CN101526994A - Fingerprint image segmentation method irrelevant to collecting device - Google Patents

Abstract

The invention discloses a fingerprint image segmentation method irrelevant to a collecting device, which not only has better segmentation effect, but also avoids the defect that when the traditional algorithm segments fingerprint images from different devices, different threshold values need to be set or different classifiers are trained. The fingerprint image segmentation method basically has collecting device independency and is more suitable for Internet application environment which uses multiple types of fingerprint collecting devices. The method comprises the following steps: 1) an input picture is divided into W*W no-overlapping image blocks; 2) CMV indexes of the image blocks are obtained and are normalized; 3) K-means clustering is carried out for the image blocks; 4) the classification of foreground blocks and background blocks is ensured to obtain a primary segmentation result; and 5) a final segmentation result is obtained after the morphologic processing.

Description

A kind of fingerprint image dividing method that has nothing to do with collecting device

Technical field

The present invention relates to automatic fingerprint recognition field, specifically a kind of fingerprint image dividing method that has nothing to do with collecting device.

Background technology

It is crucial pre-treatment step in the Automated Fingerprint Identification System that fingerprint image is cut apart, and by with accurately the cutting apart of prospect and background, makes subsequent treatment can concentrate on the effective coverage, thereby reduces calculated amount, improves the accuracy that detail characteristics of fingerprints obtains.Fingerprint image is cut apart and is at first chosen and one or more description features of calculated fingerprint, then design with realize suitable partitioning algorithm.

Be noted that existing fingerprint image partitioning algorithm when carrying out effect analysis, only on a certain single fingerprint base, do the experiment comparison and analysis usually, perhaps adopt different parameter or threshold value for different fingerprint bases.For the supervision partitioning algorithm is arranged, often test sample book and training sample be from same database, and all be to adopt same collecting device to obtain from the fingerprint image of same database.

In recent years, in the living things feature recognition field, the equipment interoperability issue had obtained a lot of concerns, and the equipment interoperability is meant that biological recognition system is to the adaptive faculty from the data of different acquisition equipment.Present most of Automated Fingerprint Identification System is all at a certain particular fingerprint collecting device design, when this system handles during from the view data of dissimilar collecting devices, because the fingerprint image of these equipment collections has different picture quality, resolution and gray shade scale usually, cause this system cut apart, decline in various degree appears in enhancing, coupling etc. on performance when handling.

There is the equipment interoperability issue equally in present fingerprint image cutting techniques, be embodied in: with regard to the feature that (1) is chosen when cutting apart, for fingerprint image from a plurality of dissimilar collecting devices, background occurs obviously intersecting with the eigenwert of prospect, same eigenwert, for a kind of fingerprint image, belong to background, for another kind of image, may belong to prospect.(2) with regard to dividing method, so that supervise algorithm to be arranged is example, because present most of algorithms are when the design category device, the image that training sample and test sample book are gathered from same equipment, cause using the appearance when cutting apart of this sorter from the fingerprint image of miscellaneous equipment bigger cut apart error.As the important preprocessing step of Automated Fingerprint Identification System, the equipment interoperability issue of fingerprint image dividing method ought to obtain enough concerns, but with regard to the knowledge that the inventor grasps at present, did not also have open source literature that this type of problem is discussed.At present all concentrating on the fingerprint matching stage about the research of fingerprint image acquisition equipment interoperability issue carries out.A good fingerprint image dividing method should be able to adapt to the fingerprint image of different acquisition equipment acquisition, realize cutting apart reliably under the prerequisite of not making an amendment, do not do parameter configuration, promptly possess the collecting device independence.

Summary of the invention

The present invention is for overcoming above-mentioned the deficiencies in the prior art, a kind of fingerprint image dividing method that has nothing to do with collecting device is provided, this method not only has good segmentation effect, and different fingerprint collecting equipment is had robustness, has possessed the collecting device independence substantially.

For achieving the above object, the present invention adopts following technical scheme:

A kind of fingerprint image dividing method that has nothing to do with collecting device, its step is:

1) input picture is divided into W * W piece zero lap image block;

2) every is asked the CMV index and carries out normalized;

3) each piece is carried out the K-means cluster;

4) determine the classification of foreground blocks and background piece, obtain preliminary segmentation result;

5) the morphology aftertreatment obtains final segmentation result.

Described step 2) in, described CMV index is also carried out the normalized process and is:

A. direction consistance C in the piece

The direction consistance of fingerprint image piece has been described the consistent degree of each pixel direction of this piece; If represent the direction consistance of an image block, then with Coh:

$\mathrm{Coh}=\frac{\left|{\mathrm{\Σ}}_W(G_{s,x},G_{s,y})\right|}{{\mathrm{\Σ}}_W|(G_{s,x},G_{s,y}\left)\right|}=\frac{\sqrt{{(G_{\mathrm{xx}}-G_{\mathrm{yy}})}^2+4G_{\mathrm{xy}}^2}}{G_{\mathrm{xx}}+G_{\mathrm{yy}}}---\left(1\right)$

(G wherein _{S, x}, G _{S, y}) be pixel (x, two times of gradients y), $G_{\mathrm{xx}}=\underset{W}{\mathrm{\Σ}}{G}_x^2,$ $G_{\mathrm{yy}}=\underset{W}{\mathrm{\Σ}}{G}_y^2,$ $G_{\mathrm{xy}}=\underset{W}{\mathrm{\Σ}}{G}_x{G}_y,$ (G _x, G _y) be gradient vector, W is the size of piece, default value gets 8 * 8;

B. piece gray average M:

The gray average of an image block is defined as follows:

$\mathrm{Mean}=\underset{W}{\mathrm{\Σ}}I---\left(2\right)$

Wherein, I is the gray-scale value of each point in the piece, and the W definition is the same;

C. piece gray variance V

The gray variance of a piece is defined as follows:

$\mathrm{Var}=\underset{W}{\mathrm{\Σ}}{(I-\mathrm{Mean})}^2---\left(3\right)$

Wherein I, W definition is the same:

After asking for the CMV value of piece,, make it value all between [0,1] to the normalization of feature value.

Utilize the K-means method that each image block is carried out cluster according to its CMV eigenwert, the cluster parametric description is: each image block is a sample, and each sample has 3 features, is respectively C, M, V; The classification number is set at 2; 2 image blocks of picked at random are as the initial distance center; Distance calculation between all the other each pieces and 2 the piece centers adopt Euclidean (Euclidean) distance square.

In the described step 4), the net result of K-means method is just classified each image block with classification value (1/2), therefore need definite which classification value to represent prospect or background, the method of determining prospect and background classification value is: the eight neighborhood pieces of getting the picture centre piece, what certain classification value number was many is foreground blocks classification value, as classification value in the eight neighborhood pieces 6 " 2 " values are arranged, 2 " 1 " value thinks that then classification value " 2 " is a prospect classification value.

Utilize morphology to cut apart aftertreatment in the described step 5), owing to be subjected to noise effect, some isolated background piece and foreground blocks can appear in the preliminary segmentation result, should take principle to handle: for isolated foreground blocks, if for the number of foreground blocks is less than 4, then this foreground blocks is labeled as the background piece in its eight neighborhoods piece; For isolated background piece, if in its eight neighborhoods piece be the number of foreground blocks more than or equal to 5, then this background piece is labeled as foreground blocks.

Fingerprint image among the present invention is cut apart Feature Selection: in existing fingerprint image dividing method, mostly utilize half-tone information and texture information during feature selecting, during computation of characteristic values based on piece or based on pixel.Document [Bazen, A., Gerez, S.:Segmentation of fingerprint images.In:Proc.Workshop on Circuits Systems and Signal Processing (ProRISC 2001). (2001) pp.276-280] employing direction consistance (Coherence,), local gray level average (Mean) and gray variance (Variance) (hereinafter to be referred as CMV) be as the parameter of describing each fingerprint pixel feature, make up a linear classifier with this and realize that fingerprint cuts apart.The CMV feature combines half-tone information and texture information, in view of mostly adopting block-based feature in the existing fingerprint image dividing method, so the present invention adopts block-based CMV index to describe fingerprint characteristic.

Whole fingerprint image is resolved into the zero lap image block of W * W size, and for each image block, its CMV index definition is as follows:

A. direction consistance (C) in the piece

The direction consistance of fingerprint image piece has been described the consistent degree of each pixel direction of this piece.If represent the direction consistance of an image block, then with Coh:

$\mathrm{Coh}=\frac{\left|{\mathrm{\Σ}}_W(G_{s,x},G_{s,y})\right|}{{\mathrm{\Σ}}_W|(G_{s,x},G_{s,y}\left)\right|}=\frac{\sqrt{{(G_{\mathrm{xx}}-G_{\mathrm{yy}})}^2+4G_{\mathrm{xy}}^2}}{G_{\mathrm{xx}}+G_{\mathrm{yy}}}---\left(1\right)$

(G wherein _{S, x}, G _{S, y}) be pixel (x, two times of gradients y), $G_{\mathrm{xx}}=\underset{W}{\mathrm{\Σ}}{G}_x^2,$ $G_{\mathrm{yy}}=\underset{W}{\mathrm{\Σ}}{G}_y^2,$ $G_{\mathrm{xy}}=\underset{W}{\mathrm{\Σ}}{G}_x{G}_y,$ (G _x, G _y) be gradient vector, W is the size of piece, this paper default value gets 8 * 8.

B. piece gray average (M)

The gray average of an image block is defined as follows:

$\mathrm{Mean}=\underset{W}{\mathrm{\Σ}}I---\left(2\right)$

Wherein, I is the gray-scale value of each point in the piece, and the W definition is the same.

C. piece gray variance (V)

The gray variance of a piece is defined as follows:

$\mathrm{Var}=\underset{W}{\mathrm{\Σ}}{(I-\mathrm{Mean})}^2---\left(3\right)$

Wherein I, W definition is the same.

After asking for the CMV value of piece,, make it value all between [0,1] to the normalization of feature value.

Cluster is one of important research content of research directions such as data mining, pattern-recognition, has important role aspect the immanent structure of recognition data.Cluster belongs to the unsupervised learning method in essence, cuts apart for fingerprint image, is typical two class clustering problem: prospect and background.

According to data gathering rule and using these regular methods in cluster, clustering algorithm roughly is divided into the stratification clustering algorithm, divides the formula clustering algorithm, based on clustering algorithm and other clustering algorithms of density and grid.Division formula clustering algorithm need be specified clusters number or cluster centre in advance, by the computing that iterates, progressively reduces the error amount of objective function, when target function value is restrained, obtains final cluster result.The K-means method that is proposed by Mac Queen is a kind of classical division formula clustering algorithm that solves the cluster analysis problem, and this algorithm thought is simple, easily realize, and fast convergence rate.If difference obviously between each family, and DATA DISTRIBUTION is dense, then this algorithm is more effective.If but shape of each bunch and size differences are little, then may occur bigger bunch and cut apart phenomenon.

Find that by statistics the foreground blocks of fingerprint image and background piece form tangible two bunches, and DATA DISTRIBUTION is dense, so the present invention selects the K-means method to carry out fingerprint image piece cluster and cut apart.

The fingerprint image of employing K-means cluster thought is cut apart and roughly is divided into following steps, and algorithm flow chart is seen Fig. 1.

1) image block and CMV value are asked for

Fingerprint image is resolved into the zero lap image block of W * W size, for each image block, according to formula (1), (2),

(3) ask for its CMV value, and the CMV of all pieces is carried out normalized.

2) image block cluster

Utilize the K-means method that each image block is carried out cluster according to its CMV eigenwert, the cluster parametric description is: each image block is a sample, and each sample has 3 features, is respectively C, M, V; The classification number is set at 2; 2 image blocks of picked at random are as the initial distance center; Distance calculation between all the other each pieces and 2 the piece centers adopt Euclidean (Euclidean) distance square.

3) determine prospect classification value and background classification value according to cluster result, obtain preliminary segmentation result.

The net result of K-means method is just classified each image block with classification value (1/2), therefore need definite which classification value to represent prospect or background.This paper determines that the principle of prospect and background classification value is: the eight neighborhood pieces of getting the picture centre piece, what certain classification value number was many is foreground blocks classification value, as classification value in the eight neighborhood pieces 6 " 2 " values are arranged, 2 " 1 " value thinks that then classification value " 2 " is a prospect classification value.

4) utilize morphology to cut apart aftertreatment.

Owing to be subjected to noise effect, some isolated background piece and foreground blocks can appear in the preliminary segmentation result, this paper takes following principle to handle: for isolated foreground blocks, if in its eight neighborhoods piece for the number of foreground blocks is less than 4, then this foreground blocks is labeled as the background piece; For isolated background piece, if in its eight neighborhoods piece be the number of foreground blocks more than or equal to 5, then this background piece is labeled as foreground blocks.

The invention has the beneficial effects as follows: solved the equipment interoperability issue that exists in the existing fingerprint image dividing method, this method not only can correctly be cut apart fingerprint image, and avoided traditional algorithm will set different threshold values when cutting apart or trained the defective of different sorters from the fingerprint image of distinct device, accomplished and device independent, very be fit to use the internet, applications environment of polymorphic type fingerprint collecting equipment.

Description of drawings

Fig. 1 is an algorithm flow chart of the present invention.

Embodiment

The invention will be further described below in conjunction with accompanying drawing and embodiment.

A kind of fingerprint image dividing method that has nothing to do with collecting device, its step is:

1) input picture is divided into W * W piece zero lap image block;

2) every is asked the CMV index and carries out normalized;

3) each piece is carried out the K-means cluster;

4) determine the classification of foreground blocks and background piece, obtain preliminary segmentation result;

5) the morphology aftertreatment obtains final segmentation result.

Described step 2) in, described CMV index is also carried out the normalized process and is:

A. direction consistance C in the piece

The direction consistance of fingerprint image piece has been described the consistent degree of each pixel direction of this piece; If represent the direction consistance of an image block, then with Coh:

$\mathrm{Coh}=\frac{\left|{\mathrm{\Σ}}_W(G_{s,x},G_{s,y})\right|}{{\mathrm{\Σ}}_W|(G_{s,x},G_{s,y}\left)\right|}=\frac{\sqrt{{(G_{\mathrm{xx}}-G_{\mathrm{yy}})}^2+4G_{\mathrm{xy}}^2}}{G_{\mathrm{xx}}+G_{\mathrm{yy}}}---\left(1\right)$

(G wherein _{S, x}, G _{S, y}) be pixel (x, two times of gradients y), $G_{\mathrm{xx}}=\underset{W}{\mathrm{\Σ}}{G}_x^2,$ $G_{\mathrm{yy}}=\underset{W}{\mathrm{\Σ}}{G}_y^2,$ $G_{\mathrm{xy}}=\underset{W}{\mathrm{\Σ}}{G}_x{G}_y,$ (G _x, G _y) be gradient vector, W is the size of piece, default value gets 8 * 8;

B. piece gray average M:

The gray average of an image block is defined as follows:

$\mathrm{Mean}=\underset{W}{\mathrm{\Σ}}I---\left(2\right)$

Wherein, I is the gray-scale value of each point in the piece, and the W definition is the same;

C. piece gray variance V

The gray variance of a piece is defined as follows:

$\mathrm{Var}=\underset{W}{\mathrm{\Σ}}{(I-\mathrm{Mean})}^2---\left(3\right)$

Wherein I, W definition is the same;

After asking for the CMV value of piece,, make it value all between [0,1] to the normalization of feature value.

Utilize the K-means method that each image block is carried out cluster according to its CMV eigenwert, the cluster parametric description is: each image block is a sample, and each sample has 3 features, is respectively C, M, V; The classification number is set at 2; 2 image blocks of picked at random are as the initial distance center; Distance calculation between all the other each pieces and 2 the piece centers adopt Euclidean (Euclidean) distance square.

In the described step 4), the net result of K-means method is just classified each image block with classification value (1/2), therefore need definite which classification value to represent prospect or background, the method of determining prospect and background classification value is: the eight neighborhood pieces of getting the picture centre piece, what certain classification value number was many is foreground blocks classification value, as classification value in the eight neighborhood pieces 6 " 2 " values are arranged, 2 " 1 " value thinks that then classification value " 2 " is a prospect classification value.

Utilize morphology to cut apart aftertreatment in the described step 5), owing to be subjected to noise effect, some isolated background piece and foreground blocks can appear in the preliminary segmentation result, should take principle to handle: for isolated foreground blocks, if for the number of foreground blocks is less than 4, then this foreground blocks is labeled as the background piece in its eight neighborhoods piece; For isolated background piece, if in its eight neighborhoods piece be the number of foreground blocks more than or equal to 5, then this background piece is labeled as foreground blocks.

Claims (5)

1. One kind with the irrelevant fingerprint image dividing method of collecting device, it is characterized in that its step is:

   1) input picture is divided into W * W piece zero lap image block;

   2) every is asked the CMV index and carries out normalized;

   3) each piece is carried out the K-means cluster;

   4) determine the classification of foreground blocks and background piece, obtain preliminary segmentation result;

   5) the morphology aftertreatment obtains final segmentation result.
2. 2. the fingerprint image dividing method that as claimed in claim 1 and collecting device are irrelevant is characterized in that described step 2) in, described CMV index is also carried out the normalized process and is:

   A. direction consistance C in the piece

   The direction consistance of fingerprint image piece has been described the consistent degree of each pixel direction of this piece; If represent the direction consistance of an image block, then with Coh:

   $\mathrm{Coh}=\frac{\left|{\mathrm{\Σ}}_W(G_{s,x},G_{s,y})\right|}{{\mathrm{\Σ}}_W\left|(G_{s,x},G_{s,y})\right|}=\frac{\sqrt{{(G_{\mathrm{xx}}-G_{\mathrm{yy}})}^2+{4G}_{\mathrm{xy}}^2}}{G_{\mathrm{xx}}+G_{\mathrm{yy}}}---\left(1\right)$

   (G wherein _{S, x}, G _{S, y}) be pixel (x, two times of gradients y), $G_{\mathrm{xx}}=\underset{W}{\mathrm{\Σ}}{G}_x^2,$ $G_{\mathrm{yy}}=\underset{W}{\mathrm{\Σ}}{G}_y^2,$ $G_{\mathrm{xy}}=\underset{W}{\mathrm{\Σ}}{G}_x{G}_y,$ (G _x, G _y) be gradient vector, W is the size of piece, default value gets 8 * 8;

   B. piece gray average M:

   The gray average of an image block is defined as follows:

   $\mathrm{Mean}=\underset{W}{\mathrm{\Σ}}I---\left(2\right)$

   Wherein, I is the gray-scale value of each point in the piece, and the W definition is the same;

   C. piece gray variance V

   The gray variance of a piece is defined as follows:

   $\mathrm{Var}=\underset{W}{\mathrm{\Σ}}{(I-\mathrm{Mean})}^2---\left(3\right)$

   Wherein I, W definition is the same;

   After asking for the CMV value of piece,, make it value all between [0,1] to the normalization of feature value.
3. 3. the fingerprint image dividing method that has nothing to do with collecting device as claimed in claim 1, it is characterized in that, utilize the K-means method that each image block is carried out cluster according to its CMV eigenwert, the cluster parametric description is: each image block is a sample, each sample has 3 features, is respectively C, M, V; The classification number is set at 2; 2 image blocks of picked at random are as the initial distance center; Distance calculation between all the other each pieces and 2 the piece centers adopt Euclidean (Euclidean) distance square.
4. 4. the fingerprint image dividing method that has nothing to do with collecting device as claimed in claim 1, it is characterized in that, in the described step 4), the net result of K-means method is just classified each image block with classification value (1/2), therefore need definite which classification value to represent prospect or background, the method of determining prospect and background classification value is: the eight neighborhood pieces of getting the picture centre piece, what certain classification value number was many is foreground blocks classification value, as classification value in the eight neighborhood pieces 6 " 2 " values are arranged, 2 " 1 " value thinks that then classification value " 2 " is a prospect classification value.
5. 5. the fingerprint image dividing method that has nothing to do with collecting device as claimed in claim 1, it is characterized in that, utilize morphology to cut apart aftertreatment in the described step 5), owing to be subjected to noise effect, some isolated background piece and foreground blocks can appear in the preliminary segmentation result, should take principle to handle: for isolated foreground blocks, if in its eight neighborhoods piece for the number of foreground blocks is less than 4, then this foreground blocks is labeled as the background piece; For isolated background piece, if in its eight neighborhoods piece be the number of foreground blocks more than or equal to 5, then this background piece is labeled as foreground blocks.

Images