CN102222220B - Method for extracting vein-mode hand back texture - Google Patents

Abstract

The invention aims at providing a method for extracting the vein-mode hand back texture. The method is characterized by comprising the steps of: establishing a local second-order differential structural model, namely a VLSDM model, and adopting a noise filtering method based on multiscale analysis to process filtering response of the multiscale VLSDM model so as to obtain the final vein-mode hand back texture. By the adoption of the method, not only can the vein texture and the background area be distinguished, but also the extracted vein texture maintains the original form of the local curved surface, and simultaneously the direction and the scale information of the vein texture can be obtained.

Description

Method for extracting vein-mode hand back texture

Technical field

What the present invention relates to is a kind of recognition methods of biological characteristic.

Background technology

The identification of human body hand back vein has stability, uniqueness, high antifalsification, untouchable and precision advantages of higher, is the advanced subject of area of pattern recognition.Be with a wide range of applications and economic worth.

The method of extracting at present vein texture mainly contains two kinds:

1, the method for the statistical property based on vein texture area pixel grey value profile: by the statistical property definite area threshold value of vein texture area pixel grey value profile, vein sample image is carried out to binary segmentation, obtain a width and can distinguish the bianry image of vein texture region and background area, as shown in Fig. 1 (A).This method advantage is simple and quick, but higher to the quality requirements of vein sample.

The method of 2, following the tracks of based on vein texture transversal section minimum gradation value point: vein texture transversal section minimum gradation value point is carried out to " repeating linear tracking ", from a certain end points, carrying out transversal section minimum gradation value point follows the tracks of, form transversal section minimum gradation value line, after an end points is followed the tracks of and is finished, from another one end points, carry out same tracking operation, the curve that the transversal section minimum gradation value point finally obtaining forms, by overlapping, forms vein textural characteristics, as shown in Fig. 1 (B).When following the tracks of, write down the number of times that transversal section minimum gradation value point is traversed, to distinguish noise spot and vein point, the probability that noise spot is traversed is put little than vein.The relative first method of the method, its extraction effect has some improvement.

The second-order differential architectural characteristic that does not also have to discuss based on image in vein identification field is extracted the pertinent literature of vein texture, but studies often in field of medical image processing ^[1-3], wherein most methods is with document ^[1]the model proposing is basis, document ^[1]eigenwert based on texture second-order differential structure, propose one blood vessel degree computation model, but while directly utilizing this model to process vein image, there are two problems: 1, insensitive to close grain in vein image; 2, not obvious to the performance of intersecting blood vessels place texture.

Open report related to the present invention has:

[1]A.Frangi,W.Niessen,K.Vincken,and M.Viergever,Multiscale vessel enhancement filtering,Medical Image Computing and Computer-Assisted Interventation in1998,1998,pp.130-137.

[2]C.Ca ero and P.Radeva,Vesselness enhancement diffusion,Pattern Recognition Letters,vol.24,2003,pp.3141-3151.

[3]R.Manniesing,M.Viergever,and W.Niessen,Vessel enhancing diffusion::A scale space representation of vessel structures,Medical Image Analysis,vol.10,2006,pp.815-825.

Summary of the invention

The object of the present invention is to provide and can distinguish vein texture and background area, can access the direction of vein pattern reason and the method for extracting vein-mode hand back texture of yardstick information.

The object of the present invention is achieved like this:

Method for extracting vein-mode hand back texture of the present invention, is characterized in that:

(1) set up the local second-order differential structural model-VLSDM of vein texture model: VLSDM model is comprised of the remarkable second-order differential structure characteristic parameter of the local curve form restricted model of vein texture and vein texture, according to the different feature of the local curve form constraint index value of vein texture, method in conjunction with sampling thheorem and Weighted Fusion obtains merging shape constraining index value computation model-SICM model, choose the maximum principal curvatures of curved surface as the remarkable second-order differential structure characteristic parameter of the vein texture in VLSDM model, the array mode of employing based on maximal possibility estimation model refinement merges SICM model and remarkable second-order differential structure characteristic parameter, thereby form VLSDM model,

(2) adopt the noise elimination method based on multiscale analysis to process multiple dimensioned VLSDM model filtering response, thereby obtain last vein-mode hand back texture.

The present invention can also comprise:

1, the method for building up of described SICM model is: adopt sampling function Shannon to sample to shape indexing value SapeIndex, form neighborhood of pixel points curved surface and sampled targets texture curve form index value shape _dcalculating formula of similarity, adopts the method for Weighted Fusion, by with intersects vein texture and dark ridged vein texture and obtains similarity and merge the last vein shape constraining model of formation.

2, described neighborhood of pixel points curved surface and sampled targets texture curve form index value shape _dcalculating formula of similarity mathematic(al) representation is:

$\mathrm{shapeness}=\left\{\begin{array}{cc}0,& \mathrm{SapeIndex}\≤0\\ \frac{\sin \left(\mathrm{\α}(\mathrm{SapeIndex}-{\mathrm{shape}}_D)\right)}{\mathrm{\α}(\mathrm{SapeIndex}-{\mathrm{shape}}_D)},& \mathrm{SapeIndex}>0\end{array}\right.,$

In formula, α represents the attenuation coefficient of Shannon function, and shapeness represents neighborhood of pixel points curved surface and shape _drepresented texture curve form similarity.

3, the method mathematic(al) representation of described Weighted Fusion is:

vein_shape＝β*ridge+(1-β)*concavity

In formula, vein_shape represents final shape constraining value, and ridge represents the dark rib region sampled value in vein image, and concavity represents the recessed bag shape area sampling value in vein image, and β is combination coefficient.

4, the array mode mathematic(al) representation of described maximal possibility estimation model refinement is:

$\mathrm{vein}\_\mathrm{ness}=(1-\exp (-\frac{\mathrm{vein}\_{\mathrm{<figure-callout\; id="2"\; label="shape"\; filenames="HDA00002823980000012.png"\; state="\{\{state\}\}">shape</figure-callout>}}^2}{{\mathrm{\&sigma;}}_1}))(1-\exp (-\frac{\mathrm{vein}\_{\mathrm{<figure-callout\; id="2"\; label="feature"\; filenames="HDA00002823980000012.png"\; state="\{\{state\}\}">feature</figure-callout>}}^2}{{\mathrm{\&sigma;}}_2})),$

In formula, vein_shape, vein_feature are two eigenvalue λ of curved surface second-order differential structure matrix Hessian Matrix ₁, λ ₂in maximal value, σ ₁and σ ₂represent respectively shape constraining model sensitivity coefficient, structure characteristic parameter sensitivity coefficient, σ ₁estimation formulas as follows:

In formula, m represents shape constraining item (1-exp (vein_shape ²/ σ ₁)) value and 1 approximation quality, namely 0.999 ... 9 number after 9 radix point,

Structure characteristic parameter sensitivity coefficient σ ₂estimation formulas as follows:

${\mathrm{\&sigma;}}_2=\frac{{\mathrm{\&lambda;}}_{\max }^2}{m\ln 10},$

In formula, m represents shape constraining item (1-exp (vein_feature ²σ ₂)) value and 1 approximation quality, λ _maxfor the maximum value of image medium sized vein region vein_feature,

Array mode to maximal possibility estimation model refinement is further improved, and the computing formula of the vein degree vein_ness ' after improvement is:

$\mathrm{vein}\_{\mathrm{ness}}^{\text{'}}=\sqrt{{\mathrm{\&sigma;}}_2\&times;\ln \left(\frac{1}{1-\mathrm{vein}\_\mathrm{ness}}\right)}.$

5, the mathematic(al) representation of described VLSDM model is:

$\mathrm{vein}(i,j)=\underset{s_{\min }\&le;s\&le;s_{\max }}{\max }\mathrm{vein}\_\mathrm{ness}(s,i,j),$

The transverse and longitudinal coordinate figure of pixel in i, j presentation video in formula, vein (i, j) is illustrated in the texture vein degree that pixel (i, j) obtains by multiple dimensioned VLSDM model, s _minrepresent minimum differential yardstick, s _maxrepresent maximum differential yardstick, the vein degree of pixel (i, j) when vein_ness in formula (s, i, j) expression differential yardstick is s.

6, the described noise elimination method based on multiscale analysis is:

(1) eliminate small scale glitch noise and speckle noise: make differential yardstick gray-scale map I corresponding to maximum vein degree in differential metric space _sin background area and differential yardstick be s _maxregion be white, differential yardstick is less than s _maxregion be black, form bianry image I _b, glitch noise region and thin vein region are all at I _bin become the speckle regions of black, add up the region maximum value in multiple dimensioned VLSDM model response gray-scale map corresponding to each speckle regions, if its value is greater than the threshold value T that thin vein connected domain and multiple dimensioned VLSDM model corresponding in noise connected domain respond gray-scale value maximum value _max, this speckle regions is filled by white, the image I after speckle regions is filled filtering processing ' _bwith original image I _sdo and operation, obtain only comprising the image I of vein texture region and large scale noise region ' _s;

(2) eliminate large scale noise: make image I ' _smiddle background area is black, other regions are white, large scale noise has just become black perforated and white dot region, by connected domain area threshold, carry out this class large scale noise of elimination, the image medium sized vein unity and coherence in writing region of removing after large scale noise is white portion, background area is black, the large scale that is eliminated noise filtering template;

(3): multiple dimensioned VLSDM model response is carried out to filtering, multiply operation is done in the Filtering Template that utilization obtains and the response of multiple dimensioned VLSDM model, noise region in multiple dimensioned VLSDM model response all becomes null value, venosomes vein degree value remains unchanged, thereby obtains filtered multiple dimensioned VLSDM model filtering response.

Advantage of the present invention is: the present invention not only can distinguish vein texture and background area, and the vein texture extracting kept the original form of its local curved surface, can access direction and the yardstick information of vein pattern reason simultaneously.

Accompanying drawing explanation

Fig. 1 is λ ₁, λ ₂while getting different value, function f (x, y)=(λ ₁x ²+ λ ₂y ²) at (0,0) some place neighborhood curve form three-dimensional plot;

Fig. 2 is SapeIndex (λ ₁, λ ₂, p _o) with λ ₁, λ ₂the chromatogram changing;

Fig. 3 is curved surface minimum curvature direction

distribution situation at two-dimentional vein texture region;

Fig. 4 is hand back vein zones of different shape indexing value distribution situation;

Fig. 5 is β while getting different value, and vein_shape responds gray-scale map;

Fig. 6 is four kinds of structure characteristic parameter response gray-scale maps of same sample;

Fig. 7 is that differential yardstick is respectively 4,5 o'clock, two groups of four kinds of structure characteristic parameters response gray-scale maps of same sample image;

Fig. 8 is four kinds of different structure characteristic quantity response gray-scale maps that strengthen sample image;

Fig. 9 is σ ₁while getting different value, shape constraining item response gray-scale map;

Figure 10 is σ ₂while getting different value, structure characteristic parameter bound term response gray-scale map;

Figure 11 is that differential yardstick is 3 o'clock, vein degree model response diagram;

Figure 12 is the response gray-scale map that Figure 11 (A) obtains after formula (3.16) conversion;

Figure 13 improves the response gray-scale map of array mode under different differential yardsticks;

Figure 14 is multiple dimensioned VLSDM model response noises specificity analysis, and wherein A is multiple dimensioned model response, and B is differential yardstick yardstick response diagram, and C is the histogram curve of figure A;

Figure 15 is the post-processed process illustration of multiple dimensioned VLSDM model response;

Figure 16 is the dissimilar vein sample texture of multiple dimensioned VLSDM model extraction result figure;

Figure 17 is process flow diagram of the present invention.

Embodiment

Below in conjunction with accompanying drawing, for example the present invention is described in more detail:

In conjunction with Fig. 1～17, method for extracting vein-mode hand back texture of the present invention, comprises the foundation of the local second-order differential structural model of vein texture (VLSDM model) and based on multiple dimensioned VLSDM model, vein texture is extracted to two parts.VLSDM model is comprised of the remarkable second-order differential structure characteristic parameter of the local curve form restricted model of vein texture and vein texture, first, according to the different feature of the local curve form constraint index value of vein texture, method in conjunction with sampling thheorem and Weighted Fusion obtains merging shape constraining index value computation model (SICM model), and the shape indexing value that SICM model can make to input sample medium sized vein texture region and the texture zone of intersection is less compared with analog value large and background area; Then, according to the maximum principal curvatures of the curved surface feature stronger to the descriptive power of vein texture, choose the maximum principal curvatures of curved surface as the remarkable second-order differential structure characteristic parameter of the vein texture in VLSDM model; Finally, based on can make filter response medium sized vein region and the difference between background area of VLSDM model can reduce the criterion of vein texture Character losing as far as possible greatly and simultaneously, the array mode of employing based on maximal possibility estimation model refinement merges SICM model and remarkable second-order differential structure characteristic parameter, thereby forms VLSDM model.The multiple dimensioned characteristic that the present invention is intrinsic according to vein texture, right vein texture is analyzed to adopt multiple dimensioned VLSDM model, make full use of noise behavior in multiple dimensioned VLSDM model filtering response simultaneously, the noise elimination method of employing based on multiscale analysis processed multiple dimensioned VLSDM model filtering response, thereby obtains last vein-mode hand back texture.

In conjunction with the method for sampling thheorem and Weighted Fusion, obtaining merging SICM model is: first, adopt sampling function-Shannon function to sample to shape indexing value SapeIndex, form neighborhood of pixel points curved surface and sampled targets texture curve form index value shape _dcalculating formula of similarity; Then, adopt the method for Weighted Fusion, by with intersects vein texture and dark ridged vein texture and obtains similarity and merge the last vein shape constraining model of formation.

Neighborhood of pixel points curved surface and sampled targets texture curve form index value shape _dcalculating formula of similarity mathematic(al) representation is as follows:

$\mathrm{shapeness}=\left\{\begin{array}{cc}0,& \mathrm{SapeIndex}\&le;0\\ \frac{\sin \left(\mathrm{\&alpha;}(\mathrm{SapeIndex}-{\mathrm{shape}}_D)\right)}{\mathrm{\&alpha;}(\mathrm{SapeIndex}-{\mathrm{shape}}_D)},& \mathrm{SapeIndex}>0\end{array}\right.$

In formula, SapeIndex represents shape indexing value, shape _dindicate the to sample ideal form index value of texture shape, for the back of the hand, α represents the attenuation coefficient of Shannonh function, shapeness represents neighborhood of pixel points curved surface and shape _drepresented texture curve form similarity, requires SapeIndex>0 to be because the texture of two types of required sampling is for intersecting vein texture and dark ridged vein texture, and their ideal form index value is respectively 0.5 and 1, is all greater than 0.

The attenuation coefficient α of Shannonh function is: itself and ideal value shape _dneighborhood length of an interval degree is relevant, and computation rule is to guarantee that shapeness value corresponding to the interior shape indexing value of neighborhood is in [0.9,1], and with the estimation of sampling function: sin (α * r)/(α * r)=0.9, r represents shape _dthe interval radius of neighborhood, because the span of SapeIndex is [1,1], so the span of r is [0,1], for normal vein texture region, we get r=0.2, so by estimating that the value that can obtain α is about 4, for intersecting unity and coherence in writing region, owing to occurring that the interval of maximum value is less, so we get r=0.07, the value that calculates α is approximately 10.

The method of Weighted Fusion is: its mathematic(al) representation is as follows:

vein_shape＝β*ridge+(1-β)*concavity

In formula, vein_shape represents final shape constraining value; Ridge represents the dark rib region sampled value in vein image, by α=4, shape _d=0.5 substitution texture similarity shapeness computing formula obtains; Concavity represents the recessed bag shape area sampling value in vein image, now by α=10, shape _d=1 substitution texture similarity shapeness computing formula obtains; β is combination coefficient, and its value is taken as at 0.8 o'clock, and syncretizing effect is best, and when now can guarantee to reduce in image dark speckle regions, the vein_shape value of vein texture area is also comparatively approaching.

VLSDM model is: it adopts the array mode based on maximal possibility estimation model refinement that SICM model is merged and formed with remarkable second-order differential structure characteristic parameter, and the array mode mathematic(al) representation of maximal possibility estimation model refinement is as follows:

$\mathrm{vein}\_\mathrm{ness}=(1-\exp (-\frac{\mathrm{vein}\_{\mathrm{<figure-callout\; id="2"\; label="shape"\; filenames="HDA00002823980000012.png"\; state="\{\{state\}\}">shape</figure-callout>}}^2}{{\mathrm{\&sigma;}}_1}))(1-\exp (-\frac{\mathrm{vein}\_{\mathrm{<figure-callout\; id="2"\; label="feature"\; filenames="HDA00002823980000012.png"\; state="\{\{state\}\}">feature</figure-callout>}}^2}{{\mathrm{\&sigma;}}_2}))$

In formula, vein_shape, vein_feature are two eigenvalue λ of curved surface second-order differential structure matrix Hessian Matrix ₁, λ ₂in maximal value, if gained vein_feature value is less than 0, make vein_feature equal 0.σ ₁and σ ₂represent respectively shape constraining model sensitivity coefficient, structure characteristic parameter sensitivity coefficient, σ ₁estimation formulas as follows:

In formula, m represents shape constraining item (1-exp (vein_shape ²/ σ ₁)) value and 1 approximation quality, namely 0.999 ... 9 number after 9 radix point.Generally get m=3 and just can meet the demands, by T _shape=0.5 substitution above formula, can obtain σ ₁estimated value be approximately 0.036.

Structure characteristic parameter sensitivity coefficient σ ₂estimation formulas as follows:

${\mathrm{\&sigma;}}_2=\frac{{\mathrm{\&lambda;}}_{\max }^2}{m\ln 10}$

In formula, m represents shape constraining item (1-exp (vein_feature ²σ ₂)) value and 1 approximation quality, namely 0.999 ... 9 number after 9 radix point.Generally getting m=3 just can meet the demands; λ _maxmaximum value for image medium sized vein region vein_feature.

In order to guarantee that the maximum principal curvatures of curved surface, to the good descriptive power of vein texture shape, further improves the array mode of maximal possibility estimation model refinement, the computing formula of the vein degree vein_ness ' after improvement is as follows:

$\mathrm{vein}\_{\mathrm{ness}}^{\text{'}}=\sqrt{{\mathrm{\&sigma;}}_2\&times;\ln \left(\frac{1}{1-\mathrm{vein}\_\mathrm{ness}}\right)}$

Multiple dimensioned VLSDM model is: while adopting VLSDM model to process vein image, in the response of gained texture form, thin vein texture can disappear along with the increase of filtering differential yardstick, its reason is: thin vein texture exists the feature that width is little, contrast is low, this has determined that its local curved-surface shape itself is obvious not, after the Gaussian filter smoothing processing of large scale, make the characteristic parameter of its local curve form less, thereby cause its vein degree to diminish, so the response of thin vein texture form can diminish when differential yardstick becomes large.Only have when filter window width is close with its texture width, now can regard gaussian filtering as matched filter, because Gauss's window Two-dimensional Surfaces shape is similar to vein texture curve form, therefore can reach a larger value, this has also illustrated a bit: when differential yardstick is large from little change, can reach maximal value at a certain yardstick; This point sets out, and the vein degree of corresponding pixel points during by more different differential yardstick is got its maximal value as last vein degree, thereby it is as follows to obtain the mathematic(al) representation of multiple dimensioned VLSDM model:

$\mathrm{vein}(i,j)=\underset{s_{\min }\&le;s\&le;s_{\max }}{\max }\mathrm{vein}\_\mathrm{ness}(s,i,j)$

In formula, i, the transverse and longitudinal coordinate figure of pixel in j presentation video, vein (i, j) is illustrated in the texture vein degree (also can be described as the maximum vein degree in differential metric space) that pixel (i, j) obtains by multiple dimensioned VLSDM model, s _minrepresent minimum differential yardstick, s _maxrepresent maximum differential yardstick, when veinness in formula (s, i, j) represents that differential yardstick is s, the vein degree of pixel (i, j).

Noise elimination method based on multiscale analysis is: the key of the multiple dimensioned VLSDM model response of elimination is thin vein region and the noise region of distinguishing in response, and the feature that noise is communicated with district is: vein degree corresponding in region is generally less; The feature that thin vein is communicated with district is: center vein degree is larger, and frontier district is less, especially with other vein textures Fang Geng little across; Although the multiple dimensioned VLSDM model response that some pixel of thin vein connected region and noise region is corresponding equates, but in these two kinds of connected regions, the maximum value difference of corresponding multiple dimensioned VLSDM model response is larger, can they be distinguished by threshold value, distinguish gray-scale value corresponding to threshold value that thin vein connected domain and multiple dimensioned VLSDM model corresponding in noise connected domain respond gray-scale value maximum value and histogram extreme value relevant, if use T _maxrepresent this threshold value, use G _h-maxthe gray-scale value that represents the multiple dimensioned VLSDM model response that Histogram Maximum point is corresponding, by hand back vein sample is analyzed, multiple dimensioned VLSDM model response is carried out after 256 grades of gray scale normalizations, and dark noise is communicated with district's response gray-scale value mostly and G _h-maxapproach, the peak response gray-scale value in bright noise connected region is generally G _h-max2 ~ 3 times large, the peak response gray-scale value that thin vein is communicated with district generally can reach G _h-max6 ~ 8 times, for guaranteeing to filter noise, do not affect thin vein texture simultaneously, can choose T _max=4G _h-max, use I _vrepresent that the response of multiple dimensioned VLSDM model carries out 256 grades of images after gray scale normalization, use I _srepresent differential yardstick gray-scale map corresponding to maximum vein degree in differential metric space, the noise elimination method concrete steps based on multiscale analysis are as follows:

The first step: eliminate small scale glitch noise and speckle noise, concrete grammar is to allow I _sin background area and differential yardstick be s _maxregion be white, differential yardstick is less than s _maxregion be black, form bianry image I _b, the glitch noise region in Fig. 3 .15 and thin vein region are all at I like this _bin become the speckle regions of black, add up the region maximum value in multiple dimensioned VLSDM model response gray-scale map corresponding to each speckle regions, if its value is greater than threshold value T _max, this speckle regions is filled by white.Image I after speckle regions is filled filtering processing ' _bwith original image I _sdo and operation, finally just obtained only comprising the image I of vein texture region and large scale noise region ' _s.

Second step: eliminate large scale noise, concrete grammar is, allow image I ' _smiddle background area is black, other regions are white, large scale noise has just become black perforated and white dot region like this, black hole noise brings because the thick vein of intersection sticks together, it is larger that speckle noise is that the maximum vein degree in differential metric space appears at the ，Qie white dot district gray-scale value that the Noise texture in large scale causes.The area of large scale white dot noise region and black hole noise region is generally less, therefore can carry out this class large scale noise of elimination by connected domain area threshold, area threshold for white dot noise can be selected larger value, because now the connection area of venosomes is generally larger, for ROI, carry out the hand back vein sample after 256 * 256 size normalization, threshold value can be taken as 100, area is less than the white connected domain of this value and thinks noise, and the pixel value of this connected domain is set to 0; The threshold area in black cavity can not be got too large, because may being also many vein crossings, black cavity forms, for ROI, carry out the hand back vein sample after 256 * 256 size normalization, the threshold area of black hole can be taken as 30, think that the black hole that is less than this value is noise, the pixel value of noise perforated is set to 1.The image medium sized vein unity and coherence in writing region of removing after large scale noise is white portion (pixel value is 0), and background area is black (pixel value is 0), thereby can be eliminated large scale noise filtering template.

The 3rd step: multiple dimensioned VLSDM model response is carried out to filtering, utilize Filtering Template and the response of multiple dimensioned VLSDM model that second step obtains to do multiply operation, noise region in multiple dimensioned like this VLSDM model response just all becomes null value, and venosomes vein degree value remains unchanged, thereby obtain filtered multiple dimensioned VLSDM model filtering response.

Make a concrete analysis of below:

1. vein texture Second Order Partial differential structrue characteristic

Certain 1 p in analysis image I (p) ₀neighborhood in the gray-scale value I (p of other pixels _o+ δ p _o) and p _opoint is during being related to of gray-scale value, a kind of conventional method is, by image I (p) at a p _oplace carries out Taylor expansion:

$I(p_o+\mathrm{\&delta;}{p}_o,s)\&ap;I(p_o,s)+\mathrm{\&delta;}{p}_o^T{\&dtri;}_{o.s}+\mathrm{\&delta;}{p}_o^T{H}_{o.s}\mathrm{\&delta;}{p}_o---\left(1\right)$

In formula represent p _othe gradient vector of point, H _o,srepresent p _othe Hassian matrix (Hessian Matrix) of point, second-order differential structure or shape gating matrix also referred to as image, s represents differential yardstick, according to linear-scale Space Theory by the thought with different scale Gauss partial differential core convolutional calculation Image Multiscale partial differential

and H _o,scomputing formula as follows:

${\&dtri;}_{o,s}=(s^r\&CenterDot;I\left(p\right)*\frac{\&PartialD;}{\&PartialD;x}G(p,s),s^r\&CenterDot;I\left(p\right)*\frac{\&PartialD;}{\&PartialD;y}G(p,s),){|}_{p=p_o}$

$H_{o,s}=s^{\mathrm{\&gamma;}}\&times;\left[\begin{array}{cc}I\left(p\right)*\frac{{\&PartialD;}^{2}G(p,s)}{{\&PartialD;x}^2}& I\left(p\right)*\frac{{\&PartialD;}^{2}G(p,s)}{\&PartialD;x\&PartialD;y}\\ I\left(p\right)*\frac{{\&PartialD;}^{2}G(p,s)}{\&PartialD;x\&PartialD;y}& I\left(p\right)*\frac{{\&PartialD;}^{2}G(p,s)}{{\&PartialD;\mathrm{<figure-callout\; id="2"\; label="y"\; filenames="HDA00002823980000012.png"\; state="\{\{state\}\}">y</figure-callout>}}^2}\end{array}\right]{|}_{p=p_o---\left(2\right)}$

In formula, r, γ are the differential calculus ruler degree normalization factor, and it equals the exponent number of required partial differential, therefore r=1 here,

γ=2, G (p, s) represents two-dimensional Gaussian function, expression formula is as follows:

$G(p,s)=\frac{1}{2\mathrm{\&pi;}{s}^2}\exp (-\frac{\left|\right|p\left|\right|}{2s^2})---\left(3\right)$

1.1 Hassian matrix H _o,sthe characteristic of eigenwert aspect the local curve form of sign vein

Use λ ₁, λ ₂represent respectively p _othe Hassian matrix H of point _o,stwo eigenwerts, λ ₁, λ ₂characterized p ₀the very important geometric properties information of neighborhood curved surface.Function f (x, y)=(λ ₁x ²+ λ ₂y ²) (0,0) vertex neighborhood curved surface at λ ₁, λ ₂geometric configuration under different value condition, as shown in Figure 1.

The relation of the local set of the eigenwert function shape of Hassian matrix, Koenderink is described below with formula:

$\mathrm{SapeIndex}({\mathrm{\&lambda;}}_1,{\mathrm{\&lambda;}}_2,p_o)=\frac{2}{\mathrm{\&pi;}}\arctan \left(\frac{{\mathrm{\&lambda;}}_1+{\mathrm{\&lambda;}}_2}{|{\mathrm{\&lambda;}}_1-{\mathrm{\&lambda;}}_2|}\right)---\left(4\right)$

In formula, SapeIndex (λ ₁, λ ₂, p _o) p in presentation video _overtex neighborhood curve form index value.

Fig. 2 (A) is depicted as SapeIndex (λ ₁, λ ₂, p _o) with λ ₁, λ ₂the chromatogram changing, can find out shape indexing value SapeIndex (λ ₁, λ ₂, p _o) by λ ₁, λ ₂various combination be divided into many different colour bands, the colour band that in figure, color is identical represents identical local curve form, colour band color difference has represented different local curve forms; When Fig. 2 (C) is differential yardstick s=3, sample image shape indexing value response gray scale, can find out that the shape indexing value response gray-scale value of venosomes curved surface is more approaching, and the local curve form of figure venosomes is more consistent.

1.2 Hassian matrix H _o,sproper vector in the characteristic representing aspect vein grain direction

To λ ₁, λ ₂relation stipulate: | λ ₁|>=| λ ₂|, use v ₁, v ₂represent respectively λ ₁, λ ₂characteristic of correspondence vector, has following characteristic: | λ ₁| and

represent respectively p ₀point curved surface maximum curvature and direction; | λ ₂| and represent respectively p _opoint curved surface minimum curvature and direction, and

orthogonal, when Fig. 3 is differential yardstick s=3,

distribution situation figure in vein sample image, can see, local vein texture region (rectangle frame B2 region in Fig. 3)

direction distributes more consistent, and can roughly reflect the trend of vein texture, at vein crossings place (rectangle frame B1 region in Fig. 3)

distribution is the multi-direction distribution of scattering, all directions more consistent.

2. vein texture local shape restricted model

Fig. 4 (A) is depicted as the chromatogram of Fig. 4 (B); Fig. 4 (B) is depicted as the pixel shape indexing value distribution curve on the different vein texture region straight lines in the left side and two, the right; Fig. 4 (C) is depicted as thin vein region (Fig. 4 (A) upper left side rectangle frame marks) partial pixel and puts concrete shape indexing value; Fig. 4 (D) is depicted as vein image background area (in Fig. 4 (A), below rectangle frame marks) partial pixel and puts concrete shape indexing value.Fig. 4 (E) and Fig. 4 (F) are depicted as the vein texture curve form index value of two vein crossings regions (two, upper right side of Fig. 4 (A) rectangle frame marks) partial pixel point.

The 4th suite face in Fig. 1, what it represented is desirable dark ridged texture curved surface, now two of center of surface point Hassian matrix eigenwert features are: one of them is larger, another is zero, and be non-negative, its described implication is that local curved surface changes greatly along maximum curvature direction curved surface, and level and smooth along minimum curvature direction curved surface; Hand back vein texture is dark ridged curved surface (shown in Fig. 3 (A)) just, so the eigenwert state of vein texture partial points Hassian matrix also should meet the characteristic of two eigenwerts of the 4th suite face central point Hassian matrix in Fig. 1, that is: λ ₁be worth larger, and λ ₁> 0, λ ₁> > λ ₂, by λ ₁, λ ₂this relation generation in formula (3.4), have:

$\mathrm{SapeIndex}({\mathrm{\&lambda;}}_1,{\mathrm{\&lambda;}}_2,p_o)=\frac{2}{\mathrm{\&pi;}}\arctan \left(\frac{{\mathrm{\&lambda;}}_1+{\mathrm{\&lambda;}}_2}{|{\mathrm{\&lambda;}}_1-{\mathrm{\&lambda;}}_2|}\right)\&ap;\frac{2}{\mathrm{\&pi;}}\arctan \left(\frac{{\mathrm{\&lambda;}}_1}{{\mathrm{\&lambda;}}_1}\right)=0.5---\left(5\right)$

The shape indexing value obtaining in above formula is ideal value, and in fact local vein texture is not all desirable ridged texture, so the shape indexing value of venosomes curved surface is concentrated in 0.5 neighborhood, but neighborhood is interval little.

From Fig. 4 (A), the shape indexing value chromatogram of background area can be found out, the shape indexing value of the most pixels in background area is negative value, concrete numerical value from Fig. 4 (D), they are distributed in-0.5 around, by the method for analyzing venosomes texture curve form, can know that the curved surface of background area is approximately bright ridged texture; The concrete distribution situation that is depicted as the shape indexing value of two vein texture infalls from Fig. 4 (E) and 4 (F) can be found out, there is local maximum in the shape indexing value in this region, and maximum value has arrived more than 0.9, approach 1, the curve form in this explanation vein crossings region is approximately recessed bag shape (the blackening point curved surface in Fig. 3 .2 in second group of image).

Shape indexing value response to vein image during mode that the present invention takes is sampled, and near the pixel that shape value is distributed in 0.5 and 1 retains, and other parts are removed.The present invention has selected a kind of conventional sampling function-Shannon function, and for without loss of generality, formula is expressed as follows:

$\mathrm{shapeness}=\left\{\begin{array}{cc}0,& \mathrm{SapeIndex}\&le;0\\ \frac{\sin \left(\mathrm{\&alpha;}(\mathrm{SapeIndex}-{\mathrm{shape}}_D)\right)}{\mathrm{\&alpha;}(\mathrm{SapeIndex}-{\mathrm{shape}}_D)},& \mathrm{SapeIndex}>0\end{array}\right.---\left(6\right)$

In formula, SapeIndex represents shape indexing value, shape _dindicate the to sample ideal form index value of texture shape, α represents the attenuation coefficient of Shannonh function, shapeness represents sampled point neighborhood curved surface and shape _drepresented texture curve form similarity, requires SapeIndex>0.

Shape indexing value SapeIndex is by calculating the eigenwert substitution formula (4) of sampled point to obtain; Shape _dvalue can in Fig. 1, calculate estimated value by the curved surface features value substitution formula (4) of corresponding types, for most of vein textures, shape _d=0.5; Concerning intersecting vein texture, the value of calculating is 1, but 1 be the boundary value of shape indexing value SapeIndex interval [1,1], so select shape _d=0.95; The attenuation coefficient α of Shannonh function and ideal value shape _dneighborhood length of an interval degree is relevant, and computation rule is to guarantee that shapeness value corresponding to the interior shape indexing value of neighborhood is in [0.9,1], and with the estimation of sampling function: sin (α * r)/(α * r)=0.9, r represents shape _dthe interval radius of neighborhood, because the span of SapeIndex is [1,1], so the span of r is [0,1], for normal vein texture region, get r=0.2, so by estimating that the value that can obtain α is about 4, for intersecting unity and coherence in writing region, owing to occurring that the interval of maximum value is less, so we get r=0.07, the value that calculates α is approximately 10.

To intersect the sampled value of vein texture and dark ridged vein texture in conjunction with forming the last about mould of vein shape

Type, formula is expressed as follows:

vein_shape＝β*ridge+(1-β)*concavity (7)

In formula, ridge represents the dark rib region sampled value in vein image, by α=4, shape _d=0.5 substitution formula (6) calculates; Concavity represents the recessed bag shape area sampling value in vein image, now α=10, shape _d=1; β is combination coefficient; Vein_shape represents final shape constraining value.

Figure 5 shows that differential yardstick is 3, when β gets different value, the response image of vein_shape, can see along with β increases, little blackening region in figure tails off, in β=0.8, some effect is best, now affects in image dark speckle regions seldom, and the shape constraining value of vein texture area is also more consistent simultaneously, when β is less than 0.8, now caused the dark speckle regions in image to become bright speckle regions, can affect equally the effect of model, so in the present invention, get β=0.8.

3. the remarkable second-order differential structure characteristic parameter of vein texture

Curved surface maximum curvature is | λ ₁|, the formula of average curvature (meanness), degree of structuration (structureness), maximum principal curvatures (maximumprincipalcurvature) is described below:

mean＝(λ ₁+λ ₂)/2

$\mathrm{Cur}=\sqrt{({\mathrm{\&lambda;}}_1^2+{\mathrm{\&lambda;}}_2^2)/2}---\left(8\right)$

λ＝max(λ ₁,λ ₂）

In formula, mean, Cur, λ represent respectively average curvature, degree of structuration, maximum principal curvatures.

3.1 distinguish the ability comparison of background and vein texture

As can be seen from Figure 6, maximum principal curvatures response strengthens effect to vein texture will get well, vein texture the smooth of the edge, and thin vein clean mark, but there is Noise texture in background area, the contrast still and between vein texture is higher.Compare, the ability that background and vein texture are distinguished in maximum principal curvatures response is stronger than other two kinds of modes.

3.2 are subject to the impact of differential dimensional variation

As can be seen from Figure 7 along with the increase of differential yardstick, the texture strengthening becomes more and more level and smooth, and vein texture region has the situation of chap, and the maximum principal curvatures response of this trend shows more obviously; The thick vein texture of curved surface maximum curvature response becomes more clear, but thin vein texture become fuzzyyer (in Fig. 7 (A1), rectangle frame marks) in its response; There is serious adhesion phenomenon in the response of average curvature, the problem (in Fig. 7 (D1), rectangle frame marks) that thin vein disappears has appearred in degree of structuration response; Compare other three kinds of responses, although the response of maximum principal curvatures has the trend of chap, its background area becomes more level and smooth, and still can keep vein texture clear in structure.

The susceptibility of 3.3 pairs of sample contrasts

Figure 8 shows that sample image in Fig. 6 (A), after strengthening and processing, is at differential yardstick at 3 o'clock, four kinds of different structure characteristic quantities response gray-scale maps.Can find out, the enhancing of sample image is processed the quality of four kinds of different structure characteristic quantities responses is all made moderate progress, especially maximum principal curvatures response, and the Noise texture of its background area becomes seldom, almost surplus vein texture only; But strengthen to process the vein the bringing problem (region that in Fig. 8 (A), empty rectangle frame marks) that attenuates, in other three kinds of responses, all respond, wherein larger on degree of structuration response and curved surface maximum curvature response impact, there is phenomenon of rupture, compared maximum principal curvatures response and be subject to this impact minimum.

To sum up, maximum principal curvatures is stronger to the descriptive power of vein texture, can guarantee preferably the integrality of vein texture structure.Therefore the present invention selects maximum principal curvatures to be used as the remarkable second-order differential structure characteristic parameter of vein texture in VLSDM model, and the structure characteristic parameter in last model is formulated as follows:

$\mathrm{vein}\_\mathrm{featrue}=\left\{\begin{array}{cc}0,& \mathrm{\&lambda;}\&le;0\\ \mathrm{\&lambda;},& \mathrm{\&lambda;}>0\end{array}\right.---\left(9\right)$

In formula, vein_feature represents VLSDM model medium sized vein texture and structural characteristic amount, and λ is the same with meaning in formula (8), represents the maximum principal curvatures of pixel, restrictive condition λ >0 is in order further to reduce the impact of ground unrest in quiet image.

4. setting up VLSDM model and parameter selects

4.1 array modes based on maximal possibility estimation model

About binding occurrence vein_shape is larger for vein shape, and vein degree is also larger.Vein degree (vein_ness) computing formula is as follows:

$\mathrm{vein}\_\mathrm{ness}=(1-\exp (-\frac{\mathrm{vein}\_{\mathrm{<figure-callout\; id="2"\; label="shape"\; filenames="HDA00002823980000012.png"\; state="\{\{state\}\}">shape</figure-callout>}}^2}{{\mathrm{\&sigma;}}_1}))(1-\exp (-\frac{\mathrm{vein}\_{\mathrm{<figure-callout\; id="2"\; label="feature"\; filenames="HDA00002823980000012.png"\; state="\{\{state\}\}">feature</figure-callout>}}^2}{{\mathrm{\&sigma;}}_2}))---\left(10\right)$

In formula, σ ₁and σ ₂represent respectively shape constraining model sensitivity coefficient, structure characteristic parameter sensitivity coefficient, the meaning of vein_shape, vein_feature is identical with formula (7), formula (9) respectively.

Sensitivity coefficient σ is discussed below ₁, σ ₂method of estimation.

(1) select sensitivity coefficient σ ₁

σ ₁determine the influence degree of shape constraining item in model, because the shape of vein texture is comparatively close, caused the vein_shape value of vein texture area also comparatively approaching, so the shape constraining item (1-exp (vein_shape in vein degree computing formula ²/ σ ₁)) numerical values recited little on the impact of vein degree, in shape constraining item can being regarded as like this, be the effect of mask, be mainly responsible for the Noise texture in elimination vein image background, guarantee the continuity of vein texture area simultaneously, from this point of view, we can select a threshold value T _shape, as vein_shape>=T _shapetime, the value of the shape constraining item in vein degree computing formula is approximately 1, represents that current pixel neighborhood of a point curve form is that the confidence level of vein texture approaches 1, works as vein_shape<T _shapetime, the shape constraining item in vein degree computing formula is less than 1, and there are two kinds of possibilities in corresponding point region now: 1, Noise texture region; 2, because causing vein texture, the back of the hand extended configuration produces the region of distortion.In order to guarantee that breakpoint, T do not appear in vein texture area _shapevalue can not be too little, otherwise may make to be out of shape texture region (1-exp (vein_shape ²/ σ ₁)) value be zero, so the intermediate value of getting the interval [0,1] of vein_shape in the present invention, i.e. T _shape=0.5, σ like this ₁estimation formulas can be expressed as follows:

In formula, m represents shape constraining item (1-exp (vein_shape ²/ σ ₁)) value and 1 approximation quality, namely 0.999 ... 9 number after 9 radix point.The present invention gets m=3, by T _shape=0.5 substitution above formula, can obtain σ ₁estimated value be approximately 0.036.

Figure 9 shows that differential yardstick s=3, σ ₁get respectively shape constraining item (1-exp (vein_shape in model at 100,50,15,5,0.5,0.05,0.005,0.0005 o'clock ²/ σ ₁)) response gray-scale map.

From Fig. 3 .10, can find out, along with σ ₁reduce, in model, the more value added convergence of the value of shape constraining item is consistent, works as σ ₁the shape constraining value in=0.05 Shi， vein district is very approaching, and phenomenon of rupture also reduced, and this is very approaching with estimated value 0.036 above, and this illustrates that this estimated value is comparatively reasonable.

(2) select sensitivity coefficient σ ₂

Structure characteristic parameter bound term (1-exp (vein_feature in vein degree computing formula ²σ ₂)) main topological structure and the local curved surface change shape information of being responsible for reaction vein texture, they are all textural characteristics that vein recognition system is played an important role, σ ₂determined the influence degree of structure bound term in model, so σ ₂choose very important.

Figure 10 shows that differential yardstick s=3, σ ₂get respectively 50,15,5,0.5,0.05 o'clock, in vein degree computing formula, structure characteristic parameter bound term (1-exp (vein_feature ²/ σ ₂)) response gray-scale map.The σ that requirement is chosen ₂value should guarantee the contrast of vein texture region and background area, can not occur white area simultaneously, estimates σ ₂method and σ ₁similar, collateral security response image does not occur that white area sets about, the structure characteristic parameter bound term (1-exp (vein_feature from vein degree formula ²σ ₂)), the region that occurs at first white area is the maximal value region of vein_feature, and vein_feature is relevant with the size of maximum principal curvatures λ, and the maximum value of note image medium sized vein region λ is λ _max, σ ₂estimated value, can be formulated as follows:

${\mathrm{\&sigma;}}_2=\frac{{\mathrm{\&lambda;}}_{\max }^2}{m\ln 10}---\left(12\right)$

In formula, the meaning of m is identical with formula (11).Estimate σ ₂, σ ₁after, formula (10) is just determined.

(3) vein degree model response

Figure 11 is that vein degree model is the response gray-scale map of 3 o'clock at differential yardstick.As shown in rectangle frame in figure, thin vein texture contrast is inadequate.In principal curvatures response, thin vein texture contrast is fine, and this is because in vein degree computing formula: work as σ ₁=0.036 o'clock, in formula (10), the numerical value of shape constraining item levels off to 1 and 0, the value of vein_ness is divided into two parts, a part is 0, another part be on the occasion of, vein_ness be on the occasion of time, the numerical value of corresponding shape constraining item levels off to 1, therefore the value of vein_ness is approximately the value of structure characteristic parameter bound term in formula (10), and from its expression formula, the response of vein degree is the equal of that structure characteristic parameter vein_feature is carried out to (1-exp (vein_feature ²/ σ ₂)) result after exponential transform, so cause the response of vein degree and maximum principal curvatures non_uniform response.

4.2 array modes based on maximal possibility estimation model refinement

Use vein_ness ⁺represent in vein degree on the occasion of, vein_ness ⁺can be formulated as follows with the relation of vein_feature:

vein_ness ⁺＝(1-exp(-vein_feature ²/σ ₂)) (13)

By known vein_feature >=0 of formula (9), by formula (13), be can be derived from:

$\mathrm{vein}\_\mathrm{feature}=\sqrt{{\mathrm{\&sigma;}}_2\&times;\ln \left(\frac{1}{1-\mathrm{vein}\_{\mathrm{ness}}^{+}}\right)}---\left(14\right)$

Maximum principal curvatures response when vein_feature is λ > 0, this explanation utilizes formula (14) to vein_ness ⁺convert, can obtain similarly responding with maximum principal curvatures; When vein_ness=0, the vein degree improving in array mode also should be 0, and in conjunction with formula (14), the formula that improves array mode is expressed as follows:

$\mathrm{vein}\_{\mathrm{ness}}^{\text{'}}=\sqrt{{\mathrm{\&sigma;}}_2\&times;\ln \left(\frac{1}{1-\mathrm{vein}\_\mathrm{ness}}\right)}---\left(15\right)$

In formula, vein_ness ' expression improves the vein degree in array mode, σ ₂, consistent with formula (10) of vein_ness; This built-up pattern can be understood as the structure characteristic parameter bound term in maximum principal curvatures response through type (10), and after transforming to and processing in vein ness territory, inverse transformation is returned.Figure 12 shows that the vein degree that Figure 11 is corresponding responds after formula (15) conversion, the response gray-scale map obtaining.

As can be seen from Figure 12, figure medium sized vein clean mark, vein crossings district is also very clear, compare with the region in red rectangle frame in Figure 11, at thin vein and low contrast venosomes, still can guarantee vein clean mark, illustrate that the array mode based on maximal possibility estimation model refinement can be reacted vein texture information better.

4. extract hand back vein texture

The impact of 4.1 differential yardstick opponent dorsal vein texture VLSDM model responses

As Figure 13, thick vein texture is insensitive for the differential yardstick that is less than its width, therefore can select to be less than the differential yardstick of its width; Thin vein texture, due to comparatively responsive to differential yardstick, causes too greatly its out-of-shape, and too little meeting causes it to rupture, and therefore concerning thin vein, differential yardstick is suitable; Background area Noise texture appears in less differential yardstick response, and the increase along with differential yardstick, fades away, and therefore for fear of the interference of noise, can not select too little differential yardstick.Vein texture edge noise generally appears in the differential yardstick close with thin vein response, but its vein degree is generally less, and along with the increase of differential yardstick, edge noise can reduce gradually.

4.2 multiple dimensioned VLSDM models

Thin vein texture in vein image can disappear along with differential yardstick increases thin vein, its reason can be understood from formula (2) and formula (3), from formula, can find out, before asking for pixel second-order differential, with the represented Gaussian filter of formula (3) is existing, image has been carried out to smoothing processing, because thin vein texture exists width little, the feature that contrast is low, this has determined that its local curved-surface shape itself is obvious not, after the Gaussian filter smoothing processing of large scale, make the characteristic parameter of its local curve form less, thereby cause its vein degree to diminish, so thin vein texture can diminish when differential yardstick becomes large, only have when filter window width is close with its texture width, now can regard gaussian filtering as matched filter, because Gauss's window Two-dimensional Surfaces shape is similar to vein texture curve form, therefore can reach a larger value, this has also illustrated a bit: when differential yardstick is large from little change, can reach maximal value at a certain yardstick, this point sets out, and the vein degree of corresponding pixel points during by more different differential yardstick is got its maximal value as last vein degree, and formula is expressed as follows:

$\mathrm{vein}(i,j)=\underset{s_{\min }\&le;s\&le;s_{\max }}{\max }\mathrm{vein}\_\mathrm{ness}(s,i,j)---\left(16\right)$

In formula, i, the transverse and longitudinal coordinate figure of pixel in j presentation video, vein (i, j) is illustrated in the texture vein degree (the present invention is called the maximum vein degree in differential metric space) that pixel (i, j) obtains by multiple dimensioned VLSDM model, s _minrepresent minimum differential yardstick, s _maxrepresent maximum differential yardstick, when vein_ness in formula (s, i, j) represents that differential yardstick is s, the vein degree of pixel (i, j).

After processing by formula (16), can guarantee that thin vein texture can not disappear, but background area Noise texture and vein texture edge noise can still exist, because they also can be under occurring that a maximal value is retained, for reduce the interference of noise as far as possible, so s _mincan not be too little; Same s _maxcan not obtain too greatly, otherwise can bring thick vein pattern reason to deform, intersect the consequences such as vein sticks together, counting yield is affected.S _min, s _maxsize relevant with minimum widith and the breadth extreme of hand back vein texture respectively, under normal circumstances, vein sample is after size normalization, it is wide that thin vein unity and coherence in writing is approximately 11 pixels, it is wide that the width of the thickest vein texture is approximately 36 pixels.Therefore, the minimum in formula (3), largest Gaussian one filter window width can be similar to and think 11 and 36, and the relation of gaussian filtering window width and differential yardstick s can use formula (17) to estimate:

In s=(n/2-1) * 0.3+0.8 (17) formula, n represents the width of filter window.Through type (3.18) can estimate s _min=2.15, s _max=5.9; Because thick vein texture is insensitive to being less than the differential yardstick of its width, in order to prevent, intersect the problem of vein adhesion and thick vein deformation texture, can get the value that is less than 5.9, the present invention gets s _max=4; In order to guarantee to extract the impact that reduces ground unrest when thin vein texture extracts, the present invention gets s _min=2, be less than the estimated value that through type (17) draws.

In order to guarantee to extract compared with thin vein texture, the present invention gets s _min=2, this has brought a problem: because vein unity and coherence in writing edge noise can not get suppressing, therefore must carry out filtering to noise.

The post-processed of 4.3 multiple dimensioned VLSDM model responses

Figure 14 (A) is depicted as the response gray-scale map after formula (16) is processed, Figure 14 (B) is depicted as differential yardstick gray-scale map corresponding to maximum vein degree in differential metric space, and Figure 14 (C) is depicted as the histogram curve (disregarding gray-scale value is zero pixel) of multiple dimensioned VLSDM model response gray-scale map.

From Figure 14 (A), can find out, in multiple dimensioned VLSDM model response, have noise, this point can be more clearly visible from Figure 14 (B), and this is because in multiple dimensioned VLSDM model, get s _min=2, can only remove a part of background area noise, and to vein texture edge noise to improve effect limited, so there is a small amount of speckle noise in Figure 14 (B), also have many glitch noises that are connected with vein texture simultaneously.From Figure 14 (B), can find out that gray-scale value corresponding to most of noise region is less, the maximum vein degree in this explanation noise region differential metric space appears in small-scale responses.In Figure 14 (B), there is a darker region (oval frame marks), its correspondence be thin vein texture region, this illustrates that the maximum vein degree in its differential metric space also appears in small-scale responses.From Figure 14 (C), histogram curve can be found out, there is a maximum point in it, what this point was corresponding is the number summation of dark pixel point, part dark noise point and the thin vein area part dark pixel point at figure A medium sized vein texture edge, obviously can not directly utilize this value to carry out threshold filter processing to Figure 14 (A), but can and Figure 14 (B) join together image Figure 14 (A) to carry out filtering processing, they are combined to the response to multiple dimensioned VLSDM model below and carry out filtering.

From Fig. 3 .15B, can find out, noise region in figure and thin vein region can be divided into many different connected regions (the blackening region in Figure 14 (B)) according to differential yardstick corresponding to the maximum vein degree of differential metric space, and the feature that noise is communicated with district is: the vein degree generally less (Figure 14 (A) bottom right rectangle frame marks) of correspondence in region, the feature that thin vein is communicated with district is: center vein degree is larger, and frontier district is less, especially with other vein textures Fang Geng little (Figure 14 (A) upper left rectangle frame marks) across, although the multiple dimensioned VLSDM model response that some pixel of thin vein connected region and noise region is corresponding equates, but in these two kinds of connected regions, the maximum value difference of corresponding multiple dimensioned VLSDM model response is larger, if can find a threshold value separates them, just can filter the noise connected domain in Figure 14 (B), known from introduction above, near the response gray value interval of correspondence maximum point in Figure 14 (C), be exactly that thin vein is communicated with district's response and is communicated with district with noise and responds internal jugular vein degree gray-scale value and have the interval intersecting, therefore distinguish gray-scale value corresponding to threshold value that thin vein connected domain and multiple dimensioned VLSDM model corresponding in noise connected domain respond gray-scale value maximum value and histogram extreme value relevant, if use T _maxrepresent this threshold value, use G _h-maxthe gray-scale value that represents the multiple dimensioned VLSDM model response that Histogram Maximum point is corresponding, by hand back vein sample is analyzed, multiple dimensioned VLSDM model response is carried out after 256 grades of gray scale normalizations, and dark noise is communicated with district's response gray-scale value mostly and G _h-maxapproach, the peak response gray-scale value in bright noise connected region is generally G _h-max2 ~ 3 times large, the peak response gray-scale value that thin vein is communicated with district generally can reach G _h-max6 ~ 8 times, for guaranteeing to filter noise, do not affect thin vein texture simultaneously, get T _max=4G _h-max, use I _vrepresent that the response of multiple dimensioned VLSDM model carries out 256 grades of images after gray scale normalization, use I _srepresent differential yardstick gray-scale map corresponding to maximum vein degree in differential metric space, the method concrete steps of the multiple dimensioned VLSDM model of elimination of the present invention response noises are as follows:

The first step: eliminate small scale glitch noise and speckle noise, concrete grammar is to allow I _sin background area and differential yardstick be s _maxregion be white, differential yardstick is less than s _maxregion be black, form bianry image I _b, the glitch noise region in Figure 14 and thin vein region are all at I like this _bin become the speckle regions (as shown in Figure 15 (A)) of black, add up the region maximum value in multiple dimensioned VLSDM model response gray-scale map corresponding to each speckle regions, if its value is greater than threshold value T _max, this speckle regions is filled by white.Image I after speckle regions is filled filtering processing ' _bwith original image I _sdo and operation, finally just obtained only comprising the image I of vein texture region and large scale noise region ' _s.

Second step: eliminate large scale noise, concrete grammar is, allow image I ' _smiddle background area is black, other regions are white, large scale noise has just become black perforated (in Figure 15 (B), oval frame marks) and white dot region (in Figure 15 (B), rectangle frame marks) like this, black hole noise brings because the thick vein of intersection sticks together, speckle noise is that the Noise texture that the maximum vein degree in differential metric space appears in large scale causes, the region that green rectangle frame in contrast Figure 14 (B) and 15 (B) marks, can find out that the gray-scale value of corresponding region in Figure 14 (B), white dot district is larger.The area of large scale white dot noise region and black hole noise region is generally less, therefore can carry out this class large scale noise of elimination by connected domain area threshold, area threshold for white dot noise can be selected larger value, because now the connection area of venosomes is generally larger, for the hand back vein sample carrying out after size normalization, it is 100 that the present invention gets its threshold value, and area is less than the white connected domain of this value and thinks noise, and the pixel value of this connected domain is set to 0; The threshold area in black cavity can not be got too large, because black cavity may be also many vein crossings, forms, and the threshold area of getting black hole in the present invention is 30, thinks that the black hole that is less than this value is noise, and the pixel value of noise perforated is set to 1.The image medium sized vein unity and coherence in writing region of removing after large scale noise is white portion (pixel value is 0), and background area is black (pixel value is 0), is also called Filtering Template, in filtering algorithm of the present invention as shown in Fig. 3 .16C.

The 3rd step: multiple dimensioned VLSDM model response is carried out to filtering, utilize Filtering Template and the response of multiple dimensioned VLSDM model that second step obtains to do multiply operation, noise region in multiple dimensioned like this VLSDM model response just all becomes null value, and venosomes vein degree value remains unchanged, filtered response is as shown in Figure 15 (D).

5. Comparison of experiment results and analysis

Respectively several dissimilar vein sample images have been carried out to vein texture and extracted, Figure 16 shows that the response gray-scale map of dissimilar sample after multiple dimensioned VLSDM model is processed.

In Figure 16, the first picture group medium sized vein texture thickness is uneven, the numerical value of the VLSDM model response in thick vein unity and coherence in writing region is greater than the vein degree value in thin vein region, therefore the grey scale pixel value that causes responding thick vein texture region in gray-scale map is larger than thin vein texture region, but extraction algorithm can guarantee the structural information of thin vein; In Figure 16, in the second picture group, the texture of sample image is thicker, can see that the width of thick venosomes in the response of VLSDM model is also larger, and extraction algorithm of the present invention also can keep the structural information in sample image preferably; In Figure 16, in the 3rd picture group, the texture of sample image is comparatively fuzzy, still can extract vein texture structure, although this is because local contrast is low, but the second-order differential of image also can detect the variation of local pixel gray-scale value.The 4th picture group is the result figure to the vein sample extraction vein texture of Noise, can find out that multiple dimensioned VLSDM model has very strong anti-noise ability, this also illustrates that the filtering algorithm in this extracting method is comparatively applicable to multiple dimensioned VLSDM model response to carry out noise reduction process.

Claims (6)

1. method for extracting vein-mode hand back texture, is characterized in that:

(1) set up the local second-order differential structural model-VLSDM of vein texture model: VLSDM model is comprised of the remarkable second-order differential structure characteristic parameter of the local curve form restricted model of vein texture and vein texture, according to the different feature of the local curve form constraint index value of vein texture, method in conjunction with sampling thheorem and Weighted Fusion obtains merging shape constraining index value computation model-SICM model, choose the maximum principal curvatures of curved surface as the remarkable second-order differential structure characteristic parameter of the vein texture in VLSDM model, the array mode of employing based on maximal possibility estimation model refinement merges SICM model and remarkable second-order differential structure characteristic parameter, thereby form VLSDM model,

(2) adopt the noise elimination method based on multiscale analysis to process multiple dimensioned VLSDM model filtering response, thereby obtain last vein-mode hand back texture;

The mathematic(al) representation of described VLSDM model is:

$\mathrm{vein}(i,j)=\underset{s_{\min }\&le;s\&le;s_{\max }}{\max }\mathrm{vein}\_\mathrm{ness}(s,i,j),$

The transverse and longitudinal coordinate figure of pixel in i, j presentation video in formula, vein (i, j) is illustrated in the texture vein degree that pixel (i, j) obtains by multiple dimensioned VLSDM model, s _minrepresent minimum differential yardstick, s _maxrepresent maximum differential yardstick, the vein degree of pixel (i, j) when vein_ness in formula (s, i, j) expression differential yardstick is s.

2. method for extracting vein-mode hand back texture according to claim 1, it is characterized in that: the method for building up of described SICM model is: adopt sampling function Shannon to sample to shape indexing value SapeIndex, form neighborhood of pixel points curved surface and sampled targets texture curve form index value shape _dcalculating formula of similarity, the method for employing Weighted Fusion, by intersecting vein texture and the secretly similarity fusion of ridged vein texture, forms last vein shape constraining model.

3. method for extracting vein-mode hand back texture according to claim 2, is characterized in that: described neighborhood of pixel points curved surface and sampled targets texture curve form index value shape _dcalculating formula of similarity mathematic(al) representation is:

$\mathrm{shapeness}=\left\{\begin{array}{cc}0,& \mathrm{SapeIndex}\&le;0\\ \frac{\sin \left(\mathrm{\&alpha;}(\mathrm{SapeIndex}-{\mathrm{shape}}_D)\right)}{\mathrm{\&alpha;}(\mathrm{SapeIndex}-{\mathrm{shape}}_D)},& \mathrm{SapeIndex}>0\end{array}\right.,$

In formula, α represents the attenuation coefficient of Shannon function, and shapeness represents neighborhood of pixel points curved surface and shape _drepresented texture curve form similarity.

4. method for extracting vein-mode hand back texture according to claim 3, is characterized in that: the method mathematic(al) representation of described Weighted Fusion is:

vein_shape=β*ridge+(1-β)*concavity

In formula, vein_shape represents final shape constraining value, and ridge represents the dark rib region sampled value in vein image, and concavity represents the recessed bag shape area sampling value in vein image, and β is combination coefficient.

5. method for extracting vein-mode hand back texture according to claim 4, is characterized in that: the array mode mathematic(al) representation of described maximal possibility estimation model refinement is:

$\mathrm{vein}\_\mathrm{ness}=(1-\exp (-\frac{\mathrm{vein}\_\mathrm{shap}{e}^2}{{\mathrm{\&sigma;}}_1}))(1-\exp (-\frac{\mathrm{vein}\_{\mathrm{feature}}^2}{{\mathrm{\&sigma;}}_2})),$

In formula, vein_shape, vein_feature are two eigenvalue λ of curved surface second-order differential structure matrix Hessian Matrix ₁, λ ₂in maximal value, σ ₁and σ ₂represent respectively shape constraining model sensitivity coefficient, structure characteristic parameter sensitivity coefficient, σ ₁estimation formulas as follows:

In formula, m represents shape constraining item (1-exp (vein_shape ²σ ₁)) value and 1 approximation quality, namely 0.999 ... 9 number after 9 radix point,

Structure characteristic parameter sensitivity coefficient σ ₂estimation formulas as follows:

${\mathrm{\&sigma;}}_2=\frac{{\mathrm{\&lambda;}}_{\max }^2}{m\ln 10},$

In formula, m represents shape constraining item (1-exp (vein_feature ²σ ₂)) value and 1 approximation quality, λ _maxfor the maximum value of image medium sized vein region vein_feature,

Array mode to maximal possibility estimation model refinement is further improved, and the computing formula of the vein degree vein_ness ' after improvement is:

$\mathrm{vein}\_{\mathrm{ness}}^{\&prime;}=\sqrt{{\mathrm{\&sigma;}}_2\&times;\ln \left(\frac{1}{1-\mathrm{vein}\_\mathrm{ness}}\right)}.$

6. method for extracting vein-mode hand back texture according to claim 5, is characterized in that: the described noise elimination method based on multiscale analysis is:

(1) eliminate small scale glitch noise and speckle noise: make differential yardstick gray-scale map I corresponding to maximum vein degree in differential metric space _sin background area and differential yardstick be s _maxregion be white, differential yardstick is less than s _maxregion be black, form bianry image I _b, glitch noise region and thin vein region are all at I _bin become the speckle regions of black, add up the region maximum value in multiple dimensioned VLSDM model response gray-scale map corresponding to each speckle regions, if its value is greater than the threshold value T that thin vein connected domain and multiple dimensioned VLSDM model corresponding in noise connected domain respond gray-scale value maximum value _max, this speckle regions is filled by white, the image I after speckle regions is filled filtering processing ' _dwith original image I _sdo and operation, obtain only comprising the image I of vein texture region and large scale noise region ' _s;

(2) eliminate large scale noise: make image I ' _smiddle background area is black, other regions are white, large scale noise has just become black perforated and white dot region, by connected domain area threshold, carry out this class large scale noise of elimination, the image medium sized vein unity and coherence in writing region of removing after large scale noise is white portion, background area is black, the large scale that is eliminated noise filtering template;

(3): multiple dimensioned VLSDM model response is carried out to filtering, multiply operation is done in the Filtering Template that utilization obtains and the response of multiple dimensioned VLSDM model, noise region in multiple dimensioned VLSDM model response all becomes null value, venosomes vein degree value remains unchanged, thereby obtains filtered multiple dimensioned VLSDM model filtering response.

Images