CN104063880B - PSO based multi-cell position outline synchronous accurate tracking system - Google Patents

Abstract

The invention provides a particle swarm optimization based multi-cell position and outline synchronous accurate tracking system which comprises three main modules including a PSO based tracking module, a PSO based discovery module and a PSO based outline module. For existing-cell tracking, the PSO based tracking module obtains an initial state of a cell in a current frame on the basis of an existing-cell prior state, then calculates a cell outline through the PSO based outline module and meanwhile achieves accurate tracking by utilizing the iterative mass center updating process. For newly-emerging cell tracking, the PSO based discovery module discovers the position of a new cell in a whole image through appropriate particle swarm initialization and searching mechanism and further obtains the cell outline.

Description

A kind of many cells position profile synchronization accurate tracking system based on pso

Technical field

The invention provides a kind of many cells position based on pso and profile synchronization accurate tracking system, belong to cell with Track field.

Background technology

Cell is the ultimate unit of vital movement, is main shower in many biological processes, for any organic The fetal development of life, evolution and life maintain, and the propagation of cell, differentiation and migration are all requisite links.Therefore, right The research of cell behavior analysis is all very valuable in a lot of fields, opens including stem-cell research, tissue engineering, medicine Send out, genetics and proteomics etc..For example, early stage being inflamed, leukocyte can be in postcapillary venule Surface scrolls, its rolling speed can reflect the intensity of inflammatory reaction, disclose the celelular mechanism during these and contribute to inflammation machine The understanding of system and the exploitation of inflammation treatment medicine.Traditional analysis to cell behavior are the artificial sides being observed by professional Method or the semi-artificial method of interactive computer auxiliary.When a large amount of cells need to be traced in long-time, such method needs Want the image processing work of substantial amounts of user mutual, process is very uninteresting and time-consuming.Additionally, this kind of method probably introduces User's prejudice and loss important information, so, a kind of accurate method of exploitation automatically to follow the tracks of cell to be extremely important and has Meaning.In in the past few decades, with developing rapidly of data processing and computer vision technique, have developed many thin Born of the same parents' automatic tracking method.

In tracking cell field, realize the automatization of tracking and accuracy is faced with many challenges, essentially from two The factor of individual aspect, i.e. cellular factor and image factor.Cellular factor refers to that those occur the complexity in cell life cycle Situation and multiple intercellular interaction scenario, such as because the division of cell enters with dead or cell or leaves image-region Cause the change of cell quantity, the complicated cell topology such as cell shape change, neighbour and overlap, lack consistent in addition Cell movement model, these are all emphasis and the difficult points of tracking cell.Image factor typically refers to low picture quality, due to life The breathing of object and cause tremble, cell enters or leaves confocal plane and causes the change of contrast so that acquired figure As Quality Down that is to say, that microcytoscope image has low signal to noise ratio (snr) or contrast, comprise strong noise, in addition Also there is the very big factor of image data amount, increase the difficulty of tracking cell.

The cell automatic tracking method that presently, there are can be divided three classes, respectively based on detection and the track side associating Method, the tracking based on model evolution and the tracking based on filtering and sampling.Based on detection and the tracking cell associating Method comprises two key steps, that is, detect and associate.In the first step, respectively the cell detection in each frame out, same When obtain the state (how center, area etc.) of the number of cell and individual cells, conventional detection in tracking cell system Method includes threshold method, gradient method (rim detection), topology operation and watershed algorithm.In second step, in two continuous frames Or the cell that detects in multiframe is associated, so can be obtained by the movement locus of cell, additionally can calculate The kinematic parameter of cell such as instantaneous velocity and acceleration etc., association is typically based on a certain specific object function of optimization, for example Nearest neighbor method and smooth motion criterion.But in some cases, as cell density is very big in image, cell division event and sending out During raw segmentation errors, such tracking is likely to failure.Threshold method is the most frequently used detection method, but for vision system Strength Changes in system and picture noise, it is also a kind of method being easiest to occur mistake.Threshold method can not split phase mutual connection Tactile cell, and watershed transform provides a kind of effective solution of contact problems, its shortcoming be can because noise and It tends to produce over-segmentation phenomenon.

In the tracking cell method based on model evolution, first cell is modeled, then in successive frame, updates this Individual model and reach the purpose of tracking, the tracking result in former frame is used for the initialization of present frame relevant parameter, can represent The model of the outward appearance of cell or profile is developed in interframe like this.Method according to setting up model, in such tracking Algorithms most in use comprise Active contour, level set method, the tracking based on core or average drifting tracking.With Active contour As a example, the energy function of an internal energy related to cell outline and external energy sum defined in former frame, So that energy function is minimized in present frame, and then find out cell outline.Compared with watershed transform, due to the wheel of neighbour's cell Wide or profile is easily fused a cell in this evolutionary process, has generation image based on the tracking that profile develops Less divided phenomenon, it needs to carry out subsequent treatment to result.

Human visual system's estimating target motion in image sequence is imitated based on the tracking cell method of filtering and sampling Stream, can be by integrating high-dimensional space, time and prior information come solve problem, more preferably using time sequence information with studied The priori of cyto-dynamics feature.Particle filter (pf) is commonly used for tracking cell, but is estimating the current shape of target Before the Posterior distrbutionp of state, need to know measurement model and motion model.Juang is first mixed Gaussian probability hypothesis density letter Number (gm-phd) wave filter is applied to many cells and follows the tracks of, and finds that this wave filter can be well in the research to cell family tree Follow the tracks of out the pedigree of cell and the motion conditions of cell.Rezato fighi proposes an enclosed of lgjms-phd wave filter Solution, contains transition probability and the division transition probability of state independence, this wave filter considerably reduces thin in a large number for existing Process time when born of the same parents and detection noise.Reza proposes a kind of sequential Monte Carlo method of many bernoulli wave filter, the party Method does not need detection module, directly using the space time information extracting from low-quality image sequence, follows the tracks of before belonging to detection Technology.Although need to expend substantial amounts of amount of calculation, tracking based on filtering and sampling is compared to based on detecting and associate Tracking can better profit from space time information, particularly can obtain more robust in low-quality cell image data Tracking result.

Content of the invention

Present invention seek to address that under low comparison diagram image sequence the position of many cells and a Contour extraction difficult problem, that is, to how thin Born of the same parents' dynamicss have differences, many cells deform, cell number time-varying, the situation such as cell neighbour, are examining without cell Survey module, need not substantial amounts of cell training sample, by search and the optimization of population, set up rational location finding strategy and Skeleton pattern, solves position and the Contour extraction difficult problem of many cells.

In order to solve the above problems, the solution of proposition is the present invention: provides a kind of many based on particle cluster algorithm Cell position and profile synchronization accurate tracking system, comprise three main modular: the tracking module based on pso, sending out based on pso Existing module and the profile module based on pso；The described tracking module based on pso is in the base using the prior state that there is cell On plinth, obtain cell original state in the current frame；The described discovery module based on pso is initialized by suitable population The original state of neoblast is found with search mechanisms in whole image；Described obtained at the beginning of cell based on the profile module of pso Calculate the profile of cell on the basis of beginning state, reach accurate tracking using the barycenter renewal process of an iteration simultaneously.

Described system is to be divided into two classes for the cell in each frame image: there is cell and newly cell occurred, if There is cell in previous frame, then follow the tracks of these in the current frame first and there is cell, then search for new again in image Cell occurs.Described system, for the tracking that there is cell, for the cell that significant indexes are maximum, first passes through based on pso Tracking module obtain the original state of cell, then obtain accurate cell outline and position using the profile module based on pso Put, if successfully tracking this cell, cell exist with present frame in, otherwise this vanished cell；Described system is for newly thin The tracking of born of the same parents, first passes through the original state obtaining cell based on the discovery module of pso, then utilizes the profile module based on pso Obtain accurate cell outline and position, for tracking result, if can not be associated with there is cell (nearest neighbor method) then it is assumed that Find a neoblast, if continuous tracking result several times be clutter (false-alarm) it is believed that all cells in present frame Tracked.

Described is to combine product by cell area with profile information based on the cell significant indexes in the tracking module of pso Raw.

Described system is for the tracking that there is cell, using ordered tracking, the tracking module based on pso Concretely comprise the following steps:

1) initialization of population: known have m cell in t-1 frameWherein Represent cell k_t-1K () state (barycenter and profile) in t-1 frame, (begs for for wherein significant indexes in profile module By) maximum cell k_t-1K (), one population scale of initialization is n_tPopulation Particle state isWhereinWithIt is respectively the matter of the potential cell of i-th particle representative Heart abscissa, barycenter vertical coordinate, width and height, the original state of particle isWherein σ is a predefined diagonal matrix,Represent cell k_t-₁ (k) predicted state in t frame, WithIt is respectively cell k_t-1(k) corresponding state of global optimum's particle in t-1 and t-2 frame, grain The initial velocity of son is random and is distributed in

2) fitness function based on pso tracking module is (for simplifying formula, subscript t and subscript k_t-1K () is omitted):

$f_t\left(x_i^n\right)=\left\{\begin{array}{cc}\mathrm{\ϵ}& \mathrm{ifa}(x_i^n\∩\{c_{k_t\left(k\right)}^t\left\}\underset{k=1}{m}\right)>0\\ {\hat{f}}_{c_{k_{t-1\left(k\right)}}^t}\left(x_i^n\right)\×f_p\left(x_i^n\right)& \mathrm{otherwise}\end{array}\right.,$ Wherein, ε is a minimum,Represent particleThe rectangular area representing and identified cell in present frameWeight Folded area, f_p() is the similarity function of rgb image, ${\hat{f}}_{c_{k_{t-1}\left(k\right)}^t}\left(\hat{x}\right)=\frac{\mathrm{\κ}\left(\hat{x}\right)-\min \left(\mathrm{\κ}\right)}{\max \left(\mathrm{\κ}\right)-\min \left(\mathrm{\κ}\right)},$ k_f() is gaussian kernel functionσ By cell k_t-1K the radius of () obtains；Then particle is individual optimumWith global optimumFor: $p_i^n=\arg \underset{j=0,...,n}{\max }{f}_t\left(x_i^j\right),$ $p_g^n=\arg \underset{i=1,...,n_t}{\max }{f}_t\left(p_i^n\right).$

3) the pso algorithm revised: the speed of particle and state more new formula are (for simplifying formula, subscript t and subscript k_t-1 K () is omitted)

$v_i^{n+1}=\mathrm{\χ}(v_i^n+c_1{\mathrm{\η}}_1(p_i^n-x_i^n)+c_2{\mathrm{\η}}_2(p_g^n-x_i^n))$

$x_i^{n+1}=x_i^n+v_i^{n+1}$

Wherein,C=(c₁+c₂) ＞ 4.0, η₁,η₂∈ (0,1) is to produce in iteration each time Raw random number；

4) condition of convergence: populationThe average fitness of middle major part particle is more than threshold value(wherein, l and w is respectively image y_tLength and width, λ is adjustment factor), or reach To maximum iteration time mi, the output of population is designated asIfThen represent and trace into cellOtherwise, this vanished cell, and be marked as losing cell

Described is to search for the position of neoblast in global image based on the discovery module of pso, until continuous t_ndSecondary search Process does not find neoblast, concretely comprises the following steps:

1) initialization of population: one population scale of initialization is n in t two field picture_dPopulationParticle state isParticle is randomly distributed in image y_tIn, the initial velocity of particle with Machine and being distributed inJ=1 ..., 4；

2) fitness function of the discovery module based on pso: $f_d\left(x_i^{d,n}\right)=\left\{\begin{array}{cc}\mathrm{\ϵ}& \mathrm{ifa}(x_i^{d,n}\∩\{c_{k_t\left(k\right)}^t\left\}\underset{k=1}{m}\right)>0\\ f_p\left(x_i^{d,n}\right)& \mathrm{otherwise}\end{array}\right.,$ Wherein, ε is a minimum,Represent particleKnown in the rectangular area representing and present frame Other cellOverlapping area, f_p() is the similarity function of rgb image；

3) the pso algorithm revised: with reference to tracking module (3rd) step based on pso；The condition of convergence:Or PersonOr reach maximum iteration time mi, n₁It is a positive integer.The output of population is designated asAs ReallyThen thinkIt is a clutter (false-alarm), otherwiseIt is a real cell, ifCan not be with loss cellAssociated (arest neighbors correlation method), then be found to a neoblast, and be labeled as k_t(m)+1.

The described profile module based on pso utilizes the cell original state of the output of the first two module, the wheel of more neoblast Exterior feature and position, it concretely comprises the following steps:

1) initialization of population: the original state of cell isM direction of initializationWhereinOn every direction id, one population scale of initialization is n_cPopulationParticle state is d_i,id=d_i,id, represent particle i and cell barycenter [x₀,y₀] distance, according to cell size [w₀,h₀] initialization particle search space be s=[max (r_min,αr_c),min(r_max,βr_c)], wherein r_c=(w₀+h₀)/4, r_minAnd r_maxIt is the cell least radius and maximum radius observing in image sequence, α and β is adjustment factor, then particle Original stateParticle rapidity is initialized as 0；

2) fitness function based on pso profile module: Wherein, $f_d(x,y)=\mathrm{\δ}{\mathrm{\σ}}_{(x,y)}/\max \left(\mathrm{\δ}{\mathrm{\σ}}_{({.}_{,}.)}\right),\mathrm{\δ}{\mathrm{\σ}}_{(x,y)}=\frac{1}{\left|n_{(x,y)}\right|}{\mathrm{\σ}}_{(x^{\text{'}},y^{\text{'}})\&element;n(x,y)}{(i_{(x^{\text{'}},y^{\text{'}})}-{\stackrel{\&overbar;}{i}}_{(x,y)})}^2,$ n_(x,y)Represent pixel The 8 neighborhood territory pixel set of (x, y), | n_(x,y)| represent set n_(x,y)Middle element number, i_(·)For gray value,Represent set n_(x,y)Average gray value, individual optimum with global optimum's more new formula: $p_{i,\mathrm{id}}^{c,n}=\arg \underset{j=0,...,n}{\max }{f}_c\left(d_{i,\mathrm{id}}^j\right),$ $p_{g,\mathrm{id}}^{c,n}=\arg \underset{i=1,...,n_c}{\max }{f}_c\left(p_{i,\mathrm{id}}^{c,n}\right);$

3) improved pso algorithm: the speed of particle and state more new formula is

$d_{i,\mathrm{id}}^{n+1}=d_{i,\mathrm{id}}^n+v_{i,\mathrm{id}}^{c,n+1}$

Wherein, It is groupAdjacent group defeated Go out, for groupWhen not occurring barycenter to updateIt is pso in last all directions after occurring barycenter to update The meansigma methodss of Search Results, for other groups, $p_{\mathrm{id},\mathrm{nb}}^{c,\mathrm{output}}=\left\{\begin{array}{cc}{p}_{g,\mathrm{id}-1}^{c,\mathrm{output}}& 2\≤\mathrm{id}\≤m-1\\ (p_{g,1}^{c,\mathrm{output}}+p_{g,m-1}^{c,\mathrm{output}})/2& \mathrm{id}=m\end{array}\right.,$ WhereinRepresent group Output, η₁,η₂,η₃∈ (0,1) is the random number producing in each iteration, groupIteration mi_cOutput after secondary is as id Configuration sampling point on direction, then can obtain m sampled point

4) barycenter renewal process: calculate the not sampled point number on cell outlineIts Middle th is in f_dThe calculated threshold value of upper utilization otsu method,It is converted into 2 dimension state (x_g,id,y_g,id), if [x₀,y₀] in extracellular, i.e. n_outMore than a predefined threshold value, update centroid position according to the following formula,Wherein, $[\hat{x},\hat{y}]=\underset{\mathrm{id}=1}{\overset{m}{\mathrm{\σ}}}\left((x_{g,\mathrm{id}}-x_0,y_{g,\mathrm{id}}-y_0)f_c\left(p_{g,\mathrm{id}}^{c,\mathrm{output}}\right)\right)/\underset{\mathrm{id}=1}{\overset{m}{\mathrm{\σ}}}{f}_c\left(p_{g,\mathrm{id}}^{c,\mathrm{output}}\right),$ If [x₀,y₀] portion in the cell, centroid position more new formula isEach time more Newly [x₀,y₀] after, need to re-use improved pso algorithm search and go out configuration sampling point, until the side-play amount of barycenter is very small When stop update.Final configuration sampling point isIt is linked in sequence after these are put and obtain complete cell wheel Wide It is the ith pixel point on profile, the significant indexes of cell areWherein a (p_c) represent profile p_cThe cell area representing, f_c() is fitness function.

The system being provided using the present invention, is had the following characteristics that

1) the pso tracking system that the present invention provides can accurately and synchronously follow the tracks of position and the profile of many cells.

2) present invention solves the problems, such as many cells position and Contour extraction using pso algorithm first, unknown for multiple numbers And cell, the cell of different dynamic characteristic, deformation cell and the neighbour's cell changing, reach accurate tracking effect.

3) tracking stability of system provided by the present invention is high, and false alarm rate, loss are low.How thin with many bernoulli wave filter Born of the same parents' tracking is compared with ant system many cells tracking, has higher precision and tracking stability, takes between two Between person.

Brief description

Fig. 1 the system structure chart；

The original sequence of Fig. 2 tracking cell；

There is tracking cell process in Fig. 3；

Fig. 4 neoblast follows the tracks of process；

Fig. 5 tracking cell result (1) cellular sequences 1, (2) cellular sequences 2；

Fig. 6 cell position estimated result (1) cellular sequences 1, (2) cellular sequences 2；

Fig. 7 cell area result (1) cellular sequences 1, (2) cellular sequences 2；

Fig. 8 root-mean-square error compares (left: sequence 1；Right: sequence 2)；

Specific embodiment

Fig. 1 is the structure chart of the system.As shown in figure 1, germinal cell (t cell) each frame of image sequence for Fig. 2 Original image, there is cell in that first follows the tracks of former frameBy the tracking module based on pso with based on pso's There is cell to each and be tracked in profile module, obtain cell position in the current frame and profile successively.Afterwards, then The new appearance repeatedly found in the picture in present frame by the discovery module based on pso and the profile module based on pso is thin Born of the same parents.

There is the tracking of cell in embodiment 1

As shown in Fig. 3 (a), in previous frame, there are 4 cells, need in the current frame according to cell significant indexes according to These cells are tracked, first with the tracking module based on pso, step is secondaryly:

1) initialization of population: known have m cell in t-1 frameWhereinTable Show cell k_t-₁K () state (barycenter and profile) in t-1 frame, (begs for for wherein significant indexes in profile module By) maximum cell k_t-₁K (), one population scale of initialization is n_tPopulation Particle state isWhereinWithIt is respectively the potential thin of i-th particle representative The barycenter abscissa of born of the same parents, barycenter vertical coordinate, width and height, the original state of particle isWherein σ is a predefined diagonal matrix,Represent cell k_t-₁(k) Predicted state in t frame, WithIt is respectively cell k_t-1(k) corresponding state of global optimum's particle in t-1 and t-2 frame, particle initial Speed is random and is distributed inAs shown in Fig. 3 (b), Follow the tracks of initialization population distribution during cell 3.

2) fitness function based on pso tracking module is (for simplifying formula, subscript t and subscript k_t-1K () is omitted):

$f_t\left(x_i^n\right)=\left\{\begin{array}{cc}\mathrm{\ϵ}& \mathrm{ifa}(x_i^n\∩\{c_{k_t\left(k\right)}^t\left\}\underset{k=1}{m}\right)>0\\ {\hat{f}}_{c_{k_{t-1\left(k\right)}}^t}\left(x_i^n\right)\×f_p\left(x_i^n\right)& \mathrm{otherwise}\end{array}\right.,$ Wherein, ε is a minimum,Represent particleThe rectangular area representing and identified cell in present frameWeight Folded area, f_p() is the similarity function of rgb image, ${\hat{f}}_{c_{k_{t-1}\left(k\right)}^t}\left(\hat{x}\right)=\frac{\mathrm{\κ}\left(\hat{x}\right)-\min \left(\mathrm{\κ}\right)}{\max \left(\mathrm{\κ}\right)-\min \left(\mathrm{\κ}\right)},$ k_f() is gaussian kernel functionσ by Cell k_t-1K the radius of () obtains；Then particle is individual optimumWith global optimumFor: $p_i^n=\arg \underset{j=0,...,n}{\max }{f}_t\left(x_i^j\right),$ $p_g^n=\arg \underset{i=1,...,n_t}{\max }{f}_t\left(p_i^n\right).$

3) the pso algorithm revised: the speed of particle and state more new formula are (for simplifying formula, subscript t and subscript k_t-1 K () is omitted)

$v_i^{n+1}=\mathrm{\χ}(v_i^n+c_1{\mathrm{\η}}_1(p_i^n-x_i^n)+c_2{\mathrm{\η}}_2(p_g^n-x_i^n))$

$x_i^{n+1}=x_i^n+v_i^{n+1}$

Wherein,C=(c₁+c₂) ＞ 4.0, η₁,η₂∈ (0,1) is to produce in iteration each time Raw random number；

5) condition of convergence: populationThe average fitness of middle major part particle is more than threshold value(wherein, l and w is respectively image y_tLength and width, λ is adjustment factor), or reach To maximum iteration time mi, the output of population is designated asIfThen represent and trace into cellOtherwise, this vanished cell, and be marked as losing cellAs shown in Fig. 3 (c), the end-state of all particles And population optimum state.

Then, the profile module based on pso utilize tracking module optimal particle state, obtain the accurate position of cell with Profile, step is:

1) initialization of population: the original state of cell isM direction of initializationWhereinOn every direction id, one population scale of initialization is n_cPopulationParticle state is d_i,id=d_i,id, represent particle i and cell barycenter [x₀,y₀] distance, according to cell size [w₀,h₀] initialization particle search space be s=[max (r_min,αr_c),min(r_max,βr_c)], wherein r_c=(w₀+h₀)/4, r_minAnd r_maxIt is the cell least radius and maximum radius observing in image sequence, α and β is adjustment factor, then particle Original stateParticle rapidity is initialized as 0；

2) fitness function based on pso profile module: Wherein, $f_d(x,y)=\mathrm{\δ}{\mathrm{\σ}}_{(x,y)}/\max \left(\mathrm{\δ}{\mathrm{\σ}}_{({.}_{,}.)}\right),\mathrm{\δ}{\mathrm{\σ}}_{(x,y)}=\frac{1}{\left|n_{(x,y)}\right|}{\mathrm{\σ}}_{(x^{\text{'}},y^{\text{'}})\&element;n(x,y)}{(i_{(x^{\text{'}},y^{\text{'}})}-{\stackrel{\&overbar;}{i}}_{(x,y)})}^2,$ n_(x,y)Represent pixel The 8 neighborhood territory pixel set of (x, y), | n_(x,y)| represent set n_(x,y)Middle element number, i_(·)For gray value,Represent set n_(x,y)Average gray value, individual optimum with global optimum's more new formula: $p_{i,\mathrm{id}}^{c,n}=\arg \underset{j=0,...,n}{\max }{f}_c\left(d_{i,\mathrm{id}}^j\right),$ $p_{g,\mathrm{id}}^{c,n}=\arg \underset{i=1,...,n_c}{\max }{f}_c\left(p_{i,\mathrm{id}}^{c,n}\right);$

3) improved pso algorithm: the speed of particle and state more new formula is

$d_{i,\mathrm{id}}^{n+1}=d_{i,\mathrm{id}}^n+v_{i,\mathrm{id}}^{c,n+1}$

Wherein, It is groupAdjacent group defeated Go out, for groupWhen not occurring barycenter to updateIt is pso in last all directions after occurring barycenter to update The meansigma methodss of Search Results, for other groups, $p_{\mathrm{id},\mathrm{nb}}^{c,\mathrm{output}}=\left\{\begin{array}{cc}{p}_{g,\mathrm{id}-1}^{c,\mathrm{output}}& 2\≤\mathrm{id}\≤m-1\\ (p_{g,1}^{c,\mathrm{output}}+p_{g,m-1}^{c,\mathrm{output}})/2& \mathrm{id}=m\end{array}\right.,$ WhereinRepresent group Output, η₁,η₂,η₃∈ (0,1) is the random number producing in each iteration, groupIteration mi_cOutput after secondary is as id Configuration sampling point on direction, then can obtain m sampled point

5) barycenter renewal process: calculate the not sampled point number on cell outlineIts Middle th is in f_dThe calculated threshold value of upper utilization otsu method,It is converted into 2 dimension state (x_g,id,y_g,id), if [x₀,y₀] in extracellular, i.e. n_outMore than a predefined threshold value, update centroid position according to the following formula,Wherein, $[\hat{x},\hat{y}]=\underset{\mathrm{id}=1}{\overset{m}{\mathrm{\σ}}}\left((x_{g,\mathrm{id}}-x_0,y_{g,\mathrm{id}}-y_0)f_c\left(p_{g,\mathrm{id}}^{c,\mathrm{output}}\right)\right)/\underset{\mathrm{id}=1}{\overset{m}{\mathrm{\σ}}}{f}_c\left(p_{g,\mathrm{id}}^{c,\mathrm{output}}\right),$ If [x₀,y₀] portion in the cell, centroid position more new formula isEach time more Newly [x₀,y₀] after, need to re-use improved pso algorithm search and go out configuration sampling point, until the side-play amount of barycenter is very small When stop update.Final configuration sampling point isIt is linked in sequence after these are put and obtain complete cell outline It is the ith pixel point on profile, the significant indexes of cell areWherein a (p_c) represent profile p_cThe cell area representing, f_c() is fitness function.Fig. 3 D () and Fig. 3 (e) are respectively the cell outline sampled point after updating before barycenter updates with barycenter, these dot sequencies are connected and obtains Complete cell outline (as shown in Fig. 3 (f)).

By above-mentioned 5 steps, complete the tracking (as shown in Fig. 3 (g)) that other have had cell.

Newly the tracking of cell in embodiment 2

For the tracking that cell newly occurs in present frame, by the discovery module based on pso and the profile module based on pso Realize.The step of the discovery module based on pso is:

1) initialization of population: one population scale of initialization is n in t two field picture_dPopulationParticle state isParticle is randomly distributed in image y_tIn, the initial velocity of particle with Machine and being distributed inJ=1 ..., 4, shown in initial distribution such as Fig. 4 (a) of population.

2) fitness function of the discovery module based on pso: $f_d\left(x_i^{d,n}\right)=\left\{\begin{array}{cc}\mathrm{\ϵ}& \mathrm{ifa}(x_i^{d,n}\∩\{c_{k_t\left(k\right)}^t\left\}\underset{k=1}{m}\right)>0\\ f_p\left(x_i^{d,n}\right)& \mathrm{otherwise}\end{array}\right.,$ Wherein, ε is a minimum,Represent particleKnown in the rectangular area representing and present frame Other cellOverlapping area, f_p() is the similarity function of rgb image；

3) the pso algorithm revised: with reference to tracking module (3rd) step based on pso；The condition of convergence:Or PersonOr reach maximum iteration time mi, n₁It is a positive integer.The output of population is designated asAs ReallyThen thinkIt is a clutter (false-alarm), otherwiseIt is a real cell, ifCan not be with loss cellAssociated (arest neighbors correlation method), then be found to a neoblast, and be labeled as k_t(m)+1.Fig. 4 (b) is the convergence state of population, and optimal particle state goes out for a potential cell.

Then, by previously described, cell outline is obtained based on the profile module of pso, such as Fig. 4 (c-d).By above-mentioned Step traces into all new cell, shown in such as Fig. 4 (e).

Fig. 5 provides the tracking result of two groups of cell image sequences it can be seen that the system can synchronous accurate tracking many cells Position and profile, and cell quantity can be processed can change, cell appearance changes, and cell neighbour etc. complicated with Track problem.

Fig. 6 is to provide cell centroid position, and Fig. 7 provides cell area.

The system with many bernoulli wave filter compared with ant system: the system is compared to it as can be seen from Table 1 He has more accurate tracking performance by two kinds of many cells trackings, has lower loss (fnr), false alarm rate (far), mark Sign conversion ratio (lsr) and follow the tracks of Loss Rate (ltr)；Fig. 8 compare for root-mean-square error (rmse) it can be seen that the system compared to Other two methods have more accurate tracking performance, and root-mean-square error is less；Table 2 be follow the tracks of time-consuming compare it can be seen that this System takes between, can meet requirement of real-time.

Table 1

Table 2

Method	Sequence 1 (second/frame)	Sequence 2 (second/frame)
			Many bernoulli wave filter	1.63	1.84
Ant system	20.25	11.83
			The system	5.48	6.28

Tracking performance index definition:

1) in loss (fnr): missing inspection cell number and whole frame, actual cell tires out the ratio of note sum.

2) in false alarm rate (far): false-alarm number and whole frame, actual cell tires out the ratio of note sum.

3) the label handover event number of label conversion ratio (lsr): cell is total with actual cell neighbour's event in whole frames The ratio of number.

4) follow the tracks of Loss Rate (ltr): the frame number that tracking is lost exceedes it and there is the cell number of frame number 50% and reality The ratio of cell marking number.

5) root-mean-square error (rmse): ${\mathrm{rmse}}_t=\sqrt{\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\σ}}}{\mathrm{rmse}}_{i.t}^2},$ Wherein ${\mathrm{rmse}}_{i,t}^2=\frac{1}{m}\underset{k=1}{\overset{m}{\mathrm{\σ}}}{\left|\right|x_{k,t}-{\hat{x}}_{k,t}\left|\right|}^2,$ N represents tracking number of times, and m represents the total cellular score in t frame, x_k,tWithIt is respectively the actual position of cell k and follow the tracks of knot Really.

Although the present invention is open as above with preferred embodiment, it is not limited to the present invention, any is familiar with this skill The people of art, without departing from the spirit and scope of the present invention, can do various changes and modification, therefore the protection model of the present invention Enclosing should be by being defined that claims are defined.

Claims (6)

1. a kind of many cells position based on pso and profile synchronization accurate tracking system are it is characterised in that include based on pso's Tracking module, the discovery module based on pso and the profile module based on pso；The described tracking module based on pso is using depositing On the basis of the prior state of cell, obtain cell original state in the current frame；The described discovery module based on pso is led to Cross population initialization and search mechanisms find the original state of neoblast in whole image；The described profile die based on pso Block calculates the profile of cell on the basis of obtaining cell original state, is reached using the barycenter renewal process of an iteration simultaneously Follow the tracks of to accurate.

2. many cells position according to claim 1 and profile synchronization accurate tracking system is it is characterised in that described system It is that two classes are divided into for the cell in each frame image: there is cell and newly cell has occurred, if existing thin in previous frame Born of the same parents, then follow the tracks of these in the current frame first and there is cell, then search in image again and cell newly occurs；Described system For the tracking that there is cell, for the cell that significant indexes are maximum, first pass through and cell is obtained based on the tracking module of pso Original state, then obtain accurate cell outline and position using the profile module based on pso, if it is thin to successfully track this Born of the same parents, cell persists in present frame, otherwise this vanished cell；Described system, for the tracking of neoblast, first passes through and is based on The discovery module of pso obtains the original state of cell, then using based on the profile module of pso obtain accurate cell outline with Position, for tracking result, if can not be associated with cell there is then it is assumed that finding a neoblast, if continuously several times Tracking result be clutter it is believed that all cells in present frame are tracked.

3. many cells position according to claim 2 and profile synchronization accurate tracking system are it is characterised in that described be based on Cell significant indexes in the tracking module of pso are to combine generation by cell area with profile information.

4. many cells position according to claim 1 and profile synchronization accurate tracking system are it is characterised in that described be based on The concretely comprising the following steps of the tracking module of pso:

1) initialization of population: known have m cell in t-1 frameWhereinRepresent thin Born of the same parents k_t-1K () state in t-1 frame, state is made up of barycenter and profile, for the cell that wherein significant indexes are maximum k_t-1K (), one population scale of initialization is n_tPopulationParticle state isWherein WithThe barycenter abscissa of the potential cell of respectively i-th particle representative, barycenter vertical coordinate, wide Spend and height, the original state of particle isWherein σ is predefined one Diagonal matrix,Represent cell k_t-1(k) predicted state in t frame, WithIt is respectively Cell k_t-1K () corresponding state of global optimum's particle in t-1 and t-2 frame, the initial velocity of particle is random and is distributed in

2) fitness function based on pso tracking module:

Wherein, for simplify formula, subscript t and under Mark k_t-1K () is omitted；ε is a minimum,Represent particleThe rectangular area representing With identified cell in present frameOverlapping area, f_p() is the similarity function of rgb image,k_f () is gaussian kernel functionσ is obtained by the radius of cell kt-1 (k)；

Then particle is individual optimumWith global optimumFor:

$p_i^n=\arg \underset{j=0,\mathrm{...},n}{\max }{f}_t\left(x_i^j\right)$

$p_g^n=\arg \underset{i=1,\mathrm{...},n_t}{\max }{f}_t\left(p_i^n\right)$

Wherein,η₁, η₂∈ (0,1) is to produce in iteration each time Raw random number；

4) condition of convergence: populationThe average fitness of middle major part particle is more than threshold valueWherein, l and w is respectively image y_tLength and width, λ be adjustment factor, or Person reaches maximum iteration time mi, and the output of population is designated asIfThen represent with Track is to cellOtherwise, this vanished cell, and be marked as losing cell

5. many cells position according to claim 1 and profile synchronization accurate tracking system are it is characterised in that described be based on The concretely comprising the following steps of the discovery module of pso:

1) initialization of population: one population scale of initialization is n in t two field picture_dPopulationParticle state isParticle is randomly distributed in image y_tIn, grain The initial velocity of son is random and is distributed in

2) fitness function of the discovery module based on pso:

Wherein, ε is a minimum,Represent particleThe rectangular area representing and identified cell in present frameOverlapping area, f_p() is the similarity function of rgb image；

3) the pso algorithm revised: with reference to tracking module (3rd) step based on pso；

The condition of convergenceOrOr reach maximum iteration time mi, n₁Be one just Integer, the output of population is designated asIfThen thinkIt is one Clutter false-alarm, otherwiseIt is a real cell, ifCan not be with loss cell Associated, then it is found to a neoblast, and be labeled as k_t(m)+1；It is wherein to be judged using arest neighbors correlation methodWith loss cellWhether it is associated.

6. many cells position according to claim 1 and profile synchronization accurate tracking system are it is characterised in that described be based on The concretely comprising the following steps of the profile module of pso:

1) initialization of population: the original state of cell isM direction of initializationWhereinOn every direction id, one population scale of initialization is n_c PopulationParticle state is d_i,id=d_i,id, represent particle i and cell barycenter [x₀, y₀] away from From according to cell size [w₀, h₀] initialization particle search space be s=[max (r_min, α r_c), min (r_max, β r_c)], its Middle r_c=(w₀+h₀)/4, r_minAnd r_maxIt is the cell least radius and maximum radius observing in image sequence, α and β is to adjust Section coefficient, then the original state of particleParticle rapidity is initialized as 0；

2) fitness function based on pso profile module:

Wherein, n_{(x, y)}Represent 8 neighborhood territory pixel set of pixel (x, y), | n_(x,y)| represent set n_(x,y)Middle element number, i_(·)For gray value,Represent set n_(x,y)Average gray value, individual optimum with global optimum's more new formula:

$p_{i,id}^{c,n}=\arg \underset{j=0,\mathrm{...},n}{\max }{f}_c\left(d_{i,id}^j\right)$

$p_{g,id}^{c,n}=\arg \underset{i=1,\mathrm{...},n_c}{\max }{f}_c\left(p_{i,id}^{c,n}\right)$

3) improved pso algorithm: the speed of particle and state more new formula is

$d_{i,id}^{n+1}=d_{i,id}^n+v_{i,id}^{c,n+1}$

Wherein,It is groupAdjacent group output, for groupWhen not occurring barycenter to updateBarycenter is occurred to update It is the meansigma methodss of pso Search Results in last all directions afterwards, for other groups,WhereinRepresent groupOutput, η 1, η 2, η 3 ∈ (0,1) are the randoms number producing in each iteration, groupIteration mi_cOutput after secondary is as the wheel on id direction Wide sampled point, then can obtain m sampled point

4) barycenter renewal process: calculate the not sampled point number on cell outlineWherein th is in f_dThe calculated threshold value of upper utilization otsu method,It is converted into 2 dimension state (x_g,id, y_g,id), if [x₀, y₀] in extracellular, i.e. n_outMore than a predefined threshold Value, updates centroid position according to the following formula

$\[x_0,y_0\]\←\[x_0+\hat{x},y_0+\hat{y}\]$

Wherein,If [x₀, y₀] portion in the cell, centroid position more new formula is? Update [x each time₀, y₀] after, need to re-use improved pso algorithm search and go out configuration sampling point, until the skew of barycenter Stop when amount is very small updating；

Final configuration sampling point isIt is linked in sequence after these are put and obtain complete cell outlineIt is the ith pixel point on profile, the significant indexes of cell areWherein a (p_c) represent profile p_cThe cell area in besieged city, f_c() is fitness Function.