CN105913084A - Intensive track and DHOG-based ultrasonic heartbeat video image classifying method - Google Patents

Abstract

The invention relates to an intensive track and DHOG-based ultrasonic heartbeat video image classifying method which comprises the following steps: ultrasonic heartbeat video image frames are subjected to intensive sampling operation, coverage of characteristic points in video frames can be increased, a track tracing method is adopted for tracking the characteristic points in the video frames when the characteristic points are converted into a three dimension space, a track obtained from tracing operation is described via a DHOG and HOF and MBH combination method, and therefore characteristic vectors can be obtained. A problem that conventional ultrasonic heartbeat video image classifying technologies are low in classifying accuracy and cannot satisfy requirements for computer aided cardiovascular diagnostic technologies can be addressed.

Description

A kind of ultrasonic cardiography video image sorting technique based on intensive track and DHOG

Technical field

The present invention relates to the Feature Extraction Technology in ultrasonic cardiography video image categorizing system, particularly a kind of based on The ultrasonic cardiography video image sorting technique of intensive track and DHOG.

Background technology

Current ultrasonic cardiography video image sorting technique is based on 3D-SIFT feature extraction and describes algorithm to ultrasonic Characteristic point in video image frame aroused in interest is extracted and then is described, and uses tradition SVM classifier pair simultaneously It is classified, but nowadays the extensive of computer-aided medical diagnosis system is applied and the most medical and medical The accuracy of diagnosis is increasingly paid attention to.

With current 3D-SIFT feature extraction and tradition SVM classifier technology, ultrasonic cardiography video image is entered For row classification, SIFT algorithm is promoted to three dimensions be described by 3D-SIFT from two-dimensional space Algorithm, it is not applied to the incidence relation between frame of video and frame, because heart is one becomes periodically fortune Dynamic object, if the association track effectively utilizing its kinetic characteristic to move characteristic point between different frame enters Line description will be greatly improved the effectiveness of feature extraction, describe simultaneously the most convenient effectively.3D-SIFT The maximum shortcoming of technology is that classification accuracy is low, classification effectiveness is low, this to a certain extent image it is special Levy accuracy and the effectiveness of extraction, ultimately result in the low and slow-footed result of classification accuracy, direct shadow Efficiency and the accuracy of medical science auxiliary diagnosis are rung, it is impossible to meet the application of medical science auxiliary diagnosis.

Summary of the invention

It is an object of the invention to provide the classification of a kind of ultrasonic cardiography video image based on intensive track and DHOG Method, to overcome prior art classification accuracy poor, the problem that classification effectiveness is low.

For achieving the above object, the technical scheme is that a kind of based on intensive track and DHOG ultrasonic Video image sorting technique aroused in interest,

Step S1: the two field picture of ultrasonic cardiography video image is carried out intensive sampling；

Step S2: remove the homogenizing part in the characteristic point region that intensive sampling obtains；

Step S3: in subsequent ultrasonic video image aroused in interest frame, the track of characteristic point is tracked；

Step S4: use descriptor that the track of characteristic point is described；

Step S5: will obtain through described step S4 that ultrasonic cardiography video image characteristic vector is encoded to be input to Corresponding grader, completes the classification to ultrasonic cardiography video image.

In an embodiment of the present invention, in described step S1, at a mesh space containing W1 pixel In carry out the sampling of intensive characteristic point；Sampling scale according to each metric space according to the factorConstantly increase Add, altogether eight metric spaces of sampling.

In an embodiment of the present invention, in described step S2, when the characteristic point in described characteristic point region When locally the eigenvalue of autocorrelation matrix is less than threshold value T, the point just this feature point being judged as in average region, And remove it, and threshold value T is:

$T=0.001\×\underset{i\∈I_t}{max}min({\mathrm{\λ}}_i^1,{\mathrm{\λ}}_i^2)$

Wherein,It it is frame of video I_tTwo eigenvalues of the correlation matrix M of middle characteristic point i, 0.001 is Intermediate value after the significance and density of balanced sample point.

In an embodiment of the present invention, in described step S3, by two two field pictures adjacent before and after estimating Motion model between characteristic point obtains intensive optical flow field；Use the pixel value of neck near field, obtained by extension The estimated value of motion model；Draw intensive optical flow field algorithm according to optical flow field method, and characteristic point is tracked.

In an embodiment of the present invention, step S31: the spreading coefficient of two adjacent images is expressed as A₁(x)、b₁(x)、c₁(x) and A₂(x)、b₂(x)、c₂(x), wherein, A₁(x) and A₂X () is the most unequal, And approximate:

$A\left(x\right)=\frac{A_1\left(x\right)+A_2\left(x\right)}{2}$

And Δ b (x)=-0.5 (b₂(x)-b₁(x)), and then try to achieve constraints A (x) d (x)=Δ b (x), wherein, d (x) The displacement vector being as the change of locus and change；

Step S32: solve d (x) according to the neighborhood pixels of characteristic point in two field picture, sets up the minimum in neighborhood and puts down Mode:

$\underset{\mathrm{\Δ}x\∈I}{\Σ}w\left(\mathrm{\Δ}x\right)\left|\right|A(x+\mathrm{\Δ}x)d\left(x\right)-\mathrm{\Δ}b(x+\mathrm{\Δ}x)|{|}^2$

In order to improve the robustness of displacement vector, use eight movement parameter models that displacement vector carries out parametrization:

d_x(x, y)=a₁+a₂x+a₃y+a₇x²+a₈xy

d_y(x, y)=a₄+a₅x+a₆y+a₇xy+a₈y²

Finally give:

$P={\left(\underset{i}{\Σ}{w}_i{S}_i^T{A}_i^T{A}_i{S}_i\right)}^{-1}\underset{i}{\Σ}{w}_i{S}_i^T{A}_i^T{\mathrm{\Δb}}_i;$

Wherein, the coordinate of the pixel of two field picture in i is neighborhood；

Step S33: introduce a priori displacement, makes the priori displacement beObtain actual value and prior estimate Relative displacement between value, obtains:

$A\left(x\right)=\frac{A_1\left(x\right)+A_2\left(\tilde{x}\right)}{2};$

$\mathrm{\Δ}b\left(x\right)=-0.5\left(b_2\right(\tilde{x})-b_1(x\left)\right)+A\left(x\right)\tilde{d}\left(x\right);$

$\tilde{x}=x+\tilde{d}\left(x\right);$

Step S34: be iterated calculating as priori displacement next time using previous estimation displacement, and For the first time during iteration, priori displacement is set to 0.

In an embodiment of the present invention, in described step S4, by combining HOF, algorithm, MBH are described Describe algorithm and DHOG describes algorithm and jointly describes feature point trajectory, to calculate feature point trajectory Description vectors.

In an embodiment of the present invention, described HOF describes the computational methods of algorithm and realizes in accordance with the following steps:

Step S411: Video segmentation is become local space time's SPACE V that several are the same, if continuous print on time shaft One space-time space of dry two field picture block composition；Then on image, by any point, (x y) can obtain its light stream VectorWherein, v_x(x, y) and v_y(x y) represents that light stream is in x-axis respectively Optical flow components with y-axis；Therefore pixel (x, y) size of the light stream at place is:

$F(x,y)=\sqrt{v_x^2(x,y)+v_y^2(x,y)}$

The direction of its light stream is:

$\mathrm{\α}(x,y)=arctan\left(\frac{v_y(x,y)}{v_x(x,y)}\right);$

Step S412: [0,360 °) light flow path direction decile be divided into b sector, the most each pixel Range value on component interval is:

$G_k(x,y)=\left\{\begin{array}{cc}F(x,y)& \mathrm{\α}(x,y)\∈b_k\\ 0& \mathrm{\α}(x,y)\∉b_k\end{array}\right.;$

Step S413: increase a null value region in HOF vector, by the HOF in 8 traditional regions Vector is promoted to 9 regions, for preserving the number of pixel, i.e. 9 dimensions.

In an embodiment of the present invention, described MBH describes the computational methods of algorithm and describes algorithm based on HOF, It is horizontal and vertical component by intensive light stream orthogonal division, in space to its derivation and to calculate its light stream respectively straight Fang Tu, thus obtain two orthogonal rectangular histograms of MBHx and MBHy.

In an embodiment of the present invention, described DHOG describes the computational methods of algorithm and comprises the steps:

Step S421: two field picture is normalized；

Step S422: image is carried out gamma correction；

Step S423: be divided into S interval by 0～360 °, wherein S is even number, in each cell Pixel vote, thus obtain its gradient orientation histogram H_g:

$H_{og}\left(i\right)=H_g\left(i\right)+H_g(i+\frac{S}{2}),1\≤i\≤\frac{S}{2}$

Wherein, H_og(i) and H_gI () represents HOG and H respectively_gIn i-th element value；In order to improve biography System HOG feature description ability to express, after HOG feature descriptor again series connection on one group of brand-new ladder Degree rectangular histogram H_ng, and then obtain DHOG gradient vector, the length of this histogram vectors is identical with HOG, Each of which element value obtains expression formula:

$H_{ng}\left(i\right)=|H_g\left(i\right)-H_g(i+\frac{S}{2})|,1\≤i\≤\frac{S}{2}$

Step S424: the DHOG characteristic vector in region is carried out contrast normalization；

In the pixel region of the 32 × 32 of step S425:DHOG characteristic vector only computation-intensive track HOG descriptor, the size of current block is the pixel region of 32 × 32, wherein, has 2 × 2 lists in each piece Unit's lattice, are combined intensive track frame, retouch the DHOG feature of each cell in this combined frames State symbol to sue for peace according to corresponding rectangular histogram, and use L₂Norm is normalized, and in this combined frames The dimension of DHOG feature descriptor is 16 dimensions.

Compared to prior art, the method have the advantages that

1, intensive sampling technology and track are described algorithm to be applied in ultrasonic cardiography video image, largely The classification accuracy of the ultrasonic cardiography video image of upper raising and efficiency.

2, use three kinds of tracks to describe the method that algorithm combines, each characteristic factor of feature point trajectory is entered Line description encodes, it is ensured that the effectiveness of feature point trajectory.

3, traditional HOG is described algorithm to be improved, it is proposed that DHOG algorithm.To HOG algorithm super Apply the some shortcomings of existence to be improved on sound video image aroused in interest, improve algorithm for illumination noise and The robustness judged with the feature of size opposite direction moving object.

Accompanying drawing explanation

Fig. 1 is ultrasonic cardiography video image based on intensive track and DHOG feature description classification mould in the present invention Type schematic diagram.

Fig. 2 is the whole metric spaces covered of sampling in the present embodiment.

Fig. 3 is intensive trajectory extraction schematic diagram in the present embodiment.

Fig. 4 is that in the present embodiment, DHOG characteristic vector generates process flow diagram flow chart.

Detailed description of the invention

Below in conjunction with the accompanying drawings, technical scheme is specifically described.

For current ultrasonic cardiography video image sorting technique classification accuracy rate on relatively low, it is impossible to reach meter The requirement of calculation machine auxiliary cardiovascular diagnosis technology, therefore, the present invention is directed to this problem and devises a kind of based on close Collection track and the ultrasonic cardiography video image sorting technique of DHOG feature description, to ultrasonic cardiography video image frame Carry out intensive sampling, to increase the coverage rate of characteristic point in frame of video, use the method for trajectory track to frame of video In characteristic point be tracked in being transformed into three dimensions, the side finally combined with DHOG, HOF and MBH The track that tracking obtains is described and obtains characteristic vector by method.Present invention is generally directed to ultrasonic cardiography video image Feature Extraction Technology in categorizing system improves explanation, will illustrate the implementation process of the present invention below.

Concrete, in the present embodiment, as it is shown in figure 1, comprise the steps:

Step S1: the two field picture of ultrasonic cardiography video image is carried out intensive sampling；

Step S2: homogenizing part is removed in the characteristic point region obtaining sampling；

Step S3: characteristic point track in subsequent ultrasonic video image aroused in interest frame is tracked；

Step S4: use descriptor to be described the track of characteristic point；

Step S5: obtain ultrasonic cardiography video image characteristic vector.

In the present embodiment, after obtaining ultrasonic cardiography video image, first have to first to the point of interest in frame of video Sample, obtain our helpful characteristic point information.In this embodiment, use the side of intensive sampling Method, intensive sampling can be good at guaranteeing to cover all characteristic points in ultrasonic cardiography video image.Contain at one Having in the mesh space of W1 pixel and carry out the sampling of intensive characteristic point, such sampling needs to cover whole Metric space, as shown in Figure 2.Such sample mode ensure that all of yardstick of uniform covering and space-time Position.

Further, it is thus necessary to determine that the value of W1, when W1 take bigger time, it is fast that characteristic point takes spot speed, But take and count less, the characteristic point of some keys may be missed, when W1 take less time, characteristic point takes Count more, most feature can be comprised, but it is longer to take a duration, be unfavorable for the timeliness of whole system Property.It is preferred that in originally implementing, the value of W1 is set to 5 by experiment.The most both can guarantee that system Ageing, again accuracy rate is not resulted in the biggest impact.As for the number of sampling scale, typically by video Resolution determine, each metric space is according to the factorConstantly increase, altogether eight chis of sampling Degree space.

In the present embodiment, the target of this algorithm is to follow the trail of all sampled points in image.But, at video shadow In Xiang, some average image-regions do not have any image structure information, during video, these points Point in these regions, substantially without changing, is tracked being the most in all senses by position.Therefore, Need to remove, to improve the efficiency of whole system and to reduce computation complexity the point in these regions.

In the present embodiment, be whether that the judgement in average region is as follows for characteristic point: in the middle of video image certain When the eigenvalue of the local autocorrelation matrix of certain characteristic point in one frame is less than specific threshold value T, just will Point that this point is judged as in average region also removes it.The threshold formula that the present invention sets as:

$T=0.001\×\underset{i\∈I_t}{max}min({\mathrm{\λ}}_i^1,{\mathrm{\λ}}_i^2)$

In formulaIt it is frame of video I_tTwo eigenvalues of the correlation matrix M of middle characteristic point i, and 0.001 This value is the intermediate value after balancing the significance of sampled point and density.So far, to ultrasonic cardiography video shadow The characteristic point sampling work of picture terminates, next will be to sampling characteristic point out at whole ultrasonic cardiography video shadow Track in Xiang is tracked.

In the present embodiment, intensive track is by independent each characteristic point followed the tracks of in each metric space Obtain, as shown in Figure 3.Method according to optical flow field draws intensive optical flow field algorithm and carries out characteristic point Follow the trail of.Intensive optical flow field is obtained by motion model between characteristic point in adjacent two two field pictures before and after estimating, The pixel value using neck near field obtains the estimated value of motion model by extension.

But in ultrasonic cardiography video image, some independent multinomial can not represent whole picture frame, i.e. Make to be that adjacent frame can not be associated by the overall situation.If the spreading coefficient of two adjacent images is respectively A₁(x)、b₁(x)、c₁(x) and A₂(x)、b₂(x)、c₂(x), wherein A₁(x) and A₂X () is not equal, Do a lower aprons to obtain:

$A\left(x\right)=\frac{A_1\left(x\right)+A_2\left(x\right)}{2}$

And Δ b (x)=-0.5 (b₂(x)-b₁(x)) and then try to achieve constraints A (x) d (x)=Δ b (x)

Among this, d (x) is as the change of locus and the displacement vector that changes.

Owing to the interval time between frame each in ultrasonic cardiography video image is the shortest, it can be assumed that consecutive frame Between motion vector be slowly varying, d (x) can be solved according to the neighborhood pixels of characteristic point in two field picture, The least square formula in neighborhood is set up according to this:

$\underset{\mathrm{\Δ}x\∈I}{\Σ}w\left(\mathrm{\Δ}x\right)\left|\right|A(x+\mathrm{\Δ}x)d\left(x\right)-\mathrm{\Δ}b(x+\mathrm{\Δ}x)|{|}^2$

Wherein, (Δ x) is the weighting function of adjacent domains I of pixel x to w.Above formula is minimized:

D (x)=(Σ wA^TA)^-1ΣwA^TΔb

In order to improve the robustness of displacement vector, use eight movement parameter models that displacement vector carries out parametrization:

d_x(x, y)=a₁+a₂x+a₃y+a₇x²+a₈xy

d_y(x, y)=a₄+a₅x+a₆y+a₇xy+a₈y²

And then release: d=Sp

$S=\left[\begin{array}{cccccccc}1& x& y& 0& 0& 0& x^2& xy\\ 0& 0& 0& 1& x& y& xy& y^2\end{array}\right]$

P=(a₁a₂a₃a₄a₅a₆a₇a₈)^T

Finally give:

$\underset{i}{\Σ}{w}_i\left|\right|A_i{S}_{i}p-{\mathrm{\Δb}}_i|{|}^2$

Wherein, i is the coordinate of the pixel of two field picture in neighborhood, and the solution that can obtain above formula is:

$P={\left(\underset{i}{\Σ}{w}_i{S}_i^T{A}_i^T{A}_i{S}_i\right)}^{-1}\underset{i}{\Σ}{w}_i{S}_i^T{A}_i^T{\mathrm{\Δb}}_i$

It is a partial model variable between Polynomial Expansion, when displacement vector is excessive, can produce and cannot avoid Mistake.Here, need to introduce the concept of a priori displacement, the priori displacement is made to beThen obtain true Relative displacement between real-valued and priori estimates, this relative displacement is the least certainly.Therefore, obtain:

$A\left(x\right)=\frac{A_1\left(x\right)+A_2\left(\tilde{x}\right)}{2}$

$\mathrm{\Δ}b\left(x\right)=-0.5\left(b_2\right(\tilde{x})-b_1(x\left)\right)+A\left(x\right)\tilde{d}\left(x\right)$

$\tilde{x}=x+\tilde{d}\left(x\right)$

As the algorithm of an iterative computation, prior estimate the best more accurate, its relative displacement will be the least, The most final result of calculation will be the most accurate.Previous estimation displacement is entered as priori displacement next time Row calculates, and for the first time during iteration, priori displacement is set to 0, can restrain close to exact value after general iteration 4～5 times. So far, the calculating of intensive optical flow field completes.

Further, in the present embodiment, after the track obtaining characteristic point, need each characteristic point Track be described, obtain for classification characteristic vector.In the present embodiment, by combining DHOG, HOF Describe algorithm with MBH feature point trajectory is described to calculate the description vectors of feature point trajectory.

Further, HOF is that a kind of behavioral characteristics describes algorithm, for calculating the spot speed of Moving Objects Calculate and represent the kinestate of this moving object.It can provide the distribution of the spatial shape of object and be somebody's turn to do The important information of the rate of change etc. of object form, also can show amplitude and the direction of object of which movement simultaneously.Light The computational methods of stream rectangular histogram (HOF) are that Video segmentation becomes several the same local space time's SPACE V, time One space-time space of a few two field picture of continuous print block composition on axle.Then on image by any point (x, y) permissible Obtain its light stream vectorsV therein_x(x, y) and v_y(x y) represents respectively It is that light stream is at x-axis and the optical flow components of y-axis.Therefore, pixel (x, y) size of the light stream at place is:

$F(x,y)=\sqrt{v_x^2(x,y)+v_y^2(x,y)}$

The direction of its light stream is:

$\mathrm{\α}(x,y)=arctan\left(\frac{v_y(x,y)}{v_x(x,y)}\right)$

[0,360 °) light flow path direction decile be divided into b sector, the most each pixel is interval at component On range value be:

$G_k(x,y)=\left\{\begin{array}{cc}F(x,y)& \mathrm{\α}(x,y)\∈b_k\\ 0& \mathrm{\α}(x,y)\∉b_k\end{array}\right.$

In the present embodiment, it is contemplated that the feature of ultrasonic cardiography video image causes the light stream of respective pixel or voxel Range value is less than the threshold value of regulation, therefore to retain this trickle feature, increases by one in HOF vector Individual null value region, is promoted to 9 regions by the HOF vector in 8 traditional regions, for preserving pixel Number, i.e. 9 dimensions.

Further, moving boundaries rectangular histogram (MBH) is based on light stream rectangular histogram (HOF), by intensive light Stream orthogonal division is horizontal and vertical component, the most in space to its derivation and calculate its light stream rectangular histogram respectively, Thus obtain two orthogonal rectangular histograms of MBHx and MBHy.Its computational methods are basically identical with HOF algorithm, But moving boundaries rectangular histogram without increasing a null value region like that to HOF, therefore according to different directions It is projected in 8 sectors with weight, then uses L₂Its characteristic vector is normalized by norm, institute It is all 8 dimensions with MBHx and MBHy.

Further, differential direction histogram of gradients (DHOG) is a kind of description for moving object detection Operator, its basic thought is to use direction based on gradient amplitude weight to carry out projecting and the rectangular histogram that generates is retouched Stating moving object, and then the appearance profile information of extraction moving object, generate complete characteristic set, it generates Process is as shown in Figure 4.

The first step, is normalized two field picture, primarily to improve the image light to different bright-dark degrees Robustness.For ultrasonic cardiography video image, the characteristic of its strong noise and low resolution makes video shadow Light and shade in Xiang can change because of the change of cardiac motion characteristics, therefore should try one's best detecting when Get rid of the image that bright and dark light is brought, thus promote Detection results.

Second step, before calculating the gradient of image, first has to image is carried out gamma correction, corrects it After the calculating effect of two field picture be better than the effect after Gaussian smoothing, because smoothing processing can make image The blur margin of middle motion model is clear, image identification degree, and particularly ultrasonic cardiography video image is this itself differentiates Rate is with regard to low image, and its marginal information is the most clear.Therefore, gamma correction is used can to improve meter Calculate degree of accuracy.The formula of gamma correction is f (I)=I^γ。

3rd step, is divided into S interval by 0～360 °, and wherein S is even number, in each cell Pixel is voted, thus obtains its gradient orientation histogram H_g:

$H_{og}\left(i\right)=H_g\left(i\right)+H_g(i+\frac{S}{2}),1\≤i\≤\frac{S}{2}$

Wherein H_og(i) and H_gI () represents HOG and H respectively_gIn i-th element value.In order to improve tradition The ability to express of HOG feature description, after HOG feature descriptor again series connection on one group of brand-new gradient straight Side figure H_ng, the length of this histogram vectors is identical with HOG, and each of which element value obtains expression formula and is:

$H_{ng}\left(i\right)=|H_g\left(i\right)-H_g(i+\frac{S}{2})|,1\≤i\≤\frac{S}{2}$

Finally give DHOG gradient vector.

4th step, carries out contrast normalization to the DHOG characteristic vector in region, and main purpose is to make DHOG characteristic vector has robustness for the illumination of shade and different brightness.Contrast normalization is often One intra-zone is carried out, and its function expression is:

$L_2-norm,v\←v/\sqrt{\left|\right|v|{|}_2^2+{\mathrm{\ϵ}}^2}$

In formula, ε be one for the least constant amount preventing denominator from being 0.

5th step, in the present embodiment, DHOG characteristic vector only calculates the intensive track that obtains above HOG descriptor in the pixel region of 32 × 32, so the size of current block is the pixel region of 32 × 32. Each piece has 2 × 2 cells, then intensive for several frames track frame is combined, single to each in these several frames The DHOG feature descriptor of unit's lattice is sued for peace according to corresponding rectangular histogram, and uses L₂Norm is normalized. Therefore, the dimension size of the DHOG feature descriptor of one group of intensive track frame is 16 dimensions.So far DHOG Feature description vector generates complete.

Describe the algorithm common description to feature point trajectory by three, finally give ultrasonic cardiography video image Characteristic vector, by encoded for characteristic vector be input to suitable grader after, it is possible to complete the ultrasonic heart The classification work of dynamic video image.

It is above presently preferred embodiments of the present invention, all changes made according to technical solution of the present invention, produced merit When can act on the scope without departing from technical solution of the present invention, belong to protection scope of the present invention.

Claims (9)

1. a ultrasonic cardiography video image sorting technique based on intensive track and DHOG, it is characterised in that

Step S1: the two field picture of ultrasonic cardiography video image is carried out intensive sampling；

Step S2: remove the homogenizing part in the characteristic point region that intensive sampling obtains；

Step S3: in subsequent ultrasonic video image aroused in interest frame, the track of characteristic point is tracked；

Step S4: use descriptor that the track of characteristic point is described；

Step S5: will obtain through described step S4 that ultrasonic cardiography video image characteristic vector is encoded to be input to Corresponding grader, completes the classification to ultrasonic cardiography video image.

A kind of ultrasonic cardiography video image based on intensive track and DHOG the most according to claim 1 divides Class method, it is characterised in that in described step S1, enters in a mesh space containing W1 pixel The characteristic point sampling that row is intensive；Sampling scale according to each metric space according to the factorConstantly increase, Eight metric spaces of sampling altogether.

A kind of ultrasonic cardiography video image based on intensive track and DHOG the most according to claim 1 divides Class method, it is characterised in that in described step S2, when the local of the characteristic point in described characteristic point region When the eigenvalue of autocorrelation matrix is less than threshold value T, the point just this feature point being judged as in average region, and will It is removed, and threshold value T is:

$T=0.001\×\underset{i\∈I_t}{max}\min ({\mathrm{\λ}}_i^1,{\mathrm{\λ}}_i^2)$

Wherein,It it is frame of video I_tTwo eigenvalues of the correlation matrix M of middle characteristic point i, 0.001 is Intermediate value after the significance and density of balanced sample point.

A kind of ultrasonic cardiography video image based on intensive track and DHOG the most according to claim 1 divides Class method, it is characterised in that in described step S3, by feature in two two field pictures adjacent before and after estimating Motion model between point obtains intensive optical flow field；Use the pixel value of neck near field, by extension campaign The estimated value of model；Draw intensive optical flow field algorithm according to optical flow field method, and characteristic point is tracked.

A kind of ultrasonic cardiography video image based on intensive track and DHOG the most according to claim 4 divides Class method, it is characterised in that

Step S31: the spreading coefficient of two adjacent images is expressed as A₁(x)、b₁(x)、c₁(x) and A₂(x)、b₂(x)、c₂(x), wherein, A₁(x) and A₂X () is the most unequal, and approximate:

$A\left(x\right)=\frac{A_1\left(x\right)+A_2\left(x\right)}{2}$

And Δ b (x)=-0.5 (b₂(x)-b₁(x)), and then try to achieve constraints A (x) d (x)=Δ b (x), wherein, d (x) The displacement vector being as the change of locus and change；

Step S32: solve d (x) according to the neighborhood pixels of characteristic point in two field picture, sets up the minimum in neighborhood and puts down Mode:

$\underset{\mathrm{\Δ}x\∈I}{\Σ}w\left(\mathrm{\Δ}x\right)\left|\right|A(x+\mathrm{\Δ}x)d\left(x\right)-\mathrm{\Δ}b(x+\mathrm{\Δ}x)|{|}^2$

In order to improve the robustness of displacement vector, use eight movement parameter models that displacement vector carries out parametrization:

d_x(x, y)=a₁+a₂x+a₃y+a₇x²+a₈xy

d_y(x, y)=a₄+a₅x+a₆y+a₇xy+a₈y²

Finally give:

$P={\left(\underset{i}{\Σ}{w}_i{S}_i^T{A}_i^T{A}_i{S}_i\right)}^{-1}\underset{i}{\Σ}{w}_i{S}_i^T{A}_i^T{\mathrm{\Δb}}_i$

Wherein, the coordinate of the pixel of two field picture in i is neighborhood；

Step S33: introduce a priori displacement, makes the priori displacement beObtain actual value and prior estimate Relative displacement between value, obtains:

$A\left(x\right)=\frac{A_1\left(x\right)+A_2\left(\tilde{x}\right)}{2};$

$\mathrm{\Δ}b\left(x\right)=-0.5\left(b_2\right(\tilde{x})-b_1(x\left)\right)+A\left(x\right)\tilde{d}\left(x\right);$

$\tilde{x}=x+\tilde{d}\left(x\right);$

Step S34: be iterated calculating as priori displacement next time using previous estimation displacement, and For the first time during iteration, priori displacement is set to 0.

A kind of ultrasonic cardiography video image based on intensive track and DHOG the most according to claim 1 divides Class method, it is characterised in that in described step S4, describes algorithm, MBH description by combining HOF Algorithm and DHOG describe algorithm and jointly describe feature point trajectory, to calculate retouching of feature point trajectory State vector.

A kind of ultrasonic cardiography video image based on intensive track and DHOG the most according to claim 6 divides Class method, it is characterised in that described HOF describes the computational methods of algorithm and realizes in accordance with the following steps:

Step S411: Video segmentation is become local space time's SPACE V that several are the same, if continuous print on time shaft One space-time space of dry two field picture block composition；Then on image, by any point, (x y) can obtain its light stream VectorWherein, v_x(x, y) and v_y(x y) represents that light stream is in x-axis respectively Optical flow components with y-axis；Therefore pixel (x, y) size of the light stream at place is:

$F(x,y)=\sqrt{v_x^2(x,y)+v_y^2(x,y)}$

The direction of its light stream is:

$\mathrm{\α}(x,y)=\arctan \left(\frac{v_y(x,y)}{v_x(x,y)}\right);$

Step S412: [0,360 °) light flow path direction decile be divided into b sector, the most each pixel Range value on component interval is:

$G_k(x,y)=\left\{\begin{array}{cc}F(x,y)& \mathrm{\α}(x,y)\∈b_k\\ 0& \mathrm{\α}(x,y)\∉b_k\end{array}\right.;$

Step S413: increase a null value region in HOF vector, by the HOF in 8 traditional regions Vector is promoted to 9 regions, for preserving the number of pixel, i.e. 9 dimensions.

A kind of ultrasonic cardiography video image based on intensive track and DHOG the most according to claim 6 divides Class method, it is characterised in that described MBH describes the computational methods of algorithm and describes algorithm based on HOF, will Intensive light stream orthogonal division is horizontal and vertical component, in space to its derivation and calculate its light stream Nogata respectively Figure, thus obtain two orthogonal rectangular histograms of MBHx and MBHy.

A kind of ultrasonic cardiography video image based on intensive track and DHOG the most according to claim 6 divides Class method, it is characterised in that described DHOG describes the computational methods of algorithm and comprises the steps:

Step S421: two field picture is normalized；

Step S422: image is carried out gamma correction；

Step S423: be divided into S interval by 0～360 °, wherein S is even number, in each cell Pixel vote, thus obtain its gradient orientation histogram H_g:

$H_{og}\left(i\right)=H_g\left(i\right)+H_g(i+\frac{S}{2}),1\≤i\≤\frac{S}{2}$

Wherein, H_og(i) and H_gI () represents HOG and H respectively_gIn i-th element value；In order to improve biography System HOG feature description ability to express, after HOG feature descriptor again series connection on one group of brand-new ladder Degree rectangular histogram H_ng, and then obtain DHOG gradient vector, the length of this histogram vectors is identical with HOG, Each of which element value obtains expression formula:

$H_{ng}\left(i\right)=|H_g\left(i\right)-H_g(i+\frac{S}{2})|,1\≤i\≤\frac{S}{2}$

Step S424: the DHOG characteristic vector in region is carried out contrast normalization；

In the pixel region of the 32 × 32 of step S425:DHOG characteristic vector only computation-intensive track HOG descriptor, the size of current block is the pixel region of 32 × 32, wherein, has 2 × 2 lists in each piece Unit's lattice, are combined intensive track frame, retouch the DHOG feature of each cell in this combined frames State symbol to sue for peace according to corresponding rectangular histogram, and use L₂Norm is normalized, and in this combined frames The dimension of DHOG feature descriptor is 16 dimensions.