CN102324092B - Method for automatically cutting granular object in digital image - Google Patents

Abstract

The invention discloses a method for automatically cutting a granular object in a digital image, belonging to the technical field of digital image processing. The method comprises the following steps of: firstly, separating an object from a background by applying an automatic threshold method by aiming at characteristics such as gray level, structural distribution, geometry and the like of the granular object in the digital image, particularly a microscopic image; then, calculating a gradient vector field of the object, and searching a key point in the gradient vector field, wherein the ideal key point has corresponding gradient vector distribution in eight neighborhoods, the gradient value of the key point is zero, and the acquired key point is used as the center of each granular object; next, defining a new effective energy function based on the gray level and a space position so as to calculate a direction gradient to replace the traditional gray level gradient; and finally, searching the boundary of the granular object by applying an active contour model. By using the method, the aggregated granular object can be accurately and effectively cut, particularly, a great number of adhered or overlapped micro-grains exist in a biomedical microscopic image, and therefore, help is provided for the image analysis and identification.

Description

The automatic division method of graininess object in a kind of digital picture

Technical field

The invention belongs to cutting apart automatically of graininess object in digital image analysis and the processing technology field, particularly digital picture.

Background technology

Digital picture, especially in the biomedical micro-image, structure or the microstructure of many objects are graininess, micro-image as the starch granules of starch food potato, wheat, Thermophilic Bacteria, spherical coccus, red blood cell, small lymphocyte etc., its characteristics are: (1) amounts of particles is more, is the gathering distribution; (2) grain spacing exists a large amount of adhesions and overlapping phenomenon from very little; (3) regular shape mostly is circle or sub-circular greatly.Carrying out biology, when basic medical is studied, usually needing structure and the distribution character of these particles are analyzed and handled, extracting and identification, unusual judgement, growth course monitoring and dynamic evaluation etc. as cell/grain count, virus characteristic.Be basis and the important step of carrying out above-mentioned analysis cutting apart of cell/graininess object, yet because limited illumination condition, and a large amount of graininess object sticks together or reason such as overlapping, makes become very difficult cutting apart automatically of they.

It is that image processing field causes one of focus that people pay close attention to and difficult point problem that image is cut apart always.In numerous image partition methods, the more representational dividing method that is based on gray scale and profile geometries, active contour model (active contourmodel, ACM or snake), in 1988 (Kass M is proposed by people such as Kass the earliest, Witkin A, Terzopoulos D.Snakes:Active contour models, Int.J.Comput.Vision.321-331, (1988)).It sets up a function that comprises based on energy in gradation of image external enwergy and the profile, by the driving of contour curve internal force and shade of gray external force, makes initial profile produce deformation, finally converges to real object boundary, thereby realizes cutting apart of target.Subsequently, many improved methods have appearred, as people such as Xu (Xu C, Prince J L. Snakes, shapes, and gradient vector flow.IEEE Trans ImageProcessing 7 (3): 359-369 (1998); Xu C., Prince J L. Generalized gradient vector flow externalforces for active contours, Signal Processing 71 (2): gradient vector flow (the gradientvector flow of Ti Chuing 131-139 (1998)), GVF), enlarged the catching range of traditional Snake, the robustness stronger to the noise tool.

In recent years, another kind of application is method very widely, level set (Level set), also be used to cell or graininess object micro-image cut apart (Solorzano De C, Malladi R, Lelievre S, et al.Segmentation of nuclei and cells usingmembrane related protein markers, Journal ofMicroscopy 201:404-415 (2001); Vese L. Chan T.Amultiphase level set framework for image segmentation using the Mumford and Shah model, International Journal of Computer Vision 50 (3): 271-293, (2002); Yan P, Zhou X, Shah M, et al.Automatic segmentation of high throughput RNAi fluorescent cellular images, IEEE Transaction onInformation Technology in Biomedicine 12 (1): 109-117 (2008)).Based on cutting apart of geometric deformation model, no matter be traditional Snake, or improved GVF, and balloon (Ballon) power model, or method such as B batten Snake, when they cut apart single target, through after the suitable initialization, can search object boundary rapidly and accurately; Deficiencies such as the level set algorithm though the search problem of two-dimensional silhouette can be transformed to the evolution of surface topology in the three dimensions, makes it be more suitable for cutting apart multiple goal, and there is structure rate function difficulty in it, the control parameter is more, time overhead is big.What is more important, these two class methods, for having a large amount of adhesions and overlapping object in the image, its performance is unsatisfactory, and they can't detect correct object bounds, implement to cut apart accurately and effectively.

Summary of the invention

The shortcoming that the object of the invention is to exist at prior art provides the automatic division method of graininess object in a kind of digital picture with not enough.

Purpose of the present invention is achieved through the following technical solutions:

The automatic division method of graininess object comprises the steps: in a kind of digital picture

First step separating particles shape object and background: earlier image is converted into bianry image, adopts the automatic threshold method that object to be split is separated with background again;

The center of second step search graininess object: the gradient vector field of computed image; Ask the particle central point;

The 3rd step calculated direction gradient: make up the energy function based on gray scale and locus earlier, calculate the central point of the energy maximum of each pixel correspondence, as the strongest center of this pixel homogeneity; Calculate the direction gradient at each pixel center corresponding with it;

The circular profile of the 4th step initialization: from the center of each particle, respectively along the upper and lower, left and right four direction, search for the pixel of first nonzero value, with the distance of center and this pixel, as the initialization radius of a circle;

The 5th step was cut apart the graininess object: utilize GVF Snake deformation model to implement to cut apart, the gradient of its external force item adopts the direction gradient in the 3rd step.

Effective energy function based on gray scale and locus of definition in the 3rd step:

E _effect＝k ₁E _pos+k ₂E _homo，

Wherein $E_{\mathrm{pos}}=e^{{-\mathrm{\γ}}_1}\left|\right|P-C\left|\right|,$ $E_{\mathrm{homo}}=e^{{-\mathrm{\γ}}_2|f(i,j)-{\mathrm{\μ}}_k|};$

In the formula, k ₁With k ₂Be weighting parameter;

γ ₁With γ ₁For greater than 0 constant;

P is pixel; C is the particle center;

μ _kBe particle C _kGray average.

Direction gradient in the 3rd step is defined as: a certain pixel P (p in the image _x, p _y) point to a certain center C (c _x, c _y) the direction gradient vector be d (x, y)=(d _x, d _y), wherein $d_x=\frac{c_x-p_x}{\left|\right|P-C\left|\right|},$ $d_y=\frac{c_y-p_y}{\left|\right|P-C\left|\right|}.$

Compared with prior art, the present invention has following beneficial effect:

(1) locatees the particle central point automatically, determine amounts of particles.

(2) by the homogeneity function of foundation based on gray scale and locus, the central point that the equal unique correspondence of energy of each pixel and its homogeney are the highest, namely each central point all has a homogeneity regional area, or claims field of force scope.

(3) replace traditional shade of gray with direction gradient, on the one hand, owing to only need to calculate the gradient of a direction, therefore can reduce calculated amount, on the other hand, can improve the accuracy of segmentation result.

(4) cutting procedure is without any need for artificial participation, and cutting procedure is quick, and the result is accurate.

(5) aggregation is had remarkable specificity strong cutting apart of particle, can effectively cut apart adhesion or overlapping particle.

Description of drawings

Fig. 1 is the process flow diagram of the automatic division method of graininess object in a kind of digital picture of the present invention.

Fig. 2 is similarity between the present invention's 8 neighborhood directions and the vector.

Fig. 3 is embodiment of the invention potato starch particle micro-image.

Fig. 4 is the histogram of embodiment of the invention image.

Fig. 5 is binary map after the particle of embodiment of the invention image and the background separation.

Fig. 6 is GVF field and the particle center of embodiment of the invention image.

Fig. 7 points to the vector plot at each particle center for the embodiment of the invention.

The direction gradient figure that Fig. 8 obtains for the embodiment of the invention.

Fig. 9 is the initial circular profile diagram of the embodiment of the invention.

Figure 10 is the segmentation result of embodiment of the invention potato starch particle micro-image.

Figure 11～13 are three research objects of the embodiment of the invention.

Figure 14～16 are the segmentation result of three research object correspondences of the embodiment of the invention.

Embodiment

Below in conjunction with embodiment and accompanying drawing technical scheme of the present invention is further described, but embodiments of the present invention are not limited thereto.

Example adopts the micro-image of potato starch particle, is gathered by DP71 Olympus optical microscope.

As shown in Figure 1, the present invention includes following steps:

(1) micro-image of input potato starch particle to be analyzed;

(2) separating particles shape object and background;

From Fig. 4 starch granules micro-image and histogram thereof as can be known, image mainly contains two bigger main peaks of being separated by, and is the less starch granules of gray-scale value between lower peak region, is the high brightness background between higher peak region.

Adopt overall adaptive threshold method with object and background separation: an at first given initial threshold (as gradation of image mean value), by iterative algorithm, make the threshold approaches optimal value gradually then, obtain threshold value accurately at last, image is implemented to cut apart.Detailed process is as follows:

Use this threshold value image is divided into two classes, calculate the average density of two class objects respectively, ask the average of two average densities and as new threshold value, again image is classified, whether more double threshold value poor, or judge whether iterations reaches maximal value determine the end process process.Its process can be described as:

A) select initial estimate T ₀, given very little stop value t, and maximum iteration time N _Max

B) use T ₀As threshold value image is divided into C ₁With C ₂Two classes;

C) to C ₁With C ₂In all pixels calculate average densities

${\mathrm{\μ}}_1=\frac{1}{N_{C_1}}\underset{i\∈c_1}{\mathrm{\Σ}}f\left(i\right),$ ${\mathrm{\μ}}_2=\frac{1}{N_{C_2}}\underset{j\∈c_2}{\mathrm{\Σ}}f\left(j\right)---\left(1\right)$

Wherein f () is image density, With

Be respectively C ₁With C ₂Pixel quantity.

D) calculate new threshold value

$T_1=\frac{{\mathrm{\μ}}_1+{\mathrm{\μ}}_2}{2}---\left(2\right)$

E) the difference Δ T=|T of calculating continuous quadratic threshold value ₁-T ₀|, if Δ T＜t, or iterations equals N _Max, then finish; Otherwise, with new threshold value T ₁Compose and give T ₀, repeating step b)-e).

Fig. 5 is after the automatic threshold method is handled, the result after particle and the background separation.

(3) center of search graininess object

When obtaining particle and the background bianry image after separated, at first the compute gradient vector field is asked the particle center again.

(3.1) compute gradient vector field (GVF)

Based on gradient vector field v (x, y)=(u (and x, y), the energy function of v (x, y)) is defined as:

$E_{\mathrm{gvf}}\left(v\right)=\∫\∫\mathrm{\μ}(u_x^2+u_y^2+v_x^2+v_y^2)+{|\▿f|}^2{|v-\▿f|}^2\mathrm{dxdy}---\left(3\right)$

Wherein, (x y) is image border figure, u to f _xWith u _yBe respectively u (x, y) and v (x, y) at the local derviation of x and y direction, || the mould of expression vector.

Ask the minimum value of (3) formula energy function, can use the variational method and find the solution following two Euler equations:

${\mathrm{\μ}\▿u}^2-(u-f_x)(f_x^2+f_y^2)=0---\left(4\right)$

${\mathrm{\μ}\▿v}^2-(v-f_y)(f_x^2+f_y^2)=0---\left(5\right)$

Wherein

Be Laplacian operator, f _xWith f _yBe respectively f at the local derviation of x and y direction.

Try to achieve by following process of iteration the GVF field:

$u_t(x,y,t)={\mathrm{\μ}\▿}^2{u}_t(x,y,t)-(u(x,y,t)-f_x(x,y))(f_x{(x,y)}^2+f_y{(x,y)}^2)---\left(6\right)$

$v_t(x,y,t)={\mathrm{\μ}\▿}^2{v}_t(x,y,t)-(v(x,y,t)-f_y(x,y))(f_x{(x,y)}^2+f_y{(x,y)}^2)---\left(7\right)$

Wherein t is iterations.

(3.2) ask the particle center

Because particle mostly is circle or sub-circular greatly, desirable central point has two features: (a) the GVF vector points to 8 neighborhood directions, and from the center point, the equal directed quantity of 8 outside neighborhood directions distributes; (b) Grad is 0.But in the real image, because uneven illumination is even or The noise, the GVF vector of central point is not to point to 8 neighborhood directions fully, but close with 8 neighborhood directions, therefore, if the similarity of two vectors is enough big, thinks that then the two is consistent.The similarity measure of vector is defined as follows:

S=(S wherein _x, S _y) central point investigated of expression,

Be 8 neighborhood territory pixels of pixel S, $V_i=(V_x^i,V_y^i)(i=\mathrm{1,2}...,8)$ Be N _iThe gray scale ladder, the expression dot product,

For pointing to N from S _iDirection vector:

The similarity measure of above-mentioned vector is actually the cosine value of two vector angles, and span is [0,1], and its value is more big, and then vector is more similar.In the example of the present invention, if the similar value of two vectors, thinks then that they are consistent more than or equal to 0.8, in 8 neighborhood directions, if having the GVF vector of 6 above directions, then point centered by this pixel.

Similarity between 8 neighborhood directions and the vector as shown in Figure 2, Fig. 6 is the particle central point of trying to achieve.

(4) calculated direction gradient

(4.1) calculate effective energy

Great majority are based on the profile searching method of deformation, be difficult to cut apart adhesion or overlapping particle, main cause is a large amount of agglomeration of particles, the field of force is comparatively complicated, the bigger particle of shade of gray wherein, profile is had stronger attractive force, tend to drive gray scale on every side and move to it than the profile of granule, thereby converge to wrong border.

For head it off, the present invention defines following energy function:

E _effect＝k ₁E _pos+k ₂E _homo (10)

K wherein ₁With k ₂Be weight coefficient.First term E _Pos, be defined as follows:

$E_{\mathrm{pos}}=e^{-{\mathrm{\γ}}_1\left|\right|P-C\left|\right|}---\left(11\right)$

Wherein P is pixel, and C is the particle center, γ ₁For greater than 0 constant.

Second then is:

$E_{\mathrm{homo}}=e^{-{\mathrm{\γ}}_2|f(i,j)-{\mathrm{\μ}}_k|}---\left(12\right)$

γ ₂For greater than 0 constant, μ _kBe particle C _kGray average.

(10) energy function of formula is called effective energy, E among the present invention _PosRelevant with the locus, reflected the distance relation at pixel and particle center, its value is the negative exponential function of distance between pixel and the center, and namely distance is more little between the two, and its value is more big, and along with the increase of distance, its value is negative exponent and successively decreases.Second E _HomoThe gray difference that has then reflected pixel and particle, it is the negative exponential function of the difference of pixel and particle gray average, and namely the gray scale of the two differs more little, and its value is more big, and vice versa.This energy function has reflected the homogeney of current pixel and particle in essence, if the gray scale of current pixel is close with the particle gray average, distance is less again, and then its available energy value is bigger, otherwise then less.Each pixel is each particle and center thereof relatively, and an available energy value is arranged, and gets energy the maximum, as the strongest center of current pixel homogeney, so that the calculated direction gradient.

(4.2) calculated direction gradient

Because in the step (4.1), each pixel all has the strongest particle of unique homogeney corresponding, therefore, can be from this pixel P (p _x, p _y) set out, obtain pointing to corresponding particle center C (c _x, c _y) vectorial D (d _x, d _y) be:

$d_x=\frac{c_x-p_x}{\left|\right|P-C\left|\right|}---\left(13\right)$

$d_y=\frac{c_y-p_y}{\left|\right|P-C\left|\right|}---\left(14\right)$

The present invention replaces traditional shade of gray with the shade of gray of this direction, can weaken the bigger particle of shade of gray value near the influence of Grad than granule, makes profile converge to granule boundary accurately.In the real image, vectorial D (d _x, d _y) also there be the situation not quite identical with 8 neighborhood directions, at this moment, the method when asking the particle center is similar, calculates the included angle cosine similarity of two vectors, if the similarity of the two, thinks then that the two is consistent more than or equal to 0.8.

Fig. 7, Fig. 8 are respectively the direction vectors that points to central point, and corresponding direction GVF field.Fig. 7 shows that each pixel is all pointed to the central point the strongest with its homogeney, and namely each central point has a field of force scope.

(5) the circular profile of initialization

Because particle is circle or sub-circular, therefore, the circle that profile is initialized as certain radius is comparatively suitable.For the bianry image after background and the particle separation, can be from the center of each particle, respectively along the upper and lower, left and right four direction, search for the pixel of first non-0 value, distance with center and this pixel, as the initialization radius of a circle, thus, the initial profile of each particle is an incircle.Fig. 9 is the circular profile of initialization.

(6) cut apart the graininess object

Each particle is through after the initialization, and the gradient in the external force item of GVF Snake deformation model adopts direction gradient, and under the driving of interior external force, initial profile can converge to the real border of particle automatically.

Figure 10 is the final segmentation result of Fig. 3, and Figure 11～13 have shown more starch granules micro-image, and Figure 14～16 are followed successively by the segmentation result of Figure 11～13 correspondences.

Fig. 3～10 have illustrated processing and the segmentation result of example.Figure 10 is the result of the dividing method key step of the present invention's proposition; Fig. 3 is starch granules micro-image to be split; Fig. 4 is its histogram; Fig. 5 is then for using the bianry image that the automatic threshold method obtains after with particle and background separation in the step (1); Fig. 6 is the GVF field and the particle center of trying to achieve in the step (3); Fig. 7 is the vector at each particle center of sensing of trying to achieve behind the effective energy as calculated described in the step (4); Fig. 8 is the direction gradient that step (4.2) is obtained; Fig. 8 is the initial profile in the step (5); Figure 10 is the result after finally cutting apart.

The parameter that the deal with data set sets in this example is as follows:

Threshold value initial estimate T in the step (2) ₀=image averaging density;

Stop value t=0.05 in the step (2);

Maximum iteration time N in the step (2) _Max=300;

Vectorial similarity threshold value is 0.8 in the step (3.2), namely works as two vector angle cosine values more than or equal to 0.8, thinks that then the two is consistent;

In the formula (10) of step (4.1), k ₁=0.5, k ₂=0.1;

In the formula (11) of step (4.1), γ ₁=0.75;

In the formula (12) of step (4.1), γ ₂=1.0.

Realization of the present invention is not limited to the disclosed scope of above-mentioned example, can adopt the form that is different from above-mentioned example to realize technique scheme.

Claims (2)

1. the automatic division method of graininess object in the digital picture is characterized in that comprising the steps:

First step separating particles shape object and background: earlier image is converted into bianry image, adopts the automatic threshold method that object to be split is separated with background again;

The center of second step search graininess object: the gradient vector field of computed image; Ask the particle central point;

The 3rd step calculated direction gradient: make up the energy function based on gray scale and locus earlier, calculate the central point of the energy maximum of each pixel correspondence, as the strongest center of this pixel homogeneity; Calculate the direction gradient at each pixel center corresponding with it;

The circular profile of the 4th step initialization: from the center of each particle, respectively along the upper and lower, left and right four direction, search for the pixel of first nonzero value, with the distance of center and this pixel, as the initialization radius of a circle;

The 5th step was cut apart the graininess object: utilize GVF Snake deformation model to implement to cut apart, the gradient of its external force item adopts the direction gradient in the 3rd step;

In described the 3rd step, define an effective energy function based on gray scale and locus:

E _effect=k ₁E _pos+k ₂E _homo，

Wherein $E_{\mathrm{pos}}=e^{{\mathrm{\γ}}_1}\left|\right|P-C\left|\right|,$ $E_{\mathrm{homo}}=e^{-{\mathrm{\γ}}_2|f(i,j)-{\mathrm{\μ}}_k|};$

In the formula, k ₁With k ₂Be weighting parameter;

γ ₁With γ ₂Be the constant greater than 0;

P is pixel; C is the particle center;

μ _kBe particle C _kGray average;

(x y) is image border figure to f.

2. according to the automatic division method of graininess object in the described a kind of digital picture of claim 1, it is characterized in that the direction gradient in the 3rd step is defined as: a certain pixel P (p in the image _x, p _y) point to a certain center C (c _x, c _y) the direction gradient vector be d (x, y)=(d _x, d _y), wherein $d_x=\frac{c_x-p_x}{\left|\right|P-C\left|\right|},$ $d_y=\frac{c_y-p_y}{\left|\right|P-C\left|\right|}.$

Images