CN103400134A - Non-contact method for extracting region of interest of finger vein sample - Google Patents

Abstract

The invention provides a non-contact method for extracting a region of interest of a finger vein sample. The method is characterized by carrying out threshold segmentation on a finger vein image column by column by using a maximum variance method, and denoising through the connected domain area and a method of searching the fingertip; fitting a finger centerline by using up and down contour points of a finger, and using the slope of the finger centerline as the basis for rotation and correction; acquiring an effective contour point set of a circular arc by using the characteristic that the fingertip contour approximates a circle, fitting a fingertip circle by using effective contour points to acquire the diameter of the fingertip circle so as to carry out positioning on the region of interest, wherein the transverse width of the region of interest is the fixed width, and the longitudinal width is determined by upper and lower edge inner tangent lines of the finger contour; and finally carrying out dimension normalization on the acquired region of interest. The method provided by the invention gives sufficient consideration to the non-contact finger vein acquisition process and image characteristics, provides a new solution for extracting the region of interest of the finger vein, carries out rotation and correction by using the finger centerline, carries out positioning by using the fingertip circle, and has strong rotation resistance and translation resistance.

Description

Non-contact type finger vein sample area-of-interest exacting method

Technical field

What the present invention relates to is a kind of mode identification technology, specifically a kind of finger vena sample area-of-interest exacting method.

Background technology

Existing finger vein identification method great majority, based on complete finger-image information, are directly processed and are not only needed larger storage space with whole finger, and operand is larger, and feature is concentrated, and discrimination is lower.Contrast finger tip and refer to root, the vein blood vessel in the middle of finger is more clear, and discrimination is large, with the object of this part zone as feature extraction, more is conducive to classify.Therefore, select this part area-of-interest of the vein as us (Region Of Interest, ROI).

Apply for being in 201210051702.3 patent document, a kind of " based on the finger venous image area-of-interest exacting method of rotation correction " disclosed, utilizing finger centroid to carry out image rotation proofreaies and correct, and according to the projection value on every row pixel vertical direction in image, determine finger the first joint position, carry out the ROI location.The method has strict demand to picture quality, to having significantly bimodal at two finger-joint places after the projection of high-quality vein image column direction; To low-quality vein image, after the row projection, peak of curve is not obvious, or, because the illumination reason presents the multimodal situation, can not realize the accurate location of ROI.

Summary of the invention

The object of the present invention is to provide a kind ofly can overcome the skew that exists in gatherer process, rotation difference, obtain comprising the non-contact type finger vein sample area-of-interest exacting method in the ROI zone of abundant, stable feature.

The object of the present invention is achieved like this:

Comprise by column that Threshold segmentation, image denoising, rotation correction, anchor point are determined, region of interesting extraction and size normalization, adopt the varimax finger vein image to carry out Threshold segmentation by column, and by the method for connected domain area and searching finger tip, carry out denoising, be partitioned into finger areas; Utilize finger up and down point to simulate the finger center line, use the slope of described finger center line as the foundation of rotation correction; Utilize the finger tip contour approximation to ask for the effective contour point set of circular arc for the characteristics of circle, further, with these effective contour point match finger tip circles, obtain the finger tip diameter of a circle, with this, carry out the area-of-interest location; The transverse width of area-of-interest is fixed width, vertically width is determined by lower limb internal tangent on finger contours; Finally the area-of-interest that obtains is carried out size normalization.

Described employing varimax finger vein image is carried out Threshold segmentation by column and is specially: the illumination based on each row in image acquisition process is equally distributed, with image each row as a unit, adopt varimax, calculate the segmentation threshold of these row, and the global threshold that carries out these row is cut apart.

Described method by connected domain area and searching finger tip is carried out denoising and is specially: calculate the area of each connected domain according to zone marker, keep the area in largest connected territory, eliminate little block distortion; For the noise of finger tip part,, according to the projection value on every row pixel vertical direction in image, determine the position of finger tip, eliminate the noise of finger tip front.

Described rotation correction is specially: the point on the finger center line is obtained by the mid point of the upper and lower point of finger, and these points are simulated a near linear, and the slope of straight line is positive and negative relevant with sense of rotation, and the size of slope is relevant with the anglec of rotation.

Described anchor point is determined to be specially: utilize the finger tip contour approximation to ask for for the characteristics of circle the effective contour point set that drops in the finger tip radius of circle, further round by the least square fitting finger tip with these effective contour points, obtain the finger tip diameter of a circle, take on the finger center line apart from the point of a diametral distance of finger tip as the area-of-interest anchor point.

Described region of interesting extraction is specially: part secant on the left side from anchor point, parting secant on the right side is fixed width apart from anchor point; The up and down cut-off rule is determined by the upper and lower edge internal tangent of the folded finger contours of left and right cut-off rule.

Described size normalization is specially: according to the bilinear interpolation algorithm, the area-of-interest that extracts is carried out the length and width normalized, what realize different samples are obtained is all the identical zone of size.

Because the collection of finger vena is contactless, there is no the constraint of gim peg, the finger nature is put, and all variant on horizontal offset, the anglec of rotation between the different samples of same person, these factors have affected accuracy of identification.Therefore, first finger-image is rotated to level, consider that each collection of finger has skew, the finger length that obtains is different, so length information can not, as the foundation of location, be found the point that can position on finger.The present invention is directed to the characteristics of contactless collection image, propose to utilize the finger center line to be rotated to proofread and correct and method that the finger tip circular diameter positions, can extract exactly finger vein area-of-interest, solved the impact of rotation, translation factor.

Characteristics of the present invention are: take into full account non-contact type finger vein gatherer process and feature of image, for the finger vena region of interesting extraction provides new solution, utilization finger center line is rotated correction and the finger tip circle positions, and has very strong anti-rotation, anti-translation.

Description of drawings

The module of Fig. 1 non-contact type finger vein sample area-of-interest exacting method forms;

Fig. 2 (a)-Fig. 2 (b) is Threshold segmentation by column; Fig. 2 (a) original image wherein, Fig. 2 (b) be the image after Threshold segmentation by column;

Fig. 3 (a)-Fig. 3 (d) denoising process, wherein Fig. 3 (a) is the image that utilizes after the connected domain area is removed block distortion, curve after the projection of Fig. 3 (b) vein image object pixel row, Fig. 3 (c) is for removing the image of finger tip front-end noise, the finger areas that Fig. 3 (d) extracts;

Fig. 4 (a)-4 (b) rotation correction process, wherein Fig. 4 (a) is that the point that match is pointed on center line obtains pointing center line, the finger vein image after Fig. 4 (b) rotation correction;

Fig. 5 (a)-Fig. 5 (b) anchor point deterministic process, wherein Fig. 5 (a) rough estimate finger tip radius, obtain the effective contour point, and the finger tip arc diameter is determined in the match of Fig. 5 (b) circle;

Fig. 6 (a)-Fig. 6 (b) region of interesting extraction process, wherein Fig. 6 (a) area-of-interest size is determined schematic diagram, Fig. 6 (b) finger vena area-of-interest;

The normalized vein area-of-interest of Fig. 7 size.

Embodiment

Below describe the specific embodiment of the present invention in detail.

1, Threshold segmentation by column

The original image size that collects is 320 * 240, from Fig. 2 (a), can find out, background area accounts for the most of ratio of image, directly with the entire image processing, not only increases calculated amount and takies storage space, and be unfavorable for identification.Therefore need to first extract finger areas, reduce background area and disturb.The thresholding operation of image can realize background and target object are separated.

Threshold operation is the grey value difference according to different pixels, and the pixel greater than a certain threshold value on image is put together, forms target area; Pixel less than this threshold value is put together, forms background area, so gray level image becomes the bianry image that only has two grades.

Threshold segmentation is formulated and is exactly

$g(x,y)=\left\{\begin{array}{c}1,f(x,y)\&GreaterEqual;\mathrm{<figure-callout\; id="0"\; label="t"\; filenames="HDA00003431937400021.png"\; state="\{\{state\}\}">t</figure-callout>}\\ 0,f(x,y)<t\end{array}\right.---\left(1\right)$

In formula, f (x, y) is the gradation of image value, and t is segmentation threshold.As boundary, image is divided into two parts take t, target represents with 1, and background represents with 0.

Varimax is a kind of adaptive threshold choosing method, from the generic attribute in zone, analyzes, and makes target and background between-group variance after cutting apart maximum, and the t of this moment is optimal threshold.At first the image that is the m level to intensity profile is obtained the probability of each gray-scale value, then with variable T, it is divided into two groups, and calculate and respectively organize probability and between-group variance,

δ ²(T)=w ₀(μ ₀-μ) ²+w ₁(μ1-μ) ² (2)

In formula, μ ₀, μ ₁Be the average gray of two groups, w ₀, w ₁Be the probability of happening of two groups, μ is the entire image average gray.Change T between 0～m-1, as between-group variance δ ²T when maximum, be threshold value.

Research finds, in image acquisition process, the suffered intensity of illumination of each row pixel is uniformly distributed, and is identical, and target and background is easily divided, so consider to operate to realize obtaining of finger areas by maximum variance thresholding by column.

Concrete thought is: regard each row of image as a unit, adopt varimax, calculate the segmentation threshold of these row, and the global threshold that carries out these row is cut apart.Although this algorithm has increased segmentation times, because each pixel that participates in computing has lacked, computing is also uncomplicated, fast., through separate targets and background by column, just can obtain comprising the bianry image of complete object object.

The bianry image of Fig. 2 (b) for adopting Threshold segmentation by column to obtain, can be partitioned into complete finger areas.

2, image denoising

Inhomogeneous due to light in gatherer process, caused thresholding to be processed and do not reached desirable effect, although whole finger areas to be identified occurred, meanwhile some block impurity have also occurred., in order to extract complete finger areas, the noise outside finger areas completely must be removed.Through observing, the area of most of block impurity is generally all little than finger areas,, if set an area threshold, removes the extrinsic region less than this value, and so most of extrinsic region will be removed.

The connected domain area denoise algorithm of bianry image can be described as:

1, according to 8 connected domains of pixel, bianry image is carried out zone marker, be divided into different connected regions;

2, a kind of label of the pixel in same area, zones of different different labels;

3,, according to different labels, be added to corresponding array;

4, add up the number of each label, obtain the size of respective regions.If should keep in zone greater than area threshold, otherwise this zone is set to background.

Fig. 3 (a) is for utilizing area to remove the result of block distortion.

After eliminating the small impurities zone through back, also have a kind of noise region to exist, that is exactly the finger fingertip front end area.This zone, because finger fingertip is thinner, illumination is brighter, is being considered as target with background in the thresholding process by column, and this subregion becomes the maximum interference factor.The removal of this part noise, by determining finger tip point position, is set to background with the row of finger tip front.

The human finger finger tip is similar to circular arc, the process that scans by column from the finger root to finger tip in theory, more near the finger tip part, the object pixel of these row and fewer, crossed the finger tip point, each row white pixel point of noise region and unexpected the increase, according to this regularity of distribution, can determine finger tip point position, thereby with target and background separation.

At first, the bianry image after removal edge block distortion is carried out the column direction projection, as formula (3).In bianry image, represent background and target with 0 and 1, the projection value of so every row is this row object pixel number sum.

$L_i=\underset{j=0}{\overset{H-1}{\mathrm{\&Sigma;}}}p(i,j)---\left(3\right)$

Wherein, p (i, j) be the value that on bianry image, (i, j) locates be 0 or 1, H be the height of image.So L _iRepresent the number of object pixel in the i row.Fig. 3 (b) carries out curve after the column direction projection for vein image.

Find out that from Fig. 3 (b) there is a good trough in drop shadow curve, corresponding Fingers cusp.According to statistics, different finger length wave trough position are different, and majority is distributed between the 80-150 row.In order to improve precision, the value of the average of adjacent five row sums as these row got in the projection of 80-150 row in Fig. 3 (b).Then from 150 row, right-to-left scans by column, and when meeting the trough condition, these row are the row at finger tip point place, and the row before finger tip point are set to background.Effect is as shown in Fig. 3 (c).

For further reducing operand, before image is processed below, first extract complete target area, reduce as far as possible background area.Extraction comprises the minimum rectangle of finger: the lateral length of rectangular area is by row and the decision of image right side of finger tip point, and vertical width of rectangular area is determined by the minimum value of finger coboundary and the maximal value of finger lower limb.What Fig. 3 (d) showed is the finger areas image that extracts.

3, rotation correction

Because the collection of finger vena is contactless, there is no the constraint of gim peg, the finger nature is put, and all variant on horizontal offset, the anglec of rotation between the different samples of same person, these factors have affected accuracy of identification.Therefore, first finger-image is rotated to level, and then carry out the ROI location.

At first will determine the anglec of rotation of the vein image that collects, the present invention obtains by the slope of finger center line.In bianry image, the marginal point of finger forms the profile of finger, and the point on the finger center line is obtained by the mid point of finger up and down point, and these points are simulated a near linear, the slope of straight line is positive and negative relevant with sense of rotation so, and the size of slope is relevant with the anglec of rotation.The point that Fig. 4 (a) provides on the finger center line distributes, and fitting a straight line.

The finger center line can adopt least square fitting to obtain.As a kind of optimization algorithm, least square method is by making the error sum of squares of trying to achieve between data and real data minimum, finding optimum data.Corresponding to the straight-line equation match,, exactly with one group of data that meet straight-line equation, with the straight-line equation of setting, to do poorly, sum of squared errors function hour, is tried to achieve the optimum value of intercept a and slope b.

If the expression formula of straight-line equation is:

y=a+bx (4)

A and b can be obtained by formula (5):

$a=\stackrel{\&OverBar;}{y}-b\stackrel{\&OverBar;}{x},b=\frac{\stackrel{\&OverBar;}{\mathrm{xy}}-\stackrel{\&OverBar;}{x}\stackrel{\&OverBar;}{y}}{\stackrel{\&OverBar;}{x^2}-{\stackrel{\&OverBar;}{x}}^2}---\left(5\right)$

Wherein:

$\stackrel{\&OverBar;}{x}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{x}_i,\stackrel{\&OverBar;}{y}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{y}_i$ (6)

$\stackrel{\&OverBar;}{x^2}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{x}_i^2,\stackrel{\&OverBar;}{\mathrm{xy}}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{x}_i{y}_i$

A is the intercept of straight line, and b is the slope of straight line, and the vergence direction of finger and the anglec of rotation can be obtained by the positive and negative of b and size.

Next the algorithm according to image rotation is adjusted to level with finger-image.As shown in Fig. 4 (b).

4, anchor point is determined

Contain two articulations digitorum manus on finger, finger is divided into three sections, we determine that tentatively extracting middle one section of finger is area-of-interest.The rotation correction of finger areas, overcome image rotation, bias effect, and next the ROI orientation problem is discussed.On finger, available information has finger length, width, but because finger is each, gathers skew is arranged, and the finger length that obtains is different, so length information can not be as the foundation of location.

The image finger tip contour approximation that gathers is circular arc, therefore can adopt the diameter of circular arc to carry out the ROI location.Concrete grammar is: first find the effective circular arc point set on the finger tip profile, then these points are justified match, calculate the match diameter.

Effectively finger tip profile point set is the method acquisition by the rough estimate radius.In finger areas after rotation correction, regulation finger center line left end is starting point, and the center of circle is dropped on the finger center line, and starting point is radius to the distance between the center of circle.From left to right move along the finger center line from starting point in the center of circle, and radius increases gradually, until radius exceed the upper down contour point of finger apart from center line distance from, obtain the maximum estimated value of finger tip radius.As shown in Fig. 5 (a).Point in this radius is all available point.

Next these point sets are justified match, determine the exact value of finger tip arc radius by the mode of calculating.Fig. 5 (b) is for solving the schematic diagram of finger tip radius.Adopt the least square optimized algorithm.

If the point (X, Y) on circular arc meets equation of a circle:

R ²=(X-A) ²+(Y-B) ² (7)

A, B, R can be obtained by formula:

$A=-\frac{a}{2}$

$B=-\frac{b}{2}---\left(8\right)$

$R=\frac{1}{2}\sqrt{a^2+b^2-4c}$

Wherein

$a=\frac{\mathrm{HD}-\mathrm{EG}}{\mathrm{CG}-D^2}$

$b=\frac{\mathrm{HC}-\mathrm{ED}}{D^2-\mathrm{GC}}---\left(9\right)$

$c=-\frac{\mathrm{\&Sigma;}(X_i^2+Y_i^2)+\mathrm{a\&Sigma;}{X}_i+\mathrm{b\&Sigma;}{Y}_i}{N}$

C, D, E, G, formula (10) is seen in the calculating of H:

$C=(\mathrm{N\&Sigma;}{X}_i^2-\mathrm{\&Sigma;}{X}_i\mathrm{\&Sigma;}{X}_i)$

D=(N∑X _iY _i-∑X _i∑Y _i)

$E=\mathrm{N\&Sigma;}{X}_i^3+\mathrm{N\&Sigma;}{X}_i{Y}_i^2-\mathrm{\&Sigma;}(X_i^2+Y_i^2)\mathrm{\&Sigma;}{X}_i---\left(10\right)$

$G=(\mathrm{N\&Sigma;}{Y}_i^2-\mathrm{\&Sigma;}{Y}_i\mathrm{\&Sigma;}{Y}_i)$

$H=\mathrm{N\&Sigma;}{X}_i^2{Y}_i+\mathrm{N\&Sigma;}{Y}_i^3-\mathrm{\&Sigma;}(X_i^2+Y_i^2)\mathrm{\&Sigma;}{Y}_i$

R is that match obtains radius, and the finger tip circular diameter is anchor point.

5, region of interesting extraction

Try to achieve finger tip profile radius of circle by match, after rotation correction in finger areas, take finger center line left end as starting point, the cut-off rule in interesting image regions vertical direction left side is: L1:x=2*R, the cut-off rule L2 on right side is defined as according to L1: L2:x=2*R+d, in formula, d represents the ROI that extracts lateral length.According to statistics, the finger areas finger length after rotation correction is in 200 about pixels, and the finger tip diameter is 50 left and right,, if refer to that root portion removes same length, points so intermediate length and is approximately 100, therefore the d value is taken as 100.

In the contour images of finger areas, consider cut-off rule L1, the folded finger areas of L2, vertical width in ROI zone is determined by the minimum value of finger contours coboundary and the maximal value of finger contours lower limb, the rectangle that these four straight lines enclose is ROI, this process schematic diagram is as shown in Fig. 6 (a), and the actual finger vena area-of-interest that obtains is as shown in Fig. 6 (b).

6, size normalization

The ROI width of said extracted is defined as 100, and different sample height are different, in order to study conveniently, ROI is carried out height normalization, makes effective coverage corresponding to different samples measure-alike.Size normalization is exactly the method by linear transformation, image not of uniform size is become the image of prescribed level.

Size normalization relates to the gray-level interpolation algorithm, and this is to establish a capital integral point on corresponding original image because the point on target image differs, and often running into coordinate in length and breadth is all the situation of decimal, will determine that so the value of object pixel will be used interpolation algorithm.In arest neighbors interpolation, bilinear interpolation [42] and curve interpolation, nearest neighbor method is Simple fast, takes to round up for the processing of decimal and finds nearest integer; Bilinear interpolation is compose upper corresponding weight value far and near according to the distance to 4 consecutive point, calculates the value of this point, and error is less.This paper considers to adopt the bilinearity method.

To a large amount of finger venous image experiment Analysis, the altitude range of area-of-interest, in 60 left and right, therefore, is normalized to 60 with the height of image.What Fig. 7 showed is unified 100 * 60 the image that is of a size of of finger vena ROI.

Claims (3)

1. non-contact type finger vein sample area-of-interest exacting method, comprise by column that Threshold segmentation, image denoising, rotation correction, anchor point are determined, region of interesting extraction and size normalization, it is characterized in that: adopt the varimax finger vein image to carry out Threshold segmentation by column, and by the method for connected domain area and searching finger tip, carry out denoising, be partitioned into finger areas; Utilize finger up and down point to simulate the finger center line, use the slope of described finger center line as the foundation of rotation correction; Utilize the finger tip contour approximation to ask for the effective contour point set of circular arc for the characteristics of circle, further, with these effective contour point match finger tip circles, obtain the finger tip diameter of a circle, with this, carry out the area-of-interest location; The transverse width of area-of-interest is fixed width, vertically width is determined by lower limb internal tangent on finger contours; Finally the area-of-interest that obtains is carried out size normalization.

2. non-contact type finger vein sample area-of-interest exacting method according to claim 1, it is characterized in that described employing varimax finger vein image carries out Threshold segmentation by column and be specially: the illumination based on each row in image acquisition process is equally distributed, with image each row as a unit, adopt varimax, calculate the segmentation threshold of these row, and the global threshold that carries out these row is cut apart.

3. non-contact type finger vein sample area-of-interest exacting method according to claim 1 and 2, it is characterized in that describedly carrying out denoising by connected domain area and the method for finding finger tip and being specially: the area that calculates each connected domain according to zone marker, the area that keeps largest connected territory, eliminate little block distortion; For the noise of finger tip part,, according to the projection value on every row pixel vertical direction in image, determine the position of finger tip, eliminate the noise of finger tip front.

Images