device, for
example, a card or token with reduced resources. the resources available to you
a smart card or an OSP for embedded devices are much smaller
than those of a pe: 50 or 100 MHz CPUs and available memories (ROM, EEPROM Y
RAM) of a few tens of KBytes, at best.
' 5
Times increase greatly if the platform where the algorithms are implemented
It has few resources. For example, the MINOTCT algorithm adapted and executed on
an embedded lEON2 processor invests in its execution almost three orders of
magnitude more than on the PC platform (about 100 s, as reported). For this
twenty reason, feature extraction is usually done on a p and y platform
the solutions that use cards or DSPs for online recognition by fingerprint
only implement the matching algorithm between the features
stored and those that arrive from abroad. In addition, stored features
they are usually those of an individual only (pairing 1 versus 1 instead of 1 versus
25 N). This solution is often called ~ match on card. "Even so, they have been proposed
solutions in which the pairing algorithm software is redesigned, it
employs fixed point arithmetic and extends the instruction set of the
embedded processor to accelerate execution. the "match on card" algorithms
They have recently been studied in the MINEX II campaign organized by NIST. The
30 Results obtained show that San Menas precise that those that are executed
on a PC platform Another option for implementation in embedded systems
is to use FPGAs ("Field Programmable Gate Arrays"). In FPGAs you can
Implement hardware coprocessors that accelerate the execution of algorithms.
Thus, for example, there is a solution that proposes the direct correlation of images
35 grayscale, using a Virtex 4 FPGA. For indexing applications of
footprints, there is a technique that creates a footprint base whose indexes are based on the
minutiae extraction, in which the fingerprint base and the search on the basis are
implemented in PC I boards based on FPGAs while minutiae extraction is performed on the PC to which the boards are connected.
In terms of security, it is very interesting that the entire process of extraction of 5 characteristics, its storage and pairing can be performed on the same device, what is called "authentication on card ~, because thus the distinctive characteristics of individuals are circumscribed within a much smaller perimeter, which, therefore, is easier to control and defend.In this line, there are solutions where a miniature extraction algorithm is implemented in an FPGA 10 Spartan 3 and solutions where a system of implementation is implemented recognition based on the location of singular points on a Celoxica RC203E board equipped with a Virtex 11 FPGA. Instead of simplifying the algorithms to be implemented, the solution analyzed in a second option is to use FPGAs that are reconfigured according to the task to be performed (extraction of the directional image, improvement and segmentation of the
15 image of the fingerprint, binarization, smoothing, thinning, minutiae detection, alignment and pairing). This idea of ~ authentication on card ~ appears in several patents. Among them, we can cite "Siometric identity verification system and method ~, US 20080223925 A 1 and, among the most recent," Smart card system with
ergonomic fingerprint sensor and method of using ", US 8276816 82
20 Another major problem of fingerprint based identification systems, no longer related to implementation but with its own nature, is that of the lack of diversity in obtaining distinctive features. For example, a user has at most 10 fingers on their hands. If it is discovered that an imposter will
25 seizes the characteristics of one of his fingers, the individual has already lost 10% of its possible characteristics. If the imposter seizes the characteristics of the 10 fingers, the individual can no longer register in any system. This problem is also called as the low revocability of the system, that is, it is difficult to generate new features when others have been discovered or compromised.
30 To avoid that the characteristics can be compromised, systems have been proposed that transform them by means of non-invertible functions, such as hash functions, so that recovering the original characteristics from the transformed ones is practically unfeasible from the computational point of view. In these systems, the measure of similarity or matching between the characteristics of
35 entry and previously registered is done in the transformed space. That is why we must analyze very well how the transformation affects the benefits of
resulting system (for example, in terms of false acceptance rates and false rejection). This solution does not solve the problem of diversity because the transformation of characteristics does not increase the number of possible characteristics. For this, a random number (known as "salt" can be used in cryptographic techniques) that is combined with the original characteristics, so that its transformation is different even if it uses [the same biometric information The ~ saltN acts as a password that the individual must enter into the system,
In addition to your fingerprint. The disadvantage of this solution is that the password and the transformed characteristics should not be made public to increase the security of the system.
Another scheme that has the advantage of not requiring secure storage of information is that of the so-called rbiometric criptosystems') biometric encryption systems. They are based on combining the original biometric characteristics (without any transfomación) with additional information such that the resulting data, known as "helper data", can be public. The two techniques most used in encryption systems have been called Fuzzy Commitment and Fuzzy Vault. Fuzzy Comm itment is a more basic and simpler scheme than Fuzzy Vault. In return, Fuzzy Commitment requires that the feature vector be a binary, ordered and fixed-length vector. Fuzzy
20 Commitment is implemented in the following two phases:
- Registration phase: the biometric characteristic is combined (usually by an XOR operation, in the case of binary characteristics) with a code word generated by applying a correct error code to a random number, password or password (password ~ to). The result is information from
25 help that does not need to be stored safely.
- Secret verification or recovery phase: the new biometric characteristic, slightly different from the one obtained during registration (which is usual). It is combined with the help information, which is public, to retrieve the code word (applying an error correction code). From the information retrieved in
30 this phase, a cryptographic key can also be generated.
Today, most of the reported fingerprint identification methods (as well as the encryption systems based on them) employ feature vectors and techniques to e: x bring them and compare them that are not suitable for electronic devices with computing resources and reduced storage That is why valid fingerprint identification solutions are necessary for
encryption systems that, while maintaining good recognition results, are suitable for low power consumption electronic devices, with limited computing capacity and do not require a powerful and bulky hardware that uses large resources.
DESCRIPTION OF THE INVENTION
A method and a device are proposed to implement the method that allow,
10 through a series of processes and from an image captured from a fingerprint, generate a feature vector based on level 1 features, specifically in the segmentation of the directional image into homogeneous regions. The method makes it possible to obtain a fixed-length bit string from the fingerprint preferably captured online by means of a fingerprint sensor of which
15 used in automatic identification systems (optical, capacitive, etc.).
In a possible embodiment the method described herein can be adapted and used in classification applications in which the footprints of the individuals are distributed in more or less disjoint groups, according to the clustering algorithm that is applied on the feature vectors. The method can also be used in indexing and identification / authentication applications, in which individuals are registered by means of feature vectors that are generated in an indexing or registration phase. In the recovery or verification phase, given an entry fingerprint, an ordered list of candidate individuals to own that footprint is generated
25 (in the indexed) or the best candidate is identified (in identification applications). In the case of authentication, a single individual is registered and in the verification phase the similarity between the generated and stored vector is measured. If it exceeds a threshold, the individual is authenticated. It is not authenticated in another case.
30 The method described here can be used in multi-biometrics. Given several samples of fingerprints captured from different fingers of the same individual, the vectors obtained from each finger are concatenated to obtain a digital identification vector of the individual. And also, given several samples of fingerprints captured from the same finger, the vectors obtained from each sample are concatenated to obtain a vector
35 fingerprint identification.
The method can be used in double-factor identification (and authentication) applications because the vector it generates can easily be combined with vectors derived from passwords or passwords.
5 The method can be used in so-called template protection schemes nemplate ~). In particular. It is very suitable for encryption systems based on the Fuzzy Commitment technique, because the generated vector is binary, ordered and of fixed length. In these schemes, the method of the invention offers the advantages of noreversibility of the transformed vectors and diversification of the vectors
10 generated, maintaining accuracy in identification (and authentication).
The method proposed in this invention can be implemented in a low cost electronic device (with reduced computing and memory resources). as for example, an FPGA or an integrated circuit of specific applications, offering 15 good features in terms of feature extraction times (placing those below the millisecond for standard fingerprint tables), matching times and candidate ordering (negligible values of few nanoseconds per candidate) as well as memory requirements (just over 100 bytes per footprint). Ace; you can get a very safe solution because all the
The biometric information of individuals may be confined within the same electronic device and not leave it.
The object of the invention is based on an online application context, that is, the user of the device collaborates to identify, unlike other contexts of fingerprint identification application, such as forensic or police, in which the user does not collaborate (because he is deceased or does not want to be identified). In an application context in which the individual wants to register and identify. The features are extracted with quality. In any case, the device that implements the invention allows to evaluate the quality of the process and the online interaction
30 with the user to avoid defective fingerprint captures.
The identification method is based on generating a vector of fingerprint characteristics for identification from a first grayscale image that contains ridges and valleys of the fingerprint, for this purpose they are carried out
35 the following steps: a) determine for each pixel of the first image, PII (where ij refer to the row and column of the pixel in the image), the gradient of the
image intensity (of gray levels) in that pixel, b) determine the direction of the gradient by an angle a ~ with respect to a reference axis,
e) divide the range of possible values of angles, a¡¡ :, in G sub · intervals (9 "..., 9G) that do not overlap and whose union gives Play the full range of possible values, encompassing each sub- 9k angle range from a value
Or k_l until ak,
d) label each Qk sub-interval with a label, Ck,
e) associate, for each pixel Pij of the first image, the label corresponding to the sub-interval to which the angle aij corresponding to that pixel belongs,
f) generate a second image from the first image, where in said second image each pixel is associated with a label, g) perform a smoothing process to the second image to obtain areas comprising pixels with the same labels, h) locate at least one convex core point in the second image
smoothed, i) define a window centered on the convex core point, j) sample pixels included in the window, and k) generate the vector from the labels of the pixels sampled in the
previous step, in an orderly manner.
The determination of the sub-interval to which the ex angle of the gradient belongs in a pixel is determined from the calculation of the horizontal gradient (Gl and the vertical gradient (Gy) of the intensity of the image (of the gray levels) in that pixel
The determination of the sub-interval gk = [a..¡, Ok) that is associated with the pixel p¡j comprises:
• determine the sign of G.
• determine the sign of Gy
• determine that:
• o belongs to the first quadrant of angles between 0 ° and 90 °, when G. and G ~ have the same sign, and, within this first quadrant, two situations are distinguished according to the range of angles covered by each sub-interval a evaluate:
- if the sub-interval being evaluated, [Uk. , (l ~), is fully included in the first quadrant, because both ak. ' as Uk are less than or equal to 900, then
- if the sub-interval being evaluated, [ak.1, ak), is partially included in the first quadrant, because Clk_ 'is less than 90 "but ak is greater than 90", then
10 •a belongs to the second quadrant between 900 and 1800,
when Gx And Gythey havedifferent signs, and, withinEastsecond
quadrant two situations are distinguished according to the range of angles that
It covers each sub-interval to evaluate:
fifteen -if the sub-interval being evaluated, (ak.1, a k),is fully included in the
second quadrant, because so much ak., like akThey aregreater or equal
that 90 ", then
. if the sub-interval that is evaluated, Ia ....,. ak}, is partially included in the second quadrant, because ak is greater than 90 "but o.k., is less than 90", then
25 The smoothing process calculates for each pixel P'j of the second image preferably which of the labels is the one that most often appears in a window of size S x S pixels of the second image, window centered on the pixel to be smoothed, where S can be factored as S = 51 x 52 x. x sn, process that includes:
• start with s1 x s1 pixel size windows and apply the smoothing to your s1 x s1 tags,
• continue with windows of (s 1 x s2) x (s 1 x s2) pixels and apply smoothing over s2 x 52 labels softened previously in the previous step,
• proceed like this until you reach the window tamar (51 x 52 x .. x sn) x (s1 35 x s2 x. x sn) pixels and apply smoothing over sn x sn tags
previously softened in the previous step
The determination of the convex core can perform the following steps:
• make another division of the range of possible angle values, ct.j, into 4 sub
5 intervals (9 , ..., g'4) that do not overlap and whose union gives rise to the full range of possible values,
• label each g'k sub-interval with a label, C'k,
• convert the second smoothed image, in which each of the pixels is associated with one of G labels (Cl. .., cG), where-preferably G> 4, in
10 a smoothed tetra-directional image, in which each of the pixels is associated with one of four labels (C'1 • ..., C'4) and the conversion includes:
• change each Ck tag associated with the subklot of angles Qk by
that label C'k associated with the sub-interval of angles g 'k that verifies that the intersection gk n g'k is the largest, and
• determine the convex core as the point where three of four homogeneous regions of the smoothed tetra-directional image are touched, which are regions that encompass most of the crests with convex curvature
Also, the invention described herein is also directed as another object thereof to a device for generating a vector of features of a fingerprint from an image thereof, a device that is associated with image capture means of the footprint and characterized because it comprises:
• a block of assignment of labels destined to assign to each pixel of the
25 image one of G possible labels, which allows to generate the second image,
• a smoothing block intended to perform a smoothing process to the second image to obtain areas comprising pixels with the same labels,
30 • a block for determining the convex core in the fingerprint, designed to locate at least one convex core point in the second smoothed image,
• a window block designed to define a window centered on the convex core point, sample pixels included in the window, obtain the label of each sampled pixel and generate the vector from the
35 labels obtained in an orderly manner
The label assignment block comprises:
• a filter preferably of Sobel 3x3 with convolution masks with integer values and powers of 2 to calculate the horizontal gradients (G ~) 5 and the vertical gradients (Gy) of the intensity of the image (of the gray levels) in the pixels, and
• logical operators type OR and ANO, relational operators and operations of absolute value and multiplication by constant values
10 The smoothing block is adapted to process the second image by sweeping its pixels is one at a time and providing the pixels of the smoothed image also one at a time, where the smoothing block defines a window of size SxS, where S is You can factor as S = s1 x s2 x ... x sn, and where the smoothing block comprises a series of registers and n sub-blocks with a series-parallel hybrid architecture of which:
• a first sub-block with window size 51 x s1 is adapted to apply a parallel smoothing function on s1 x s1 pixel tags that have been stored in the corresponding registers, sub-block
whose resulting label is stored one after another in a series of 20 records;
• a second sub-block with window size (s1 x 52) x (51 x 52) is adapted to apply a parallel smoothing function on s2 x s2 labels previously softened by the first sub-block and available in the corresponding registers that store the output of the first sub-block, sub
25 block whose resulting label is stored one after the other in a series of registers;
• so up to a nth sub-block with window size (s1 x s2 x ... x sn) X (s1 x s2 xx sn), which again applies the smoothing function in parallel on sn x sn labels previously smoothed by the previous sub-block and
30 available in the corresponding registers that store the output of the previous sub-block, sub-block whose output provides the pixel tag in the smoothed image.
The convex core determination block comprises: • a sub-block of the convex core determination block adapted to convert the second smoothed image into a tetra-directional image,
preferably truncating the 1092G bits of each pixel to 2 bits encoding the labels of the tetra-directional image, and
• a sub-block of the convex core determination block adapted to locate at least one convex core point.
5Additionally and in various embodiments, you can have:
• A memory block intended to store the captured image of the fingerprint
• An image improvement block designed to process it, improving its quality.
• A block of fusion of information adapted to: acquire a password or password, apply a non-invertible function (hash) to said password or password and combine the result of the previous step with the vector of characteristics of the fingerprint.
Since the finger, the fingerprint, are not always oriented in the same way with respect to the sensor, the option of including an image orientation block intended to rotate or rotate it to a certain position in case 20 of that the fingerprint captured by the image capture means of the fingerprint is not in a certain orientation, a block that preferably rotates through fixed angles to apply linear transformations between original pixels (x "Yi) and pixels of the rotated images (XI, YI) with the fixed linear transformation parameters for each turn, and block that in a possible embodiment can be programmable in
25 the number of rotations as well as the parameters associated with the rotations.
In the case of template protection ("template ~), the device that implements the method may additionally comprise:
30 • An acquisition block adapted to acquire a random number, password or password and apply an error correction code encoder to generate a secret,
• A block with XOR operators adapted to calculate and store some
public support data from the vector of footprint characteristics and the secret and
• A decoder block of an error correction code adapted to
Application No. F.Erccliva F.OEPM
 10/30/2015 11/02/2015
17
retrieve a secret from an extraction of the feature vector and stored help data from the fingerprint associated with the secret
All the blocks described above can be included in a device
5 low-cost electronic that, in addition, allows to interact with the user based on the evaluation of the quality with indicators extracted mainly from the smoothing operation. The device can have LEOs to communicate to the user with a simple color code that the fingerprint has been acquired with good quality and / or the finger has been well placed on the sensor (for example, a LEO
10 lit green indicates correct process and red indicates error). The device can communicate more extensive information about the capture to the user (visually through a small "LCO panel" or audibly through a simple voice synthesizer), such as "the finger has been placed too low on the sensor" Since the application context is online and the user is present.
15 information can result in the user inserting his finger back into the sensor (higher or lower, with more or less pressure, etc., depending on the information received from the device). In this case, if the device includes a LeO panel and / or a voice synthesizer for online interaction with the user when capturing the fingerprint, they can also be used to communicate outside them
20 candidate (s) selected in the recognition process.
The online interaction with the user provides quality biometric samples, which, therefore, reduce the error rates obtained for low penetration rates at the base of fingerprints and the average penetration rate in a scenario
25 incremental search scenario in an indexing application. As a consequence, the average number of candidates among whom the entry fingerprint holder always appears is a small percentage of all [based. Also, through interaction with the user, the reasons for false rejection (FRR) and false acceptance (FAR) for an identification application are improved.
DESCRIPTION OF THE DRAWINGS
In order to complement the description that is being made and in order to help a better understanding of the characteristics of the invention, according to a preferred example of practical realization thereof, "ia" will be incorporated as an integral part of
said description, a set of drawings in which with the illustrative and non-limiting nature, the following has been represented:
Figure 1.-Shows a graph showing the error rate versus the rate of
5 Penetration obtained with the method of the invention applied to three fingerprint bases: the FVC2000 082a and the FVC2002 081 a (from the fingerprint bases of MFingerprint Verification Competition-) and a fingerprint base generated from users of an experimental system Online identification.
10 Figures 2a-2d.-They show graphs where the ratio of false acceptance (FAR) and false rejection (FRR) is represented against a threshold that measures the percentage of different labels between the characteristic vectors obtained with the method of invention applied to the online footprint base: (a) Without applying multibiometric fusion. (b) With merger (minimum operator) of 3 samples per footprint in the phase of
15 record. (e) With merger (sum operator) of 2 fingers and (minimum operator) of 3 samples per finger in the registration phase. (d) With fusion of a sample of a finger and password.
Figures 3a-3d.-They show some images where possible results of the
20 basic steps of the method to generate a vector of characteristics of a fingerprint: (a) First image from which the method starts, a grayscale image captured by an optical fingerprint sensor (taken from FVC2002 081). (b) Second smoothed image, with homogeneous regions of labels, where each pixel is associated with one of eight labels (each of the eight labels is
25 represents a different plot). (cl Window with distinctive information centered on the convex core point. (d) Pixel is sampled from the window to obtain the feature vector.
Figure 4 - Shows the functional block diagram of a device that extracts a
30 distinctive feature vector of a fingerprint. Blocks drawn with dashed lines are used or not depending on the application.
Figure 5.- Shows a representation of the 27 x 27 smoothing block that it uses
Three sub-blocks: a first sub-block of 3 x 3 window size, which softens 3 x 3
35 pixels; a second 9x9 window size sub-block, which applies smoothing over 3x3 results of the previous sub-block; and, finally, a 27 x 27 sub-block, which applies smoothing over 3 x 3 results of the previous sub-block.
Figure 6.-Shows the functional block diagram of a device that implements an indexing and identification / authentication method by fingerprint and possible key: (1A) Registration phase without key (2A) Verification phase without key (1 8) Phase from
5 password registration (28) Password verification phase. Roads that are not marked are used both in the registration and verification phases. The blocks and paths drawn with dashed lines are used or not according to the application.
Figure 7.- It shows the functional block diagram of a device that implements a biometric identification encryption system Fingerprint authentication:
(1) Registration phase (2) Verification phase. Roads that are not marked are used both in the registration phase and in the verification phase. The blocks and paths drawn with dashed lines are used or not according to the application.
PREFERRED EMBODIMENT OF THE INVENTION
In view of the figures, an embodiment of the object 20 of the invention described herein is described below.
The method of the invention has been implemented in an electronic device also object of the invention; for a particular embodiment of the device an implementation is selected in an Xilinx FPGA, a Virtex 6 XC6VLX240T-3FFG1 156,
25 containing 37680 slices and 416 RAM blocks of 36 Kbits. The method of the invention could also be implemented in a specific application integrated circuit (ASIC); in such a case, the electronic device would be even smaller, would consume less power and could be integrated with fingerprint sensors (for example, capacitive type) using eMOS technologies.
In a preferred embodiment of the method of identifying fingerprints from a feature vector extraction of the invention has been implemented as follows. A block assigns each pixel of an image, corresponding to a fingerprint, one of eight possible labels, 35 using 3 bits to encode the labels, 8 bits to encode intensities (Illuminance) of the grayscale image of the fingerprint and 14 bits for gradients (obtained using Sobel 3x3 filters), generating a second
Application No. F.Erccliva F.OEPM
 10/30/2015 11/02/2015
image.
A block that applies smoothing over the second image uses a 3x3 smoothing block cascaded with another 9x9 block, cascaded with another 27x27 block. A block detects the convex core of the footprint as the point where three of the four regions of the smoothed tetra-directional image intersect.
The second smoothed image is processed to select a distinctive window centered on the convex core. In this case, the window has dimensions of 10 129x129 pixels, sampled from 8 to 8 pixels, that is, an 867 bit string is generated per captured fingerprint. The implementation also includes blocks for calculating quality indicators, a memory for storing the image of the grayscale print and a block that applies a rotation on the grayscale image stored in the memory. All this occupies 18.31% of slices and 15.87% 15 of RAM blocks, being able to reach a maximum operating frequency of
257.7 MHz and considering a footprint with 374 rows x 388 columns (such as those of the FVC 2002 081). Applying a pixel-to-pixel processing of the fingerprint, this means that the time to obtain the 867 (17x17x3) bits of the characteristic vector of a capture (without rotations) can be 0.56 ms (374 x 388 f 2577 I-Is).
20 If 3 rotations of the fingerprint are taken into account to register a user, 2601 (3 x 17 x 17 x 3) bit vectors are stored per user. In the Virtex 6 FPGA considered, the vectors of almost 5900 users can be stored in the 416 RAM blocks of 36 Kbits.
25 The block that orders the levels of similarity between the input vector and the stored ones, which applies an insertion sludge and generates a list with 50 candidates, occupies 11.48% of the total slices of the Virtex 6 FPGA that we are considering and allows a maximum frequency of 207.5 MHz. This means that the time invested in
30 the recovery phase is quite low (several tenths of a millisecond to order 5900 users).
The same device, in this case the same FPGA, can include all the blocks required by the indexing and recovery phases of the method of the invention. At
35 case of the Virtex 6 XC6VLX240T-3FFG1156 of Xilinx as a single device, 66 RAM blocks of 36 Kbits are used to store the fingerprint and the distinctive window (considering a fingerprint with 374 rows x 388 columns like those of the FVC 2002 081)
and 350 RAM blocks of 36 Kbits are available. which allow to register more than 4950 users (considering 2601 bits per user)
The method of the invention, implemented in this FPGA embodiment example, has been evaluated with two of the fingerprint bases of ~ Fingerprint Verification Competition ~: the FVC2000 OB2a and the FVC2002 OB1 a, with 800 captures each. It should be taken into account that bases such as FVC are built with many poor quality catches and poorly acquired to prove the goodness of complex identification and indexing techniques. In addition, it has also been considered a fingerprint base with 560 captures, generated from users of an experimental online identification system.
Finger placement is important to correctly extract the feature vector. In the context of the online application in which the user wants to identify himself, the finger is usually placed properly. For example, in the online user registration experiment in which 560 fingerprints were captured, 23 captures did not allow the distinctive window to be extracted (4.11%). In the bases of fingerprints FVC2000 OB2 and FVC2002 OB1, with 800 captures, as the application context is different, the number of captures from which it is not possible to extract correctly the vectar of characteristics is much higher: 149 in the FVC20aO 082 (18. 6%) and 104 in the FVC2002 DSl (13%).
The quality of the captured image is also important to evaluate, because in a capture of 560 fingerprints, 10 captures (1.79%) were of very poor quality (because the prints were really damaged). In the bases of fingerprints FVC2000 OB2 and FVC2002 081, 16 (2%) and 24 (3%) catches are also of very poor quality (due to deteriorated fingerprints or not well acquired catches).
Figure 1 represents the error rate versus the penetration rate obtained with the technique of the invention implemented in this exemplary embodiment and applied to the three fingerprint bases considered. To obtain this figure, in all the bases the fingerprints have been removed from which its feature vector cannot be extracted correctly and that are of very poor quality (the percentages discussed above), since with the device of the invention, which interacts with the user, these percentages would have been reduced to zero. An improvement has been applied to the grayscale images (applying complex filters) and reinforced the convex core detection technique. As registered feature vectors
at the base, the first capture of each individual has been taken (with 5 rotations in the FVC2002 081, 3 rotations in the FVC 2000 082, and none in the third of the bases). As input vectors, all other captures have been taken, without any rotation. To calculate the level of similarity between the vector of
5 input and stored previously rotated (in the case of FVC2002 081 and FVC 2000 082) the maximum levels of similarity have been selected with each of the stored
The average penetration rate to be carried out when they are not wanted
10 make mistakes in the recovery of the holder of the entry fingerprint ("incremental search scenario"), under the same conditions as the results of Figure 1, has been 3.16% in the FVC2000 082, 2.88% in the FVC2002 DBl and 1.62 % in the third of the bases analyzed
15 The same device, in this case the same FPGA, may include an implementation of the winning hash function of the last NIST SHA-3 competition, Keccak, to allow identification / authentication by the double factor of ~ who you are · and " what you know · This function to generate 512 bits occupies 1188 slices (3.1 5% of the total) allowing a maximum frequency of 435.3 MHz.
20 Figure 2 represents the ratio of false acceptance (FAR) and false rejection (FRR) against a threshold that measures the level of dissimilarity (percentage of different labels between feature vectors). The results correspond to the footprint base with online captures. Figure 2a illustrates the results of an identification
25 without biometric fusion. The value where the reasons intersect (EER) is 5.4%. Figure 2b illustrates the results of a fusion identification of 3 samples captured by each individual's fingerprint in the registration phase and a sample captured in the verification phase. The value where the reasons intersect (EER) is 2.5%. Figure 2c illustrates the results of a 2 finger fusion identification
30 per individual, with 3 samples captured by each finger in the registration phase and a sample captured by each finger in the verification phase. The value where the reasons intersect (EER) is 0.9%. Figure 2d illustrates the results of a merged identification of the feature vectors with a hash function that returns 512 bits applied to a password or password for each individual. The value
35 where the reasons intersect (EER) is 0%.
The same device, in this case the same FPGA, can include all the blocks
Application No. F.Erccliva F.OEPM
 10/30/2015 11/02/2015
2. 3
required by the protection technique of the method of the invention. In the case of the Virtex 6 XC6VLX240T-3FFG1156 from Xilinx as a single device, the Reed-Solomon coding block for n "'511 yk"' 383 occupies 473 slices (1.26% of the slices) and allows operating at a maximum frequency of 415 MHz. The decoder block
5 of Reed-Solomon for n = 511 and k = 383 occupies 24,763 slices (65% of the total slices) working with a maximum frequency of 78.5 MHz.
In more detail, the method for generating a vector of characteristics of a fingerprint generates the vector which is a fixed-length bit string that 10 represents a compact fingerprint. To obtain this vector, and as detailed above, it is based on a first image such as the capture of the footprint as a grayscale image (Figure 3a), the gradient of) to intensity is determined for each pixel of the image (of the gray levels) in that pixel, and the gradient direction is determined by an angle with 15 relative to a reference axis. The range of possible values of angles is divided into G sub-intervals (gl, .., gG) that do not overlap and whose union gives rise to the full range of possible values, each sub-interval g ~ encompassing angles from Uk.l up to 0k. Each sub-interval, gk, is associated with a label, Ck. The pixel corresponding to the sub-interval to which the angle of the gradient direction in that pixel belongs belongs to each pixel of the fingerprint image. As a result, a second image is generated from the first image of the fingerprint, where in said second image each pixel is associated with a label. Next, a smoothing process is performed to the second image to obtain areas comprising pixels with the same labels (Figure 3b). At least one core point is located
25 convex in the second smoothed image and a window centered on the convex core point is defined (Figure 3c). A sampling of pixels included in the window is performed (Figure 3d). The labels of the sampled, ordered pixels generate the vector of footprint characteristics. If the labels are encoded with bits, the vector is an ordered and fixed-length bit string.
If the number of labels, G, considered is small (for example, four labels), the vector to be generated is not very distinctive of the footprint, that is, there can be many footprints with a similar vector, which translates into a high rate of false acceptance, in the case of an identification / authentication application. If a number such as four labels can be used in classification applications, in which the fingerprints are distributed in pre-established groups according to the similarity of their feature vectors, so that the input fingerprint is classified in one
of those groups or degrees of belonging to several of those groups are assigned.
On the contrary, if a high number of labels is contemplated (for example, sixteen labels), the vector to be generated is very distinctive, but changes too much.
5 for different captures of the same footprint, which translates into a high rate of false rejection, in the case of an identification / authentication application. In a preferred embodiment of the method of the invention for identification / authentication applications, eight labels have been chosen, which is the case illustrated in Figure 3.
The sub-intervals should cover the entire range of angles that the directions of the gradients may have (between 0 ° and 18 (0) in a more or less spaced way. In a preferred embodiment of the invention for identification / authentication applications with G = 8, the following have been chosen: gl = (0 °, 22.5 °), g2
15 = [22.5 ', 45'), 93 = [45 ', 67.5'), 9. = [67.5 ', 90'), 9, = [90 ', 112.5'), g, = [112.5 ', 135 '),
g7 = (135 °, 157.5 °) and g8 = (157.5 °, 1800), choosing the longitudinal axis of the footprint as the reference axis.
The size of the window centered on the convex core depends on the sensor used. 20 For example, for the footprints of the FVC2002 OSl bases (388x374 pixel images captured by an optical sensor), those of the FVC2000 OB2 (256x364 pixel images captured by a low-cost capacitive sensor) and those of an experimental base ( 440x300 pixel images captured by an optical sensor), it has been proven that a suitable window is 129x129 pixels (Figure 3c). 25 As a distinctive and compact representation of the footprint, not all pixels of the window are necessary, but a sampling l / n r down-sampling H) is applied which
,
means preferably use 1 of n consecutive pixels in each row of the window. For example, for the footprints of the bases mentioned above, a 1/8 sampling is applied on the 129x129 pixel window, which means using the
30 information of 17x17 pixels (Figure 3d). In this example, if the eight labels are encoded with 3 bits, the vector obtained for each fingerprint is a 17x17x3 string = 867 bits = 108.4 bytes. These vectors can be encrypted, for security reasons, and / or compressed (for example by applying "Run-Length Encoding"), to consume less memory and / or more easily transmitted.
35 The technique to extract the characteristic vectors of the fingerprints is implemented by using the following basic blocks (Figure 4):
Application No. F. Efffiva F. EPO
 10/30/2015 02111/2015
25
• A fingerprint sensor, which provides a grayscale image of the fingerprint. If the sensor does not apply improvements to the acquired image, an image improvement block is included
• If the position of the finger on the fingerprint sensor can rotate, it will also
5 includes a block that applies rotation to the grayscale input image
• A block that assigns each pixel of the image one of the possible G tags and generates a second image.
• A block that applies smoothing to the second image.
10 • A block to detect the convex nucleus (or several candidate points to be convex nucleus) in the footprint.
• A block to determine the distinctive window, sample its pixels and store the tag values of those pixels in a bit string.
• Additionally, a block that evaluates the quality of the entire process and allows online interaction with the user can be included.
The block that assigns each pixel of the image one of the possible G tags can be implemented using a simple digital circuit. The first step carried out by this block is to calculate the horizontal (GlI) and vertical (Gy) gradients of the image intensity (of the gray levels) with some suitable filter for its hardware implementation (for example, through filters Sobel 3 x 3 using convolution masks with integer values and powers of 2). This step is common in any feature extraction technique. Then, instead of calculating in each pixel the direction of the gradient in a more or less exact way by means of a trigonometric function (in dedicated hardware, a CORDIC processor, "COordinate Rotalion Digital Computer-, is usually used to calculate the angle at whose tangent is (G¡Gx)) and then calculate the sub-interval between the possible Gs to which it belongs to, the technique of this invention compares the values of the gradients G "and Gy Y between logical operators (OR and ANO), operators
30 relational and absolute value operations and multiplication by constant values, which is much more efficient from a hardware point of view. First, the block determines that a belongs to a first quadrant between 0 ° and 900, when G "and Gy have the same sign. Second, within this first quadrant, the block determines that:
35 -if the sub-interval that is evaluated, l (l ~. "(L ~), is fully included in the first quadrant, why throw it (l ~., As (l ~ are less than or equal to 900, then
- if the sub · interval that is evaluated. [O ~ .l, w.) Is partially included in the first quadrant, because a ~ .1 is less than 90 ° but Qk is greater than 90 °, then
The block determines that a belongs to a second quadrant between 900 10 and 1 BOo, when Gx and Gy have different signs. In this case, within this second quadrant, the block determines that:
. if the sub · interval that is evaluated, [a ... 1, O ~), is fully included in the second quadrant, because both U¡, .1 and Ok are greater than or equal to 15 than 900, then
. if the sub-interval that is evaluated, [Q¡, .1, O ~), is partially included in the second quadrant, because Ok is greater than 90 ° but a.k_l is less than 90 °, then
Where tan (a.k) and tan {ak. ') Are previously known constant values once the sub-intervals to be considered have been set, Qk = [ak.l, (lk).
25 The digital circuit that implements these operations can use fixed point arithmetic, and words of log2G bits to encode the possible G tags corresponding to the G sub-intervals gk.
30 The smoothing block applies a window size S x S, where S depends, in general, on the type of fingerprint sensor used. Since parallel smoothing S x S pixels (to achieve high processing speed) can be very expensive, you can choose to cascade several smoothing sub-blocks one after the other. If the value of S can be factored as S = s1 x 52 x. x sn: first it can be used
35 a sub-block with window size s1 x 51, which applies the smoothing function
about sl x sl pixel tags; the second sub-block with window size (sl x s2) x (s1 x s2). which applies the smoothing function on s2 x s2 labels previously softened by the previous sub-block, and so on until the last sub-block with window size (sl x s2 x. x sn) x (s1 x s2 x ... x sn), which again applies the smoothing function on sn x sn tags previously softened by the previous subblock. For example, for optical sensors that capture 388x374 images
or 440x300 pixels or capacitive sensors that capture 256x364 pixel images, it has been proven that a 27 x 27 smoothing block is suitable, which can be done with three smoothing sub-blocks connected in cascade: a first 10-size sub-block " 3x3 window io, which smoothes 3x3 pixels; a second 9x9 window size sub-block, which applies smoothing over 3x3 results of the previous sub-block: and finally a sub- 27 x 27 block, which applies smoothing over 3 x 3 results of the previous sub-block (Figure 5). In a preferred embodiment, the smoothing function sj x sj considers a window of pixels around the analyzed pixel and assigns to it the value of the label that most times appears throughout the window. The technique of connecting n sub-blocks in cascade allows reducing the required hardware and the latency of the smoothing process because processing S x S values in parallel is much more expensive than processing in parallel sj x sj pixels. For example, if the pixels of the image are processed one by one, smoothing of the entire image can be
20 perform with these cascading sub-blocks (and the necessary record banks) by inverting as many clock cycles as the pixels have.
The block that calculates the convex core (or the candidate points to be) can use widely known techniques, such as those based on the calculation of the Poincaré index 25. An advantage of the method of the invention is that it makes it possible to reinforce the detection of this point with hardly any computational cost, as described below. Starting from the second smoothed image, a softened tetra-directional image is obtained directly. For example, if G = 8 and the eight sub-intervals are 9, = (O ', 22.5'), 9, = (22.5 ', 45'), 9 '= (45', 675 '), 9. = (67 5 ', 90'), 9, = (90 ', 112.5'),
30 g6 = (112.5 °, 135 °), g7 = [135 °, 157.5 °) and g8 = [157.5 °, 180 °), the following four sub-intervals can be obtained directly g'¡ = g U g8, g'2 = g2 U g3, 9'3 = g4 U g5 Y g'4 = g6 U g7. Since each pixel of the second smoothed image is represented by 3 bits, obtaining the smoothed tetra-directional image is as simple as truncating 3 to 2 bits of each pixel, if the labels are properly encoded. He
The convex core can be determined as the point where three of four homogeneous regions of the smoothed tetra-directional image intersect, which are the three regions that encompass most of the ridges with convex curvature. As the correct detection of the convex core is important to correctly extract the feature vector, several points can be considered as candidates and extract the feature vectors associated with them.
The technique of the invention makes it possible to contemplate traces acquired with the finger rolled with respect to the longitudinal axis of the sensor. The window with information representative of the fingerprint described above is characterized by its invariance before translations of the finger on the sensor since the central point of the window is the convex core point. However, the window is not invariant to rotations. To make sure that
10 different captures of the same footprint acquired with possible rotations have a high level of similarity with their corresponding feature vector stored in the base, a solution with low hardware cost is to include a block that allows the image of the fingerprint to be rotated on a scale of gray R rotations prior to obtaining the second image can be taken into account (for example, with R = 5: -22.5 °,
15 -11.25 °, 0 °, 11 .25 ° and 22 .5 °). If the captured image of the footprint, whose pixels have Cartesian coordinates (x "Yl), an angle 3 is rotated with respect to the coordinate pixel (xc, Ycl, the pixel coordinates become now (x" y, ) . Is
Operation can be expressed mathematically as follows:
1 o Xc cos (jJ) -sen (p) O · 1 O
- xi l · x'l [X '
O 1 y, sin (p) cos (P) O O 1 -Yc Yi = Yt O O 1 O O .0 O 1 1 1.
20 For example, if the pivot point is taken as the center point of the fingerprint image (for an image of 374 rows and 388 columns, the values for Xc and Yc are 187 and 194, respectively), and the angle of rotation is chosen as 11.25 °, the previous expression can be reduced to the following: 0.9808 -0.1951 41.4407 IIX'I IX,]
10.1951 0.9808 -32.7542 ~ '= Y, For example, if each pixel of the input image is addressed by its coordinates (Xl, y,) in the memory where the capture is stored, the block that applies a roll of 11 .25 °, like the previous one, now directs that pixel by its coordinates (x "y,). Since the rotations contemplated are fixed, this block implements a linear transformation between (Xi, YI) Y (X" y,) with for constant meters. so it does not require the use of multipliers. In a possible embodiment of this block, the number of rotations can be programmable as well as the parameters associated with
the rotations
To make the technique of classification, indexing or robust identification to rotations. more or less angles are contemplated according to the level of rotation that is to be supported. Rotations can be contemplated in the indexing or recording phase and / or in the recovery or verification phase and do not have to coincide in number. Thus, for example, in the indexing or registration phase, if P convex core candidates and V pixels per distinctive window are considered, a string of P x V x 3 bits can be extracted for each rotation, as discussed above. If R is contemplated
10 rotations, the total index used to represent a fingerprint capture concatenates the R bit strings, resulting in a characteristic vector of a length of R x P x V x 3 bits.
The method generates a digital identification number (a vector of R x P x V x 3 bits)
15 which is associated with the individual who has the fingerprint, so that a base can be generated with the N numbers associated with the individual N registered (the vectors can be encrypted, for security reasons, and / or compressed, to consume less memory and / or more easily transmitted). In multi-biometric applications that use O fingers per individual, a vector of O x R x P x V x 3 bits is generated for each individual
20 concatenating the O identification numbers that are obtained from the fingerprint of each finger. In multi-biometric applications that use Z samples from the same finger of the individual, a vector of Z x R x P x V x 3 bits is generated for each individual concatenating the Z identification numbers that are obtained from each sample.
25 In the recovery phase, an ordered list of individuals registered in the database is generated, calculating a level of dissimilarity (or similarity) between the input characteristics vector and each stored vector. If the vectors have been encrypted and the tablets must be decrypted and / or decompressed to calculate the level of
30 dissimilarity. The level of dissimilarity is calculated as the percentage of labels that are different between the access vector and each stored vector. In the case of multibiometry with O fingers, the overall dissimilarity is obtained from the fusion (for example with the sum operator) of the dissimilarities of each finger. In the case of multi-biometrics with Z samples of the same finger, the overall dissimilarity is obtained as the fusion (by
35 example, with the minimum operator) of the dissimilarities with each sample. The list of registered individuals is ordered from lowest to highest level of dissimilarity, being able to truncate the list into a given number of individuals or a maximum percentage of dissimilarity.
In an identification application. the candidate is selected from the list he has
less dissimilarity (or. equivalently. greater similarity). In an application of
authentication, the level of dissimilarity is compared with a threshold.
5 To generate different digital identification numbers for the same footprint
fingerprint, the number obtained according to the method of the invention can be combined with the
result of a non-invertible (hash) function of a password or password. being the
combination: (a) a simple concatenation or (b) a given interleaving of the
bits or (c) an XOR operation between the two (for this they must have the same length of
10 bits), a combination that allows indexing and identification (or authentication) by the
double factor of "who you are · (the fingerprint) and" what you know · (the password or password) and
which allows to revoke compromised identification numbers, generating others
new with a new password or password.
fifteen The technique to carry out the recovery or verification phase is implemented
using the following basic blocks (Figure 6): (a) A memory to store
Digital ID numbers or feature vectors, Si (i = 1, .... N). of N
individuals and thus allow their registration. (b) A block to calculate the similarity between the
identification number obtained from the fingerprint to be identified, S ', and the N numbers
twenty stored, Bi (i = 1, .., N), block that calculates the equality between a label of
input and other stored (both coded with 3 bits, in the case of G = 8)
preferably with 3 XOR operators whose outputs are the inputs to a
NOR operator, then a counter calculates the number of equal tags
between the input vector and each stored vector (the level of dissimilarity is the
25 complement of similarity level). (e) In the case of verification, a block that
compare the previous result with a threshold In the case of recovery, a block
which orders the similarities with each stored vector from highest to lowest, up to
reach a maximum number, M, of candidates. There are many algorithms of
management reported in the literature (for example based on binary trees or n-
JO years, methods of insertion, etc.), being able to choose one or air depending on the
speed objectives and resources to use (the fastest algorithms usually
need more resources than Jos slower and vice versa).
The digital identification numbers, B, generated can be protected by
J 5 technique called Fuzzy Commilment. as follows'
in the registration phase: a random code word, Ci, is associated to each user
(1 = 1, .., N), of an error correction code (S is added with zeros or ones up to
that its size is the same as that of the Gi) and the calculation and storage of a hash function of Ci, hash (Ci), and the results Hi = (8 XOR Ci), which are called data of help.
in the verification phase, given an entry number, B ', the Ci' = B 'XQR Hi (if B' is similar to B, Ci 'will be similar to Ci), the error correction code is applied to Ci 'and hash is applied to its result. If the result matches any hash of the stored ones, the corresponding user will be identified (if N = 1, the user will be authenticated).
In a possible communication phase, ei or B = Ci XOR Hi can be used as secrets of those that generate cryptographic keys to encrypt or authenticate messages.
The error correction code is preferably a Reed-Solomon code, which treats the labels encoded with 3 bits as symbols.
To choose the error rate to be corrected by the Reed-Solomon code, you can apply: (a) the percentage of different labels for which it is obtained that the reason for false rejection coincides with the reason for false acceptance, if a commitment is wanted optimal between both rates; (b) the percentage of different labels for which a null false acceptance is obtained, if the intrusion is to be eliminated; or (c) the percentage of different labels for which a false null rejection is obtained, if the denial of the service is to be eliminated.
The technique to carry out the protection of the feature vectors is implemented by the following basic blocks (Figure 7): (a) An acquisition block adapted to acquire a random number, key or password and apply an encoder of a error correction code to generate a code word. (b) A block adapted to generate public support data from the vector of fingerprint characteristics and the code word, to calculate a hash function of the code word and store the results in a memory. (A decoder block of an error correction code adapted to retrieve a secret from an extraction of the feature vector and the stored help data from the fingerprint associated with the secret.

权利要求:
Claims (8)
[1]
1. Method to generate a vector of characteristics of a fingerprint for identification from a first image of it in grayscale that contains ridges and valleys of the fingerprint. method characterized because it comprises:
a) determine for each pixel of the first image, P'I (where ij refer
10 to the row and column of the pixel in the image), the gradient of the intensity of the image (of the gray levels) in that pixel, b) determine the direction of the gradient by means of a Cti angle with respect to an axis of reference, c) divide the range of possible values of angles, al !, into G sub-intervals (gl "
15 gG) that do not overlap and whose union gives rise to the full range of possible values, encompassing each sub · interval 9k angles from a value a k.1 to ak, d) label each gk sub · interval with a label, Ck, e) associate, for each pixel Pij of the first image, the label corresponding to the sub · interval to which the angle belongs to jj corresponding to that pixel,
20 f) generate a second image from the first image, where in said second image each pixel associated with a label, g) perform a smoothing process to the second image to obtain areas comprising pixels with the same labels, h) locate at least one convex core point in the second smoothed image.
25 i) define a window centered on the convex core point, j) sample the pixels included in the window. and k) generate the vector from the pixel tags sampled in the previous step, in an orderly manner
Method according to claim 1, characterized in that each sub-interval gk comprises angles between 0 ° and 180Q
•
3 Method according to claim 1 characterized in that the determination of the sub-range to which the angle a of the gradient in a pixel belongs is determined from the calculation of the horizontal gradient (Gx) Y of the vertical gradient (Gy) of the intensity of the image (of the gray levels) in that pixel.
N "requested F. Efccti, 'from F. SPTO
 30110/2015 02111/2015
33
[4]
4. Method according to claim 3 characterized in that the determination of the sub
gk interval:; [0 .... 1, at) that is associated with the pixel p ~ includes:5 • determine the sign of G ~
• determine the sign of Gy
• determine that:
• a belongs to the first quadrant of angles between Oll and 90 °, when G ~ and Gy have the same sign, and, within this first
10 quadrant distinguishes two situations according to the range of angles covered by each sub-interval to be evaluated:
- if the sub-interval being evaluated, (a "'l. a¡,), is fully included in the first quadrant, because both ak.l and Dk are less than or equal to 15 90 °, then
• if the sub-interval is evaluated. ((Ik.l, lb) is partially included in
the first quadrant, because Ot., is less than 90 ° but ak is greater than 90 °, then
25 •a belongs to the second quadrant between 90 ° and 180 °, when Gx and Gy have different signs, and, within this second quadrant, two situations are distinguished according to the range of angles covered by each sub-interval to be evaluated:
30 -if the sub-interval being evaluated, [! lIt-l,! lit), is fully included in the second quadrant, because both ak.l and nk are greater than or equal to 90 °, then
- if the sub-interval that is evaluated, [ak-l, a ~), is partially included in the second quadrant, because (Ik is greater than 90 ° but ak.l is less than 90 °, then
[5]
5. Method according to claim 1 characterized in that the smoothing process calculates for each pixel p ~ of the second image preferably which of the labels is the one that most often appears in a window of size S x S pixels of the second
5 image, window centered on the pixel to be smoothed, where S can be factored as S = s1 x s2 x ... x sn, a method comprising:
• start with windows of size s1 x 51 pixels and apply smoothing to your s1 x s1 tags,
• continue with windows of (51 x 52) x (s1 x s2) pixels and apply the
10 smoothing over 52 x 52 labels softened previously in the previous step,
• proceed like this until you reach the window tamale (s1 x s2 x ... x sn) x (s1 x s2 x .. x sn) pixels and apply the smoothing over sn x sn labels previously softened in the previous step.
Method according to claim 1 characterized in that the determination of the convex core additionally comprises:
• make another division of the range of possible angle values, to lj, in 4 sub
20 intervals (g'l, ..., 9'4) that na overlap and whose union gives rise to the full range of possible values,
• label each g'k sub-interval with a label, C'k,
• convert the second smoothed image, in which each of the pixels is associated with one of G labels (Cl ...., cG), where-preferably G> 4, in
25 a smoothed tetra-directional image, in which each of its pixels is associated with one of four labels (C · 1 • ... • c ',) and the conversion includes:
• change each Ck tag associated with the gk angle sub-interval for
that label C'k associated with the sub-range of angles g'k that verifies that the intersection gk n g'k is the largest. Y
• determine the convex nucleus as the point where three of four homogeneous regions of the smoothed tetra-directional image are touched, which regions that encompass most of the crests with convex curvature
Device for generating a vector of characteristics of a fingerprint from an image thereof according to the method described in claims 1 to 6, a device that is associated with means of image capture of the fingerprint and characterized because it includes:
• a block of assignment of labels destined to assign to each pixel of the
5 image one of G possible labels, which allows to generate the secondimage,
• a smoothing block intended to perform a smoothing process to the second image to obtain areas comprising pixels with the same labels,
10 • a block for determining the convex core in the fingerprint, intended to locate at least one convex core point in the second smoothed image,
• a window block designed to define a window centered on the convex core point, sample pixels included in the window, obtain the label of each sampled pixel and generate the vector from the
15 labels obtained in an orderly manner
[8]
8. Device according to claim 7 characterized in that additionally
It comprises a memory block intended to store the captured image of the fingerprint.
[9]
Device according to claim 7, characterized in that it comprises an image orientation block intended to rotate or rotate it to a certain position in the event that the fingerprint captured by the image capture means 25 of the fingerprint is not in a certain orientation, a block that preferably rotates through fixed angles to apply linear transformations between original pixels (x "y¡) and pixels of the rotated images (xr, Yr) can be the parameters of the fixed linear transformation for each rotation, and block that in a possible embodiment it can be programmable in the number of rotations as well as the parameters
30 associated with rotations.
[10]
10. Device according to claim 7 characterized in that the label assignment block comprises:
35 • a filter preferably of Sobel 3x3 with convolution masks with integer values and powers of 2 to calculate the horizontal gradations (G ~) and the vertical gradients (Gy) of the ridges of the footprint, and
N "application F. Efccti, · to F.OEPM
 30110/2015 02111/2015
36
• logical operators type OR and ANO, relational operators and operations of absolute value and multiplication by constant values.
5 11. Device according to claim 7 characterized in that the smoothing block is adapted to process the second image by sweeping its pixels one by one and providing the pixels of the smoothed image also one at a time, where the smoothing block defines a window of size " or SxS, where S can be factored as S = s1 x s2 x .. x sn, and where the smoothing block comprises a
10 series of records and n sub-blocks with a hybrid series-parallel architecture of which:
• a first sub-block with window size s1 x s1 is adapted to apply a parallel smoothing function on s1 x s1 pixel tags that have been stored in the corresponding registers, sub-block
15 whose resulting label is stored one after another in a series of records;
• a second sub-block with tamat "the window (sl x s2) x (s l x s2) is adapted to apply a parallel smoothing function on s2 x s2 labels previously softened by the first sub-block and available in the
20 corresponding records that store the output of the first sub-block, sub-block whose resulting label is stored one after the other in a series of records:
• so up to an enesimo sub-block with window size (s1 x s2 x. X sn) x (S1 x s2 x x sn), which again applies the parallel smoothing function
25 on sn x sn labels previously softened by the previous sub-block and available in the corresponding registers that store the output of the previous sub-block, sub-block whose output provides the pixel label in the smoothed image.
[12]
12. Device according to claim 7 characterized in that it additionally comprises an information fusion block adapted to:
• Acquire a password or password
• apply a non-invertible (hash) function to that key or password and
35 • combine the result of the previous step with the vector of footprint characteristics.
[13]
13. Device according to claim 7 characterized in that it additionally comprises the following blocks:
5 • an acquisition block adapted to acquire a random number, password or password and apply an error correction code encoder to generate a secret,
• a block with XOR operators adapted to calculate and store some
public support data from the vector of footprint characteristics and the secret 10 and
• a decoder block of an error correction code adapted to retrieve a secret from an extraction of the feature vector and stored help data from the fingerprint associated with the secret.

类似技术:

---

公开号 | 公开日 | 专利标题

---

ES2791776T3|2020-11-05|System and method for biometric authentication in connection with camera-equipped devices

---

KR102202690B1|2021-01-13|Method, apparatus and system for recognizing fingerprint

---

Moujahdi et al.2014|Fingerprint shell: Secure representation of fingerprint template

---

Jain et al.2013|Fingerprint template protection: From theory to practice

---

US10169638B2|2019-01-01|Transformed representation for fingerprint data with high recognition accuracy

---

US20040042645A1|2004-03-04|Fingerprint recognition method, and fingerprint control method and system

---

Subban et al.2013|A study of biometric approach using fingerprint recognition

---

Kumar et al.2010|Combination of symmetric hash functions for secure fingerprint matching

---

EP1410312A1|2004-04-21|Method of hash string extraction

---

Jin et al.2009|Secure minutiae-based fingerprint templates using random triangle hashing

---

CN102273128B|2015-07-15|Identification or authorisation method, and associated system and secure module

---

Aravindan et al.2017|Robust partial fingerprint recognition using wavelet SIFT descriptors

---

Ali et al.2020|Securing biometric user template using modified minutiae attributes

---

CN109598247A|2019-04-09|Two dimensional code identity identifying method based on vein image minutiae point and patterned feature

---

Selwal et al.2017|Low overhead octet indexed template security scheme for multi-modal biometric system

---

ES2556276B1|2016-11-08|Fingerprint identification method and device that uses it

---

Severance2015|Anil Jain: 25 years of biometric recognition

---

CN110956468A|2020-04-03|Fingerprint payment system

---

CN111027404A|2020-04-17|Fingerprint identification method based on fingerprint protection template

---

Arjona et al.2016|A dual-factor access control system based on device and user intrinsic identifiers

---

Ramya et al.2019|A comparative analysis of similarity distance measure functions for biocryptic authentication in cloud databases

---

KR100670244B1|2007-01-16|Fingerprint sensor for both sides at the same time, and method of maching fingerprint the same

---

Kong et al.2018|A Cylinder Code-Based Partial Fingerprint Matching Algorithm for Small Fingerprint Scanners

---

Dwivedi et al.2018|Securing fingerprint template using noninvertible ridge feature transformation

---

US10733415B1|2020-08-04|Transformed representation for fingerprint data with high recognition accuracy

同族专利:

---

公开号 | 公开日

---

EP3093793A1|2016-11-16|

---

EP3093793A4|2017-11-01|

---

ES2556276B1|2016-11-08|

---

WO2015015022A1|2015-02-05|

引用文献:

---

公开号 | 申请日 | 公开日 | 申请人 | 专利标题

---

---

US5631971A|1994-05-24|1997-05-20|Sparrow; Malcolm K.|Vector based topological fingerprint matching|

---

US5841888A|1996-01-23|1998-11-24|Harris Corporation|Method for fingerprint indexing and searching|

---

JP2815045B2|1996-12-16|1998-10-27|日本電気株式会社|Image feature extraction device, image feature analysis device, and image matching system|

---

US6185318B1|1997-08-22|2001-02-06|International Business Machines Corporation|System and method for matching images an aligned string-based representation|

---

US7046829B2|2001-05-30|2006-05-16|International Business Machines Corporation|Fingerprint verification|

---

US7013030B2|2002-02-14|2006-03-14|Wong Jacob Y|Personal choice biometric signature|

---

US7313256B2|2002-09-30|2007-12-25|Motorola, Inc.|Progressive fingerprint matching system and method|

---

US8899487B2|2005-08-18|2014-12-02|Ivi Holdings Ltd.|Biometric identity verification system and method|

---

US7327859B1|2007-02-14|2008-02-05|Lam Ko Chau|Methods and systems for automated fingerprint recognition|

---

US20080273770A1|2007-05-03|2008-11-06|Upek, Inc.|Fast Fingerprint Identification And Verification By Minutiae Pair Indexing|

---

US8276816B2|2007-12-14|2012-10-02|Validity Sensors, Inc.|Smart card system with ergonomic fingerprint sensor and method of using|

---

CN101620677B|2009-03-11|2012-03-21|刘鸣宇|Fingerprint identification method based on triangulation and LOD technology|

---

CN101996318B|2010-11-25|2013-01-30|杭州晟元芯片技术有限公司|Method for rapidly calculating fingerprint similarity|

---

CN102368242A|2011-09-07|2012-03-07|常州蓝城信息科技有限公司|New fingerprint database retrieval method|

---

CN102495886A|2011-11-28|2012-06-13|河南理工大学|Method for searching fingerprint database based on quantum algorithm|CN107203647B|2017-07-20|2020-04-07|长江大学|Student information acquisition method and device based on biological characteristics|

法律状态:
2016-11-08| FG2A| Definitive protection|Ref document number: 2556276 Country of ref document: ES Kind code of ref document: B1 Effective date: 20161108 |

---

2021-09-28| FD2A| Announcement of lapse in spain|Effective date: 20210928 |

优先权:

---

申请号 | 申请日 | 专利标题

---

ES201300721A|ES2556276B1|2013-07-31|2013-07-31|Fingerprint identification method and device that uses it|ES201300721A| ES2556276B1|2013-07-31|2013-07-31|Fingerprint identification method and device that uses it|

---

PCT/ES2014/000131| WO2015015022A1|2013-07-31|2014-07-30|Fingerprint identification method and device using same|

---

EP14832457.7A| EP3093793A4|2013-07-31|2014-07-30|Fingerprint identification method and device using same|