CN101354359A - Method for detecting, tracking and recognizing movement visible exogenous impurity in medicine liquid - Google Patents

Abstract

The invention discloses a detecting and recognition method of moveable and visible foreign matters and bubbles in liquid medicines, which is characterized in that the method comprises the following steps: (1) a continuous multi-frame image of a liquid medicine to be detected is obtained; (2) the image is preprocessed; (3) a moveable object is extracted, and a reinforced image is obtained; (4) the moveable object is partitioned as follows: an improved two-dimensional maximum entropy threshold partitioning algorithm is applied into the processing of the reinforced image and a two-dimensional histogram that is formed by grayscale and average neighborhood grayscale is adopted for choosing threshold, thus obtaining a partitioned image; (5) the movable object is traced; and (6) the image is identified and determined as follows: a movement trace of the object is obtained according to the tracing result, objects inducing foreign matters and bubbles are identified and distinguished by utilizing the continuity and the directivity of the object movement trace, and consequently whether the liquid medicine is qualified is finally determined. Under the condition of ensured algorithm precision, the detecting and recognition method of moveable and visible foreign matters and bubbles in liquid medicines adequately simplifies computation, improves the real-time property of the algorithm, greatly raises the accuracy of foreign matter and bubble identification in the liquid medicine, and lowers the false detection rate of the medicine.

Description

The detection tracking and the recognition methods of movement visible exogenous impurity in medicine liquid

Technical field

The invention belongs to Flame Image Process and technical field of automation, relate to the detection recognition methods of motion visible foreign matters and bubble in a kind of liquid drug.

Background technology

At present, widely used liquid drug comprises injection, infusion solutions, oral liquid, eye drops etc. in China, usually adopt the form can of vial or plastic bottle, these medicines are not good owing to filtering in process of production, reason such as collision when container cleans not good, encapsulation, can make visible foreign matters such as having chips of glass, suspension, hair in the soup, in order to detect in the liquid preparation after the can whether sneak into visible foreign matters, need carry out visible foreign matters to each bottle medicine and detect.China's pharmacopeia was promptly stipulated to carry out clarity from 1985 to the detection of this similar drug and is detected and the particulate matter inspection, and foreign matter in the medicine and particulate are carried out strictness control.On August 24th, 1999, national Bureau of Drugs Supervision formally prints and distributes the notice of relevant regulations " about the enforcement＜GMP〉", explicitly calls for pharmaceutical producing enterprise must pass through GMP (good) authentication before the end of the year 2000." Chinese pharmacopoeia version in 2005 is in execution on July 1 in 2005, former " clarity test detailed rules and regulations and criterion " revision is " visible foreign matters inspection technique ", further clear and definite medicines such as injection, eye drops must be produced under the condition that meets GMP (GMP), should adopt suitable method to check and reject simultaneously substandard product one by one before the product export, visible foreign matters is meant and is present in injection, the eye drops, the visual insoluble substance that can observe under rated condition, its particle diameter or length are usually greater than 50um.The visible foreign matters inspection technique has lamp test and light scattering method, general normal employing lamp test.Present domestic medicine visible foreign matters detects the method that mostly adopts artificial lamp inspection, rotate gently manually or the upset bottle container, make the visible foreign matters that may the exist motion (noting not making soup to produce bubble) in the soup, judge by the human eye range estimation, if exist visible foreign matters its manual rejecting.That can not avoid during manual detection exists efficient low, the loss height, and problem such as precision is low, the repeatability and the consistance of testing result are poor, and are engaged in testing for a long time and also bring harm to lamp inspection personnel's health.Popularization along with the GMP authentication, China's medicine is produced equipment and checkout equipment robotization and intelligent level and is greatly improved, semi-automatic lamp inspection machine and the full-automatic lamp inspection machine that the medicine visible foreign matters detects appearred being used for, adopt wherein semi-automatic lamp inspection machine simple mechanical drive and optical system realize the semi-automation of artificial lamp inspection, this pick-up unit can alleviate hand labor to a certain extent, but still examine work point by lamp and analyse judgement, accuracy of detection and accuracy rate are not improved.Full-automatic lamp inspection machine has only several companies of a few countries such as Germany, Japan, Italy to have the ability to produce, but external so far full-automatic lamp inspection machine is used still seldom in China, trace it to its cause and have following problem: the filtration in the domestic pharmacy link, wrappage etc. and have very big-difference abroad cause that to detect quality low.But along with the continuous improvement of domestic pharmaceutical environment, the advanced medical checkout equipment of foreign country certainly will change the detection present situation in domestic pharmacy market.In the face of the impact of external advanced complete medical equipment, medical checkout equipment and core detection recognition methods that research and development have independent intellectual property right have crucial value.

In medical fluid foreign body in vivo vision-based detection process, need to make the foreign matter that is deposited in the soup bottom drive liquid upper with special mechanical hook-up, so that video camera imaging, but because the motion of mechanical hook-up often makes in the image after the imaging how much have bubble and noise, and the technical barrier that is faced in foreign matter target and the bubble identification in the image mainly contains:

1, foreign matter is not of uniform size, and detection resolution will reach 50um, and the foreign matter target does not have information such as size, shape and texture, and traditional image processing method can't be used;

2, because bottle surface exists scale embossment, spot, the image background complexity after the imaging, noise is various;

3, signal to noise ratio (S/N ratio) is low in the image, and the foreign matter target often is submerged in the ground unrest;

4, the production line detection speed requires highly, and the real-time of recognizer has been proposed very high requirement.

Summary of the invention

Technical matters to be solved by this invention is, the detection recognition methods of motion visible foreign matters and bubble in a kind of liquid drug is provided, be used for detecting substandard product rapidly with foreign matter at automatic fluid medicine detection line, carry out Digital Signal Analysis and Processing with imaging hardware system is supporting, for medical checkout equipment provides a kind of reliable detection method, thereby improve medicine visible foreign matters accuracy of detection and repeatability, thoroughly solve detect in the high problem of false drop rate, satisfy the performance requirement of domestic existing lamp check system.

The present invention solves the problems of the technologies described above the technical scheme that is adopted to be:

The detection recognition methods of motion visible foreign matters and bubble is characterized in that in a kind of liquid drug, may further comprise the steps:

1) current liquid drug to be detected is obtained the multiframe consecutive image;

2) image pre-service: the circular shuttering of employing 7 * 7 carries out high cap (top-hat) morphologic filtering as structural element to the image that obtains, and to remove and less all kinds of bright noise and the dark noise of structural element, obtains the denoising image;

3) moving target extracts: the kinetic characteristic that shows as interframe according to the foreign matter target on time domain, utilize the moving target in the novel three two field picture difference algorithms extraction image, its operating process is: first frame and second frame are done difference processing, second frame and the 3rd frame are done difference processing, strengthen target image by the image accumulation then, obtain to strengthen the back image;

4) moving Object Segmentation: to described enhancing back image utilization modified Two-dimensional maximum-entropy Threshold Segmentation Algorithm, the two-dimensional histogram that adopts gray scale and neighborhood averaging gray scale to constitute is selected threshold value, obtains cutting apart the back image;

5) motion target tracking: when detect cut apart after continuous 3 frames of possible target in the image appear in certain neighborhood, assert that then movement locus is initial; According to the regularity of foreign matter target in the image and bubble target travel, adopt based on the Kalman filtering of uniform rectilinear motion model and upgrade dbjective state, and generate oval tracking gate thus; When the measuring value that falls into oval tracking gate has only one, directly carry out target trajectory and upgrade; When the measuring value that falls into oval tracking gate more than one, target trajectory by distance further the nearest measured value of predicted value upgrade; When track rejection, the possible position of a few frame targets behind the filter value outside forecast in the moment keeps stable memory tracking power before adopting;

6) image recognition and judgement: obtain target trajectory by tracking results, utilize the continuity of target trajectory and directivity cog region to tell foreign matter and bubble target, thereby judge finally whether current liquid drug is qualified.

Described step 2) be:

If the gray scale function of image be f (x, y), structural element be B (x, y), (x y) is pixel coordinate, and then 4 kinds of basic operations are defined as follows:

Corrosion: (f Θ B) (x, y)=min{f (x+i, y+j)-B (i, j) | (x+i, y+j) ∈ D _f(i, j) ∈ D _B;

Expand: $(f\⊕B)(x,y)=\max \{f(x-i,y-i)+B(i,j)|(x-i,y-i)\∈D_f;(i,j)\∈D_B\};$

Opening operation:

Closed operation: $f\·B=(f\⊕B)\mathrm{\ΘB};$

In the formula, D _fAnd D _BBe respectively the field of definition of function f and B,

Introduce cap transformation filter operator: Hat (f)=f-(f ο B);

(f ο B) carries out the gray scale opening operation for structural element B to image f in the formula, selects for use 7 * 7 circular shuttering as structural element original image to be carried out high cap morphologic filtering, obtains containing the image of foreign matter target, bubble, i.e. the denoising image.

Described step 3) is:

At the denoising image, according to image sequence first frame and second frame in every continuous three two field pictures are done difference processing, second frame and the 3rd frame are done difference processing, press following formula then and strengthen target image, obtain to strengthen the back image:

d(x，y)＝|f ₁(x，y)-f ₂(x，y)|×|f ₂(x，y)-f ₃(x，y)|；

F wherein ₁(x, y), f ₂(x, y), f ₃(x y) is respectively through filtered first frame of high cap, second frame, the 3rd two field picture, | f ₁(x, y)-f ₂(x, y) | be first frame and the second frame difference image, | f ₂(x, y)-f ₃(x, y) | be second frame and three-frame difference image.

Step 4) may further comprise the steps:

(a) image segmentation is become a series of subimages;

(b) the Two-dimensional maximum-entropy algorithm computation goes out the threshold value of each subimage;

(c) according to the threshold value of each subgraph, image is cut apart;

Wherein the acquiring method of threshold value is as follows: establishing size of images is M * N, and image has L gray level, and the field average gray of each pixel also is a L gray level, and total gray level is L * L; On two dimensional surface, make the two dimensional gray histogram, gray-scale value and neighborhood average gray form the gray scale bi-values, the coordinate figure that is designated as each pixel for (i, j), the gray-scale value of i and j difference represent pixel point and the average gray in its field, the frequency that this gray scale combination occurs is f _Ij, this pixel functional value is joint probability density function p _Ij:

$p_{\mathrm{ij}}=\frac{f_{\mathrm{ij}}}{M\×N}$ i，j＝0，1，...L-1；

If so that (s t) is segmentation threshold, to the probability distribution p of object O _Ij(i=1,2 ... s; J=1,2 ... t) carry out the normalization operation; The definition normalization operation back entropy relevant with the probability distribution of object O is:

$H\left(O\right)=-\underset{i-0}{\overset{s}{\mathrm{\Σ}}}\underset{j=0}{\overset{t}{\mathrm{\Σ}}}\frac{p_{\mathrm{ij}}}{P_{\mathrm{st}}}\ln \frac{p_{\mathrm{ij}}}{P_{\mathrm{st}}}=-\frac{1}{P_{\mathrm{st}}}\underset{i=0}{\overset{s}{\mathrm{\Σ}}}\underset{j=0}{\overset{t}{\mathrm{\Σ}}}(p_{\mathrm{ij}}\ln p_{\mathrm{ij}}-p_{\mathrm{ij}}\ln P_{\mathrm{st}})=\ln P_{\mathrm{st}}+\frac{H_{\mathrm{st}}}{P_{\mathrm{st}}};$

Wherein, $P_{\mathrm{st}}=\underset{i=0}{\overset{s}{\mathrm{\Σ}}}\underset{j=0}{\overset{t}{\mathrm{\Σ}}}{p}_{\mathrm{ij}},$ $H_{\mathrm{st}}=-\underset{i=0}{\overset{s}{\mathrm{\Σ}}}\underset{j=0}{\overset{t}{\mathrm{\Σ}}}{p}_{\mathrm{ij}}\ln p_{\mathrm{ij}};$

The entropy relevant with the probability distribution of background B is

$H\left(B\right)=-\underset{i=s+1}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=t+1}{\overset{L-1}{\mathrm{\Σ}}}\frac{p_{\mathrm{ij}}}{1-P_{\mathrm{st}}}\ln \frac{p_{\mathrm{ij}}}{1-P_{\mathrm{st}}}=-\frac{1}{1-P_{\mathrm{st}}}\underset{i=s+1}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=t+1}{\overset{L-1}{\mathrm{\Σ}}}[p_{\mathrm{ij}}\ln p_{\mathrm{ij}}-p_{\mathrm{ij}}\ln (1-P_{\mathrm{st}})];$

$=\ln (1-P_{\mathrm{st}})+\frac{H_{L-1L-1}-H_{\mathrm{st}}}{1-P_{\mathrm{st}}}$

Wherein, $H_{L-1L-1}=-\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{L-1}{\mathrm{\Σ}}}{p}_{\mathrm{ij}}\ln p_{\mathrm{ij}};$

Criterion function Ψ (s t) is taken as H (O), H (B) sum, promptly

$\mathrm{\Ψ}(s,t)=H\left(O\right)+H\left(B\right)=\ln P_{\mathrm{st}}+\frac{H_{\mathrm{st}}}{P_{\mathrm{st}}}+\ln (1-P_{\mathrm{st}}+\frac{H_{L-1L-1}-H_{\mathrm{st}}}{1-P_{\mathrm{st}}});$

$=\ln P_{\mathrm{st}}(1-P_{\mathrm{st}})+\frac{H_{\mathrm{st}}}{P_{\mathrm{st}}}+\frac{H_{L-1L-1}-H_{\mathrm{st}}}{1-P_{\mathrm{st}}}$

(s, t) (s t) promptly is the optimal threshold t that is asked to Zui Da gray-scale value to make Ψ ^*, promptly

$t^{*}=\mathrm{Arg}\left\{\underset{s,t\∈G}{\mathrm{Max}}\mathrm{\Ψ}(s,t)\right\};$ Wherein Arg{} represents the function of negating.

As improvement, described step 4) also comprises: the introducing genetic algorithm is optimized the partition threshold of Two-dimensional maximum-entropy method, to shorten the Flame Image Process time; Concrete steps are as follows:

Initialization population: produce at random, select evenly to distribute;

Population scale: by analysis, population scale is popsize=20, and maximum evolutionary generation is g _Max=100;

Fitness function: (s is t) as fitness function directly to use target function type Ψ;

Coding: because the gradation of image value between 0～255, adopts 8 gray level images, then threshold parameter satisfies 0≤s, t≤255, so individuality is encoded to 16 binary codes, most-significant byte is represented s, and least-significant byte is represented t;

Decoding: the most-significant byte of 16 binary codes is decoded as the s value, and least-significant byte is decoded as the t value;

Select: the employing ratio is selected operator;

Crossover operator: adopt two point to intersect, at random two point of crossing of Chan Shenging lay respectively at coded strings preceding 8 with back 8; Carry out interlace operation with crossover probability Pc; Early stage is chosen Pc=0.8 in search, and the search later stage is chosen Pc=0.6;

Mutation operator: mutation operation is the step-by-step negate; The probability P that will make a variation m is defined as class parabolic type mutation operator:

$P_m=P_m\left(g\right)=P_{m\_\max }-\frac{4\×(P_{m\_\max }-P_{m\_\min })}{g_{\max }^2}{(g-\frac{g_{\max }}{2})}^2$

Wherein g is an evolutionary generation, and g ∈ [1, g _Max], P _{M_min}=P _b, P _{M_max}=10P _{M_min}

P _bBe basic variation probability, estimate by following formula:

$P_b\≈\frac{175}{\mathrm{popsize}\×\sqrt{\mathrm{bits}}};$

Wherein bits represents individual figure place; Work as popsize=20, during bits=16, P _b≈ 0.02;

Work as P _b≈ 0.02, g _Max=100 o'clock, can get:

P _m＝P _m(g)＝0.2-0.00008×(g-50) ²；

Stop criterion: when algorithm was carried out the highest fitness value in maximum algebraically or the colony and stablized, algorithm was out of service, and having, the individuality of high fitness value is separating of being asked; When algorithm stopped, the individuality with the highest fitness value was as optimal threshold.

Described step 5) may further comprise the steps:

By determining to obtain the single frames testing result, that target is approximate with uniform rectilinear motion model based on modified Two-dimensional maximum-entropy threshold segmentation algorithm; Its tracker model is: writ attitude vector is $X\left(k\right)={[x\left(k\right),\stackrel{\·}{x}\left(k\right),y\left(k\right),\stackrel{\·}{y}\left(k\right)]}^T,$ Measuring vector is Z (k)=[x (k), y (k)] ^T, system equation can be write to be become

X(k+1)＝ΦX(k)+Gw(k)；

Z(k+1)＝HX(k)+v(k)；

Wherein $\mathrm{\&Phi;}=\left[\begin{array}{cccc}1& T& 0& 0\\ 0& 1& 0& 0\\ 0& 0& 1& T\\ 0& 0& 0& 1\end{array}\right],$ $G=\left[\begin{array}{cc}\frac{{\mathrm{<figure-callout\; id="2"\; label="T"\; filenames="A20081014313800231.png"\; state="\{\{state\}\}">T</figure-callout>}}^2}{2}& 0\\ T& 0\\ 0& \frac{{\mathrm{<figure-callout\; id="2"\; label="T"\; filenames="A20081014313800231.png"\; state="\{\{state\}\}">T</figure-callout>}}^2}{2}\\ 0& T\end{array}\right],$ $H=\left[\begin{array}{cccc}1& 0& 0& 0\\ 0& 0& 1& 0\end{array}\right],$

W (k), v (k) are respectively that variance is the zero-mean white Gaussian noise of Q (k) and R (k); T is the sampling period in the formula;

Track initial: if continuous 3 frames of possible target that detect appear in 3 * 3 neighborhoods, track initial then;

Tracking gate and data association: suppose being estimated as of k-1 state vector constantly

The one-step prediction of state vector is:

$\hat{X}\left(k\right|k-1)=\mathrm{\&Phi;}\hat{X}(k-1|k-1);$

Then the new breath of system is:

$d\left(k\right)=Z\left(k\right)-H\hat{X}\left(k\right|k-1);$

New breath variance battle array is:

S(k)＝HP(k|k-1)H ^T+R(k)；

Wherein, P (k|k-1) is the one-step prediction variance;

The norm of the new vectorial d of breath of order (k) is:

g(k)＝d ^T(k)S ^-1(k)d(k)；

Wherein g (k) obeys χ _M1 ²Distribute, M1 is for measuring dimension;

Tracking gate of definition in measurement space makes and measures with certain probability distribution in tracking gate;

$\tilde{V}\left(\mathrm{\&gamma;}\right)=[Z:g\left(k\right)\&le;\mathrm{\&gamma;}],$ Wherein γ can pass through χ ²The distribution difference checks in;

If after certain frame is handled, in tracking gate, have only an echo, then target trajectory directly upgrades; If have in the tracking gate more than an echo, then target trajectory upgrades by the nearest measured value of distance one-step prediction value;

The track of target upgrades by the standard Kalman filtering algorithm

X(k|k-1)＝ΦX(k-1|k-1)

P(k|k-1)＝ΦP(k-1|k-1)Φ ^T+GQ(k-1)G ^T

K(k)＝P(k|k-1)H ^T[HP(k|k-1)H ^T+R] ^-1；

X(k|k)＝X(k|k-1)+K(k)[Z(k)-H(k)X(k|k-1)]

P(k|k)＝[I-K(k)H]P(k|k-1)

In target detected and follow the tracks of after, if the of short duration disappearance of target, can be according to target positional information and motion state before this, dope next step possible position of target, when target occurs once more, but tenacious tracking and be unlikely to lose objects still, and detailed process is as follows:

Suppose being estimated as of k state vector constantly Then the predicted value in the target of k+n frame is:

$\hat{X}(k+n|k)={\mathrm{\&Phi;}}^n\hat{X}\left(k\right|k);$ Select n＜6.

Described step 6) is: suppose that for each bar that target following obtains it is that foreign matter forms or noise or bubble formation that track is distinguished track by following 3 principles, judgment criterion is as follows:

(a) because image is taken when static after medicine bottle turns over turnback, visible foreign matters moves downward, so its centre of form ordinate should be to become big successively, is true origin with the upper left corner;

(b) bubble moves upward, and direction of motion is opposite with foreign matter;

(c) geometric locus of foreign matter formation is smooth, and the track that noise forms is unordered;

Therefore, diminish successively, illustrate that this track is produced by bubble if detect the centre of form ordinate of movement locus; Become big track successively if detect the level and smooth and centre of form ordinate of movement locus, the foreign matter existence be described, judge that current liquid drug is defective.

Beneficial effect of the present invention has:

The present invention adopts the mathematics shape filtering to obtain pretreatment image, and its algorithm can pass through the hardware Parallel Implementation, has improved processing speed greatly.In moving target extracts, utilize novel three two field picture difference algorithms to overcome well that a bottle sidewall may there are differences, the simple sequence image difference is to small foreign matter target detection effect defect of bad, and significantly improved the signal to noise ratio (S/N ratio) of output image, simplified the follow-up difficulty of cutting apart detection algorithm greatly.Adopt modified Two-dimensional maximum-entropy threshold values moving Object Segmentation algorithm to make full use of the intensity profile information of object pixel and the relevant information between pixel in the image, improved the anti-noise ability of Threshold Segmentation, and be optimized by the parameter of genetic algorithm to the Two-dimensional maximum-entropy method, reduce the Flame Image Process time, improved the real-time and the robustness of algorithm greatly.According to the kinetic characteristic of moving target in the image, adopt based on the Kalman filtering of uniform rectilinear motion model and upgrade dbjective state, under the situation that guarantees tracking accuracy, simplified the calculated amount of filtering.The present invention guarantees fully to have simplified calculating under the condition of arithmetic accuracy, improved the real-time of algorithm, improved the accuracy rate that foreign matter and bubble in the soup detect identification greatly, medicine loss and false drop rate have been reduced, can be applicable to have vast market prospect and actual application value in the various detection systems such as medicine, beverage, drinks.

Description of drawings

Fig. 1 is the main-process stream block diagram of method involved in the present invention;

Fig. 2 is the original image of liquid drug in bottle that obtains among the present invention;

Fig. 3 is the enhancing image that target obtains after extracting among the present invention;

Fig. 4 is the image that obtains after the moving Object Segmentation among the present invention;

Fig. 5 is the movement locus figure that motion target tracking obtains among the present invention.

Label declaration: 1-bubble track, 2-foreign matter track.

Embodiment

Be described in further detail below with reference to Fig. 1～5 and specific embodiment.

Embodiment 1:

As shown in Figure 1, the present invention proposes a kind of motion visible foreign matters and bubble recognition methods that liquid drug detects that be used for.The concrete implementation detail of each several part is as follows:

1, Top-hat (high cap) morphologic filtering is handled

The image that obtains is adopted gray scale morphologic filtering filtering image noise.The most basic computing of gray scale morphology is exactly burn into expansion, opening operation and closed operation.If the gray scale function of image be f (x, y), structural element be B (x, y), then 4 kinds of basic operations are defined as follows:; (x y) is pixel coordinate in the formula;

Corrosion: (f Θ B) (x, y)=min{f (x+i, y+j)-B (i, j) | (x+i, y+j) ∈ D _f(i, j) ∈ D _B; Expand: $(f\&CirclePlus;B)(x,y)=\max \{f(x-i,y-i)+B(i,j)|(x-i,y-i)\&Element;D_f;(i,j)\&Element;D_B\};$

Opening operation:

Closed operation: $f\&CenterDot;B=(f\&CirclePlus;B)\mathrm{\&Theta;B};$

In the formula, D _fAnd D _BIt is respectively the field of definition of function f and B.The result of corrosion gray level image is, the part darker than background obtains expansion, and the part brighter than background shunk; The result of expansion gray level image is, the part brighter than background obtains expansion, and the part darker than background shunk; Opening operation makes the profile of image become smooth, disconnects narrow interruption and eliminates thin protrusion; Closed operation makes the profile of image become smooth equally, but opposite with opening operation, and it can eliminate narrow interruption and long thin wide gap, eliminates little hole, and fills up the slight crack in the outline line.

Be defined as based on the morphologic Top-hat conversion of gray scale filter operator:

Hat(f)＝f-(fοB)；(1)

B is a structural element in the formula, (f ο B) carries out the gray scale opening operation for structural element B to image f, its structure promptly is the background that estimates, obtain Hat (f) after subtracting each other, the target image of noise that has been filtering, certainly, also include some single-point impulsive noises through filtered image, they can filtering in the processing of back.Because the noise in the image that obtains is little speck, therefore, select for use 7 * 7 circular shuttering original image to be carried out the Top-hat morphologic filtering as structural element, obtain containing the denoising image of foreign matter target, bubble, as shown in Figure 3.

2, the moving target based on three two field picture difference algorithms extracts

At the image after the processing of Top-hat filtering and noise reduction, according to image sequence first frame and second frame are done difference processing, second frame and the 3rd frame are done difference processing, press following formula then and strengthen target image:

d(x，y)＝|f ₁(x，y)-f ₂(x，y)|×|f ₂(x，y)-f ₃(x，y)|(2)

F wherein ₁(x, y), f ₂(x, y), f ₃(x y) is respectively through filtered first frame of Top-hat, second frame, the 3rd two field picture, | f ₁(x, y)-f ₂(x, y) | be first frame and the second frame difference image, | f ₂(x, y)-f ₃(x, y) | be second frame and three-frame difference image;

After handling by above-mentioned three two field picture difference algorithms, (x has only the pixel position of moving target correspondence non-vanishing in y) to d.But in fact, because the existence of various noises, make in the difference image also non-vanishingly on a lot of pixels position outside moving target, this just requires follow-up processing can effectively identify foreign matter information.

3, moving Object Segmentation

If size of images is M * N, image has L gray level, and the field average gray of each pixel also is this L gray level, and total gray level is L * L.Make the two dimensional gray histogram in the plane, gray-scale value and neighborhood average gray form the gray scale bi-values, be designated as (i, j), i and the gray-scale value of j difference represent pixel point and the average gray in its field, the frequency that this gray scale combination occurs is f _Ij, this pixel functional value is joint probability density function p _Ij

$p_{\mathrm{ij}}=\frac{f_{\mathrm{ij}}}{M\&times;N}$ i，j＝0，1，...L-1(3)

If so that (s t) is segmentation threshold, corresponding to the probability distribution p of object O _Ij(i=1,2 ... s; J=1,2 ... t) and should be 1, so need carry out normalization operation.

The definition normalization operation back entropy relevant with the probability distribution of object O is:

$H\left(O\right)=-\underset{i-0}{\overset{s}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{t}{\mathrm{\&Sigma;}}}\frac{p_{\mathrm{ij}}}{P_{\mathrm{st}}}\ln \frac{p_{\mathrm{ij}}}{P_{\mathrm{st}}}=-\frac{1}{P_{\mathrm{st}}}\underset{i=0}{\overset{s}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{t}{\mathrm{\&Sigma;}}}(p_{\mathrm{ij}}\ln p_{\mathrm{ij}}-p_{\mathrm{ij}}\ln P_{\mathrm{st}})=\ln P_{\mathrm{st}}+\frac{H_{\mathrm{st}}}{P_{\mathrm{st}}}---\left(4\right)$

Wherein,

$P_{\mathrm{st}}=\underset{i=0}{\overset{s}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{t}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}}---\left(5\right)$

$H_{\mathrm{st}}=-\underset{i=0}{\overset{s}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{t}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}}\ln p_{\mathrm{ij}}---\left(6\right)$

In like manner, relevant with the probability distribution of background B entropy is

$H\left(B\right)=-\underset{i=s+1}{\overset{L-1}{\mathrm{\&Sigma;}}}\underset{j=t+1}{\overset{L-1}{\mathrm{\&Sigma;}}}\frac{p_{\mathrm{ij}}}{1-P_{\mathrm{st}}}\ln \frac{p_{\mathrm{ij}}}{1-P_{\mathrm{st}}}=-\frac{1}{1-P_{\mathrm{st}}}\underset{i=s+1}{\overset{L-1}{\mathrm{\&Sigma;}}}\underset{j=t+1}{\overset{L-1}{\mathrm{\&Sigma;}}}[p_{\mathrm{ij}}\ln p_{\mathrm{ij}}-p_{\mathrm{ij}}\ln (1-P_{\mathrm{st}})]---\left(7\right);$

$=\ln (1-P_{\mathrm{st}})+\frac{H_{L-1L-1}-H_{\mathrm{st}}}{1-P_{\mathrm{st}}}$

Wherein, $H_{L-1L-1}=-\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}}\ln p_{\mathrm{ij}};$

As long as in the zone (i=s+1 ... L-1 and j=1,2 ..t) and the zone (i=1,2 ..s and j=t+1 ..., L-1) middle p _Ij≈ 0, and more than Ding Yi H (B) just can set up, and hypothesis helps to save computing time like this.Because under many situations, off-diagonal probability distribution is negligible, it is rational therefore making above-mentioned hypothesis.

Criterion function Ψ (s t) is taken as H (O), H (B) sum, promptly

$\mathrm{\&Psi;}(s,t)=H\left(O\right)+H\left(B\right)=\ln P_{\mathrm{st}}+\frac{H_{\mathrm{st}}}{P_{\mathrm{st}}}+\ln (1-P_{\mathrm{st}}+\frac{H_{L-1L-1}-H_{\mathrm{st}}}{1-P_{\mathrm{st}}})$

(8)

$=\ln P_{\mathrm{st}}(1-P_{\mathrm{st}})+\frac{H_{\mathrm{st}}}{P_{\mathrm{st}}}+\frac{H_{L-1L-1}-H_{\mathrm{st}}}{1-P_{\mathrm{st}}}$

(s, t) (s t) promptly is the optimal threshold t that is asked to Zui Da gray-scale value to make Ψ ^*, promptly

$t^{*}=\mathrm{Arg}\left\{\underset{s,t\&Element;G}{\mathrm{Max}}\mathrm{\&Psi;}(s,t)\right\}---\left(9\right)$

Wherein Arg{} represents the function of negating.

At this, at (s, t) optimum of two parameter problems is found the solution the employing genetic algorithm optimization.Genetic algorithm is as a kind of global search method of finding the solution the efficient parallel of problem, its principal feature is the message exchange between colony's search strategy and the individual in population, can in search procedure, obtain automatically and the knowledge that accumulates relevant search volume, control search procedure adaptively in the hope of optimum solution or approximate optimal solution.Simple coding techniques of genetic algorithm utilization and reproduction mechanisms show complexity and the nonlinear problem that is difficult to the conventional search methods solution.

Initialization population: produce at random, select evenly to distribute.

Population scale: by analysis, population scale is popsize=20, and maximum evolutionary generation is g _Max=100.

Fitness function: directly use target function type (6) as fitness function.

Coding: because the gradation of image value between 0～255, adopts 8 gray level images to carry out emulation experiment, then threshold parameter satisfies 0≤s, t≤255, so individuality is encoded to 16 binary codes, most-significant byte is represented s, and least-significant byte is represented t.

Decoding: the most-significant byte of 16 binary codes is decoded as the s value, and least-significant byte is decoded as the t value.

Select: the employing ratio is selected operator;

Crossover operator: adopt two point to intersect, at random two point of crossing of Chan Shenging lay respectively at coded strings preceding 8 with back 8.Carry out interlace operation with the crossover probability Pc that configures in advance.Crossover probability is too high to mean that individual renewal is very fast, can reach bigger solution space, and can reduction obtain the probability of non-optimum solution, but the destructiveness of existing more excellent pattern is also increased thereupon.Crossover probability is low excessively, the search meeting because of the hunting zone reduce slow up.This paper chooses Pc=0.8 in search early stage, and the search later stage is chosen Pc=0.6.

Mutation operator: owing to adopt the binary coding mode, therefore variation is exactly the step-by-step negate.Mutation operation has been controlled the ratio that new gene is introduced population.The variation probability is too low, and some useful genes just can not be introduced, and the variation probability is too high, and promptly random variation is too many, and the offspring just may lose parents' good characteristic so.The present embodiment probability P m that will make a variation is defined as class parabolic type mutation operator:

$P_m=P_m\left(g\right)=P_{m\_\max }-\frac{4\&times;(P_{m\_\max }-P_{m\_\min })}{g_{\max }^2}{(g-\frac{g_{\max }}{2})}^2---\left(10\right)$

Wherein g is an evolutionary generation, and g ∈ [1, g _Max], P _{M_min}=P _b, P _{M_max}=10P _{M_min}

P _bBe basic variation probability, can estimate by following formula:

$P_b\&ap;\frac{175}{\mathrm{popsize}\&times;\sqrt{\mathrm{bits}}};$

Wherein bits represents individual figure place.Work as popsize=20, during bits=16, P _b≈ 0.02;

Work as P _b≈ 0.02, g _Max=100 o'clock, can get:

P _m＝P _m(g)＝0.2-0.00008×(g-50) ²；

Stop criterion: when algorithm was carried out the highest fitness value in maximum algebraically or the colony and stablized, algorithm was out of service, and having, the individuality of high fitness value is separating of being asked.When algorithm stopped, the individuality with the highest fitness value was as optimal threshold.

The basic step of piecemeal Two-dimensional maximum-entropy threshold segmentation algorithm is as follows:

(1) image segmentation is become a series of subimages;

(2) calculate the threshold values (calculating each threshold values) of each subimage with the Two-dimensional maximum-entropy algorithm;

(3) according to the threshold values of each subgraph, image is cut apart.

Image after cutting apart as shown in Figure 4.

4, motion target tracking

By determining to obtain the single frames testing result based on modified Two-dimensional maximum-entropy threshold segmentation algorithm, because the speed of the gentle motion of ducking in drink of foreign matter is slower, so target can be approximate with uniform rectilinear motion model.

The tracker model: writ attitude vector is $X\left(k\right)={[x\left(k\right),\stackrel{\&CenterDot;}{x}\left(k\right),y\left(k\right),\stackrel{\&CenterDot;}{y}\left(k\right)]}^T,$ Measuring vector is Z (k)=[x (k), y (k)] ^T, system equation can be write to be become:

X(k+1)＝ΦX(k)+Gw(k)(11)

Z(k+1)＝HX(k)+v(k)(12)

Wherein $\mathrm{\&Phi;}=\left[\begin{array}{cccc}1& T& 0& 0\\ 0& 1& 0& 0\\ 0& 0& 1& T\\ 0& 0& 0& 1\end{array}\right],$ $G=\left[\begin{array}{cc}\frac{{\mathrm{<figure-callout\; id="2"\; label="T"\; filenames="A20081014313800231.png"\; state="\{\{state\}\}">T</figure-callout>}}^2}{2}& 0\\ T& 0\\ 0& \frac{{\mathrm{<figure-callout\; id="2"\; label="T"\; filenames="A20081014313800231.png"\; state="\{\{state\}\}">T</figure-callout>}}^2}{2}\\ 0& T\end{array}\right],$ $H=\left[\begin{array}{cccc}1& 0& 0& 0\\ 0& 0& 1& 0\end{array}\right],$

W (k), v (k) are respectively that variance is the zero-mean white Gaussian noise of Q (k) and R (k).T is the sampling period in the formula.Track initial: if continuous 3 frames of possible target that detect appear in 3 * 3 neighborhoods, track initial then.Tracking gate and data association: suppose being estimated as of k-1 state vector constantly

The one-step prediction of state vector is:

$\hat{X}\left(k\right|k-1)=\mathrm{\&Phi;}\hat{X}(k-1|k-1)---\left(13\right)$

Then the new breath of system is:

$d\left(k\right)=Z\left(k\right)-H\hat{X}\left(k\right|k-1)---\left(14\right)$

New breath variance battle array is:

S(k)＝HP(k|k-1)H ^T+R(k)(15)

Wherein, P (k|k-1) is the one-step prediction variance.

The norm of the new vectorial d of breath of order (k) is:

g(k)＝d ^T(k)S ^-1(k)d(k)；(16)

Wherein g (k) obeys χ _M1 ²Distribute, M1 is for measuring dimension.

A definition ellipsoid (being called tracking gate again) in measurement space makes and measures with certain probability distribution in tracking gate

$\tilde{V}\left(\mathrm{\&gamma;}\right)=[Z:g\left(k\right)\&le;\mathrm{\&gamma;}]---\left(17\right)$

Wherein γ can pass through χ ²Distribution table checks in.

If after certain frame is handled, in tracking gate, have only an echo, then target trajectory directly upgrades; If have in the tracking gate more than an echo, then target trajectory upgrades by the nearest measured value of distance one-step prediction value.

The track of target upgrades by the standard Kalman filtering algorithm

X(k|k-1)＝ΦX(k-1|k-1)

P(k|k-1)＝ΦP(k-1|k-1)Φ ^T+GQ(k-1)G ^T

K(k)＝P(k?|k-1)H ^T[HP(k|k-1)H ^T+R] ^-1(18)

X(k|k)＝X(k|k-1)+K(k)[Z(k)-H(k)X(k|k-1)]

P(k|k)＝[I-K(k)H]P(k|k-1)

Fig. 5 has provided the movement locus figure that target following obtains.

In target detected and follow the tracks of after, if, can dope next step possible position of target according to target positional information and motion state before this because some optical considerations makes the of short duration disappearance of target, when target occurs once more, but tenacious tracking and be unlikely to lose objects still.The purpose of doing like this greatly reduces the noise erroneous judgement is the possibility of foreign matter.Because the movement velocity of foreign matter in liquid is slower, can draw the filter value of dbjective state vector by the described target following in top based on uniform rectilinear motion model, outside forecast obtains the possible position of target.

Suppose being estimated as of k state vector constantly

Then the predicted value in the target of k+n frame is:

$\hat{X}(k+n|k)={\mathrm{\&Phi;}}^n\hat{X}\left(k\right|k)---\left(19\right)$

When using in actual detected, along with the increase of time and the variation of possibility object of which movement, more total tracking data in past is more and more uncorrelated with following situation, and along with the increase of n, precision of prediction can descend.Here generally select n＜6.

5, Target Recognition and judgement

This step mainly is a technology of utilizing some image recognitions, supposes that for each bar that target following obtains it is that foreign matter forms or noise or bubble formation that track is distinguished track by following 3 principles, and judgment criterion is as follows:

(1) because image is taken when static after medicine bottle turns over turnback, visible foreign matters moves downward, so its centre of form ordinate should be to become big (is true origin with the upper left corner) successively;

(2) bubble moves upward, and direction of motion is opposite with foreign matter;

(3) geometric locus of foreign matter formation is smooth, and the track that noise forms is unordered.

Therefore, diminish successively, illustrate that this track is produced by bubble if detect the centre of form ordinate of movement locus; Become big track successively if detect the level and smooth and centre of form ordinate of movement locus, illustrate that foreign matter exists, write down judged result, and send the rejecting signal to system.The track of foreign matter and bubble and recognition result are as shown in Figure 5.

Claims (7)

1. the detection recognition methods of motion visible foreign matters and bubble in the liquid drug is characterized in that, may further comprise the steps:

1) current liquid drug to be detected is obtained the multiframe consecutive image;

2) image pre-service: the circular shuttering of employing 7 * 7 carries out high cap morphologic filtering as structural element to the image that obtains, and to remove and less all kinds of bright noise and the dark noise of structural element, obtains the denoising image;

3) moving target extracts: the kinetic characteristic that shows as interframe according to the foreign matter target on time domain, utilize the moving target in the novel three two field picture difference algorithms extraction image, its operating process is: first frame and second frame are done difference processing, second frame and the 3rd frame are done difference processing, strengthen target image by the image accumulation then, obtain to strengthen the back image;

4) moving Object Segmentation: to described enhancing back image utilization modified Two-dimensional maximum-entropy Threshold Segmentation Algorithm, the two-dimensional histogram that adopts gray scale and neighborhood averaging gray scale to constitute is selected threshold value, obtains cutting apart the back image;

5) motion target tracking: when detect cut apart after continuous 3 frames of possible target in the image appear in certain neighborhood, assert that then movement locus is initial; According to the regularity of foreign matter target in the image and bubble target travel, adopt based on the Kalman filtering of uniform rectilinear motion model and upgrade dbjective state, and generate oval tracking gate thus; When the measuring value that falls into oval tracking gate has only one, directly carry out target trajectory and upgrade; When the measuring value that falls into oval tracking gate more than one, target trajectory by distance further the nearest measured value of predicted value upgrade; When track rejection, the possible position of a few frame targets behind the filter value outside forecast in the moment keeps stable memory tracking power before adopting;

6) image recognition and judgement: obtain target trajectory by tracking results, utilize the continuity of target trajectory and directivity cog region to tell foreign matter and bubble target, thereby judge finally whether current liquid drug is qualified.

2. the detection recognition methods of motion visible foreign matters and bubble is characterized in that in the liquid drug as claimed in claim 1, described step 2) be:

If the gray scale function of image be f (x, y), structural element be B (x, y), (x y) is pixel coordinate, and then 4 kinds of basic operations are defined as follows:

Corrosion: (f Θ B) (x, y)=min{f (x+i, y+j)-B (i, j) | (x+i, y+j) ∈ D _f(i, j) ∈ D _B;

Expand: $(f\&CirclePlus;B)(x,y)=\max \{f(x-i,y-j)+B(i,j)|(x-i,y-j)\&Element;D_f;(i,j)\&Element;D_B\};$

Opening operation:

Closed operation: $f\&CenterDot;B=(f\&CirclePlus;B)\mathrm{\&Theta;B};$

In the formula, D _fAnd D _BBe respectively the field of definition of function f and B,

Introduce cap transformation filter operator: Hat (f)=f-(f о B);

(f о B) carries out the gray scale opening operation for structural element B to image f in the formula, selects for use 7 * 7 circular shuttering as structural element original image to be carried out high cap morphologic filtering, obtains containing the image of foreign matter target, bubble, i.e. the denoising image.

3. the detection recognition methods of motion visible foreign matters and bubble is characterized in that in the liquid drug as claimed in claim 1, and described step 3) is:

At the denoising image, according to image sequence first frame and second frame in every continuous three two field pictures are done difference processing, second frame and the 3rd frame are done difference processing, press following formula then and strengthen target image, obtain to strengthen the back image:

d(x，y)＝|f ₁(x，y)-f ₂(x，y)|×|f ₂(x，y)-f ₃(x，y)|

F wherein ₁(x, y), f ₂(x, y), f ₃(x y) is respectively through filtered first frame of high cap, second frame, the 3rd two field picture, | f ₁(x, y)-f ₂(x, y) | be first frame and the second frame difference image, | f ₂(x, y)-f ₃(x, y) | be second frame and three-frame difference image.

4. the detection recognition methods of motion visible foreign matters and bubble is characterized in that in the liquid drug as claimed in claim 1, and described step 4) may further comprise the steps:

(a) image segmentation is become a series of subimages;

(b) the Two-dimensional maximum-entropy algorithm computation goes out the threshold value of each subimage;

(c) according to the threshold value of each subgraph, image is cut apart;

Wherein the acquiring method of threshold value is as follows: establishing size of images is M * N, and image has L gray level, and the field average gray of each pixel also is a L gray level, and total gray level is L * L; On two dimensional surface, make the two dimensional gray histogram, gray-scale value and neighborhood average gray form the gray scale bi-values, the coordinate figure that is designated as each pixel for (i, j), the gray-scale value of i and j difference represent pixel point and the average gray in its field, the frequency that this gray scale combination occurs is f _Ij, this pixel functional value is joint probability density function p _Ij:

$p_{\mathrm{ij}}=\frac{f_{\mathrm{ij}}}{M\&times;N}$ i，j＝0，1，...L-1；

If so that (s t) is segmentation threshold, to the probability distribution p of object O _Ij(i=1,2 ... s; J=1,2 ... t) carry out the normalization operation; The definition normalization operation back entropy relevant with the probability distribution of object O is:

$H\left(O\right)=-\underset{i-0}{\overset{s}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{t}{\mathrm{\&Sigma;}}}\frac{p_{\mathrm{ij}}}{P_{\mathrm{st}}}\ln \frac{p_{\mathrm{ij}}}{P_{\mathrm{st}}}=-\frac{1}{P_{\mathrm{st}}}\underset{i=0}{\overset{s}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{t}{\mathrm{\&Sigma;}}}(p_{\mathrm{ij}}\ln p_{\mathrm{ij}}-p_{\mathrm{ij}}\ln P_{\mathrm{st}})=\ln P_{\mathrm{st}}+\frac{H_{\mathrm{st}}}{P_{\mathrm{st}}};$

Wherein, $P_{\mathrm{st}}=\underset{i=0}{\overset{s}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{t}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}},$ $H_{\mathrm{st}}=-\underset{i=0}{\overset{s}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{t}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}}\ln p_{\mathrm{ij}};$

The entropy relevant with the probability distribution of background B is

$H\left(B\right)=-\underset{i=s+1}{\overset{L-1}{\mathrm{\&Sigma;}}}\underset{j=t+1}{\overset{L-1}{\mathrm{\&Sigma;}}}\frac{p_{\mathrm{ij}}}{1-P_{\mathrm{st}}}\ln \frac{p_{\mathrm{ij}}}{1-P_{\mathrm{st}}}=-\frac{1}{1-P_{\mathrm{st}}}\underset{i=s+1}{\overset{L-1}{\mathrm{\&Sigma;}}}\underset{j=t+1}{\overset{L-1}{\mathrm{\&Sigma;}}}[p_{\mathrm{ij}}\ln p_{\mathrm{ij}}-p_{\mathrm{ij}}\ln (1-P_{\mathrm{st}})];$

$=\ln (1-P_{\mathrm{st}})+\frac{H_{L-1L-1}-H_{\mathrm{st}}}{1-P_{\mathrm{st}}}$

Wherein, $H_{L-1L-1}=-\underset{t=0}{\overset{L-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}}\ln p_{\mathrm{ij}};$

Criterion function Ψ (s t) is taken as H (O), H (B) sum, promptly

$\mathrm{\&Psi;}(s,t)=H\left(O\right)+H\left(B\right)=\ln P_{\mathrm{st}}+\frac{H_{\mathrm{st}}}{P_{\mathrm{st}}}+\ln (1-P_{\mathrm{st}}+\frac{H_{L-1L-1}-H_{\mathrm{st}}}{1-P_{\mathrm{st}}});$

$=\ln P_{\mathrm{st}}(1-P_{\mathrm{st}})+\frac{H_{\mathrm{st}}}{P_{\mathrm{st}}}+\frac{H_{L-1L-1}-H_{\mathrm{st}}}{1-P_{\mathrm{st}}}$

(s, t) (s t) promptly is the optimal threshold t that is asked to Zui Da gray-scale value to make Ψ ^*, promptly

$t^{*}=\mathrm{Arg}\left\{\underset{s,t\&Element;G}{\mathrm{Max}}\mathrm{\&Psi;}(s,t)\right\};$ Wherein Arg{} represents the function of negating.

5. the detection recognition methods of motion visible foreign matters and bubble is characterized in that in the liquid drug as claimed in claim 4, and described step 4) also comprises: the introducing genetic algorithm is optimized the partition threshold of Two-dimensional maximum-entropy method, to shorten the Flame Image Process time; Concrete steps are as follows:

Initialization population: produce at random, select evenly to distribute;

Population scale: by analysis, population scale is popsize=20, and maximum evolutionary generation is g _Max=100;

Fitness function: (s is t) as fitness function directly to use target function type Ψ;

Coding: because the gradation of image value between 0～255, adopts 8 gray level images, then threshold parameter satisfies 0≤s, t≤255, so individuality is encoded to 16 binary codes, most-significant byte is represented s, and least-significant byte is represented t;

Decoding: the most-significant byte of 16 binary codes is decoded as the s value, and least-significant byte is decoded as the t value;

Select: the employing ratio is selected operator;

Crossover operator: adopt two point to intersect, at random two point of crossing of Chan Shenging lay respectively at coded strings preceding 8 with back 8; Carry out interlace operation with crossover probability Pc; Early stage is chosen Pc=0.8 in search, and the search later stage is chosen Pc=0.6;

Mutation operator: mutation operation is the step-by-step negate; The probability P that will make a variation m is defined as class parabolic type mutation operator:

$P_m=P_m\left(g\right)=P_{m\_\max }-\frac{4\&times;(P_{m\_\max }-P_{m\_\min })}{g_{\max }^2}{(g-\frac{g_{\max }}{2})}^2$

Wherein g is an evolutionary generation, and g ∈ [1, g _Max], P _{M_min}=P _b, P _{M_max}=10P _{M_min}

P _bBe basic variation probability, estimate by following formula:

$P_b\&ap;\frac{175}{\mathrm{popsize}\&times;\sqrt{\mathrm{bits}}};$

Wherein bits represents individual figure place; Work as popsize=20, during bits=16, P _b≈ 0.02;

Work as P _b≈ 0.02, g _Max=100 o'clock, can get:

P _m＝P _m(g)＝0.2-0.00008×(g-50) ²；

Stop criterion: when algorithm was carried out the highest fitness value in maximum algebraically or the colony and stablized, algorithm was out of service, and having, the individuality of high fitness value is separating of being asked; When algorithm stopped, the individuality with the highest fitness value was as optimal threshold.

6. the detection recognition methods of motion visible foreign matters and bubble is characterized in that in the liquid drug as claimed in claim 1, and described step 5) may further comprise the steps:

By determining to obtain the single frames testing result, that target is approximate with uniform rectilinear motion model based on modified Two-dimensional maximum-entropy threshold segmentation algorithm; Its tracker model is: writ attitude vector is $X\left(k\right)={[x\left(k\right),\stackrel{\&CenterDot;}{x}\left(k\right),y\left(k\right),\stackrel{\&CenterDot;}{y}\left(k\right)]}^T,$ Measuring vector is Z (k)=[x (k), y (k)] ^T, system equation can be write to be become

X(k+1)＝ΦX(k)+Gw(k)；

Z(k+1)＝HX(k)+v(k)；

Wherein $\mathrm{\&Phi;}=\left[\begin{array}{cccc}1& T& 0& 0\\ 0& 1& 0& 0\\ 0& 0& 1& T\\ 0& 0& 0& 1\end{array}\right],$ $G=\left[\begin{array}{cc}\frac{T^2}{2}& 0\\ T& 0\\ 0& \frac{T^2}{2}\\ 0& T\end{array}\right],$ $H=\left[\begin{array}{cccc}1& 0& 0& 0\\ 0& 0& 1& 0\end{array}\right],$

W (k), v (k) are respectively that variance is the zero-mean white Gaussian noise of Q (k) and R (k); T is the sampling period in the formula; Track initial: if continuous 3 frames of possible target that detect appear in 3 * 3 neighborhoods, track initial then;

Tracking gate and data association: suppose being estimated as of k-1 state vector constantly

The one-step prediction of state vector is: $\hat{X}\left(k\right|k-1)=\mathrm{\&Phi;}\hat{X}(k-1|k-1);$

Then the new breath of system is:

$d\left(k\right)=Z\left(k\right)-H\hat{X}\left(k\right|k-1);$

New breath variance battle array is:

S(k)＝HP(k|k-1)H ^T+R(k)；

Wherein, P (k|k-1) is the one-step prediction variance;

The norm of the new vectorial d of breath of order (k) is:

g(k)＝d ^T(k)S ^-1(k)d(k)；

Wherein g (k) obeys χ _M1 ²Distribute, M1 is for measuring dimension;

Tracking gate of definition in measurement space makes and measures with certain probability distribution in tracking gate;

$\tilde{V}\left(\mathrm{\&gamma;}\right)=[Z:g\left(k\right)\&le;\mathrm{\&gamma;}],$ Wherein γ can pass through χ ²The distribution difference checks in;

If after certain frame is handled, in tracking gate, have only an echo, then target trajectory directly upgrades; If have in the tracking gate more than an echo, then target trajectory upgrades by the nearest measured value of distance one-step prediction value;

The track of target upgrades by the standard Kalman filtering algorithm:

X(k|k-1)＝ΦX(k-1|k-1)

P(k|k-1)＝ΦP(k-1|k-1)Φ ^T+GQ(k-1)G ^T

K(k)＝P(k|k-1)H ^T[HP(k|k-1)H ^T+R] ^-1；

X(k|k)＝X(k|k-1)+K(k)[Z(k)-H(k)X(k|k-1)]

P(k|k)＝[I-K(k)H]P(k|k-1)

In target detected and follow the tracks of after, if the of short duration disappearance of target, can be according to target positional information and motion state before this, dope next step possible position of target, when target occurs once more, but tenacious tracking and be unlikely to lose objects still, and detailed process is as follows:

Suppose being estimated as of k state vector constantly

Then the predicted value in the target of k+n frame is:

$\hat{X}(k+n|k)={\mathrm{\&Phi;}}^n\hat{X}\left(k\right|k);$ Select n＜6.

7. as the detection recognition methods of motion visible foreign matters and bubble in each described liquid drug of claim 1～6, it is characterized in that, described step 6) is: suppose that for each bar that target following obtains it is that foreign matter forms or noise or bubble formation that track is distinguished track by following 3 principles, judgment criterion is as follows:

(a) because image is taken when static after medicine bottle turns over turnback, visible foreign matters moves downward, so its centre of form ordinate should be to become big successively, is true origin with the upper left corner;

(b) bubble moves upward, and direction of motion is opposite with foreign matter;

(c) geometric locus of foreign matter formation is smooth, and the track that noise forms is unordered;

Therefore, diminish successively, illustrate that this track is produced by bubble if detect the centre of form ordinate of movement locus; Become big track successively if detect the level and smooth and centre of form ordinate of movement locus, the foreign matter existence be described, judge that current liquid drug is defective.

Images