CN102800107B - Motion target detection method based on improved minimum cross entropy - Google Patents

Abstract

The invention discloses a motion target detection method based on improved minimum cross entropy. The method comprises the following steps: 1) differencing images; 2) carrying out binarization: 2.1) calculating and normalizing the grey level and gradient; 2.2) calculating the grey level-gradient co-occurrence matrix and the probability thereof; 2.3) calculating two-dimensional minimum cross entropy; 2.4) selecting optimal threshold based on genetic algorithm; and 2.5) evaluating separation quality based on uniformity measure; 3) carrying out morphology operation: expanding and carrying out corrosion processing on binarization images; 4) analyzing communicated areas for acquiring the communicated domain of each object in the images; and 5) recognizing the target: starting from the pixel point at the left upper corner in the binarization images, and if the current pixel point is 1, searching for the pixel points with the pixel value of 1 for 8 fields by adopting a greedy algorithm, and storing the position information of each target and the related information into a target chain, wherein the detected target information is in the target chain. According to the invention, the accuracy is high, and the timeliness is good.

Description

A kind of moving target detecting method based on improving minimum cross entropy

Technical field

The present invention relates to the technical fields such as image procossing, computer vision, pattern-recognition, especially a kind of object detection method for motion fish volume image.

Background technology

Target following, feature extraction and target identification etc. are the important steps of image understanding in computer vision, and they are all closely related with the quality of target detection.Target detection process can be divided into four steps such as image difference, binaryzation, shape filtering and connectivity analysis, wherein threshold value be chosen in target detection process particularly important.

Kullback proposes the method for information theory of difference between two kinds of probability distribution in the same event space of tolerance, and is widely used.Cross entropy is incorporated in Iamge Segmentation by the people such as LE.C.H, utilizes minimum cross entropy criterion to select threshold value, makes the quantity of information difference between original image and segmentation image minimum.Xue Jing is great subsequently waits people to utilize posterior probability to solve cross entropy between one dimension maximum kind, and the people such as Fan Lisheng propose the polarization Cross-Entropy Algorithm being used for Polarimetric SAR Image.One dimension cross entropy threshold method is generalized to two-dimensional space by the people such as Mei Rong, and utilizes optimized algorithm to consume to avoid a large amount of computing times; The calculated amount that the people such as Chen Qi bring in view of two-dimension method, proposes FAST RECURSIVE ALGORITHM.

Meanwhile, in information theory, the concept of Shannon entropy is also applied in Iamge Segmentation, which depict the degree of uncertainty of a probability distribution.A lot of researcher both at home and abroad on this basis, proposes the maximum entropy threshold method of a peacekeeping two dimension respectively subsequently.For taking into full account the texture information of image, the people such as Zhou Delong propose maximum entropy threshold partitioning algorithm on Gray Level-Gradient Co-occurrence Matrix model basis, and the gray scale of the method combining image and gradient information provide the theoretical foundation that threshold value is chosen.The people such as Zhang Hong propose again the weighting conditions entropy threshold algorithm based on Gray Level-Gradient Co-occurrence Matrix model on this basis.Because two-dimensional entropy calculated amount is huge, consuming time longer, be difficult to practical, the congruent people of Wu one proposes the FAST RECURSIVE ALGORITHM of threshold segment in 2 D entropy image, improves arithmetic speed.The people such as Long Jianwu, for the shortcoming of the maximum entropy partitioning algorithm noise resisting ability difference based on shade of gray model, also been proposed the maximum entropy threshold segmentation algorithm based on three-dimensional model, and provide quick recursion formula.

In piece image, the gray scale of each picture point is the basis forming piece image, the internal information of reflection image; Gradient is then the key element of composing images edge contour, and the principal outline information of piece image is provided by its edge contour information.In Gray Level-Gradient Co-occurrence Matrix, the gray scale of two kinds of key elements the most basic and picture point and the mutual relationship of gradient (or edge) in image are embodied a concentrated reflection of, its element c _ijbe defined as: in normalized gray level image F (x, y) and normalized gradient image G (x, y), statistics makes gray-scale value be i and Grad to be total picture point number of j simultaneously.

In Gray Level-Gradient space, the grey value profile of image inside is relatively more even, therefore Grad is less; Marginal point obviously distinguishes object and background, therefore its Grad is larger.Therefore, the Gray level-gradient co-occurrence matrix model space very clearly depicts the border condition in image between prospect and background, and the regularity of distribution of gray scale and gradient, give also the spatial relationship of each pixel and its field pixel simultaneously, can the textural characteristics of rendering image well.

Genetic algorithm is that the evolution laws of a class reference organic sphere develops and next random search algorithm, is the optimized algorithm with robustness, concurrency and adaptivity.It is to realize easily and precision high is applied to solving the actual optimization problem in various field.Genetic algorithm main processing steps is as follows: 1) encode, one is encoded to a chromosome, and the object of coding is for the form of expression of optimization problem solution and the calculating after being beneficial in genetic algorithm.2) structure of fitness function and application, fitness function is determined according to the objective function of optimization problem substantially.After fitness function is determined, natural selection rule determines which chromosome adapts to existence with the distribution of the size decisive probability of fitness function value, and which is eliminated.The chromosome composition population survived, forms one and can multiply follow-on colony.3) chromosomal combination, the gene of parents combine be by coding between mating reach the generation of new explanation of future generation.4) make a variation, may make a variation by producer in new explanation production process, variation makes some coding change, and makes solution have larger ergodicity.

At present relatively more conventional method for testing motion can be divided three classes: based on the method for inter-frame difference, the method based on background difference and the method based on light stream.The defect existed: 1, effectively can not be partitioned into motion fish body target; 2, more responsive to illumination patterns.

Summary of the invention

In order to deficiencies such as the poor and poor real of the degree of accuracy overcoming existing object detection method, the present invention chooses in the threshold value in binaryzation stage, Gray Level-Gradient Co-occurrence Matrix is incorporated in 2-D minimum cross entropy thresholding algorithm, utilize genetic algorithm approximate solution optimal threshold, and adopt homogeneity to estimate evaluation index as picture quality, propose the moving target detecting method based on minimum cross entropy that a kind of degree of accuracy is high, real-time is good.

The technical solution adopted for the present invention to solve the technical problems is:

Based on the moving target detecting method improving minimum cross entropy, described object detection method comprises the following steps:

1) image difference: utilize image difference to assign to obtain difference image;

2) binaryzation: binarization can regard image segmentation process as again, comprises the following steps:

2.1), the gray-scale value of computed image and Grad, and gray-scale value and Grad to be normalized;

2.2), calculate each element value of Gray Level-Gradient Co-occurrence Matrix, and calculate the probability matrix of its correspondence;

2.3), according to step 2.2) Gray Level-Gradient Co-occurrence Matrix that obtains and probability matrix thereof carry out the 2-D minimum cross entropy of computed image;

2.4), utilize genetic algorithm and minimum cross entropy formula to choose the optimal threshold of image, namely meet the segmentation threshold that cross entropy in image is minimum;

3) morphological operation: first binary image is expanded and carry out corrosion treatment again.

4) connected component analysis;

5) visual inspection identification: in binary image from top left corner pixel point, if current pixel point is 1, then adopting greedy algorithm to search for its 8 field pixel value is the pixel of 1.Until search is the pixel of 1 less than pixel or has traveled through all picture positions; While search, set up object chain, the positional information of each target and relevant information all will be kept in object chain, after target identifying terminates, namely be the target information detected in object chain.

Further, in described step 1), with the background subtraction based on fuzzy reasoning.

Further again, described binarization also comprises: 2.5), for the segmentation threshold chosen, utilize homogeneity to estimate and evaluate.

Further, in described step 3), adopt the regional analysis based on region growth method and wire tag method.

Technical conceive of the present invention is: Iamge Segmentation is one of basis and important field the most in Image processing and compute machine visual field Level Visual, and it is basic premise image being carried out to visual analysis and pattern-recognition.Thresholding method is a kind of tradition and classical method in Iamge Segmentation field, because of its realize simple, calculated amount is little, performance is comparatively stable and to become in Iamge Segmentation the most most widely used cutting techniques of fundamental sum, has been applied to a lot of fields.Up to the present, various Research on threshold selection has been proposed.

The segmentation principle of thresholding method is: by setting different characteristic threshold value, the pixel of image is divided into some classes according to threshold value.The general feature used in cutting procedure comprises gray-scale value, color property etc.

Gray Level-Gradient Co-occurrence Matrix is incorporated in 2-D minimum cross entropy thresholding algorithm by this method, utilize genetic algorithm approximate solution optimal threshold, and adopt homogeneity to estimate evaluation index as picture quality, achieve a kind of moving target detecting method based on minimum cross entropy.

Beneficial effect of the present invention is mainly the computation process accelerating 2-D minimum cross entropy algorithm on the one hand, improves the validity of target detection; Choose on the other hand in threshold value the inside and marginal information that can take into account image, effectively suppress the uneven impact on segmentation effect of illumination patterns, improve the robustness of object detection method.

Accompanying drawing explanation

Fig. 1 is a kind of moving target detecting method process flow diagram based on improving minimum cross entropy.

Fig. 2 is the background subtraction schematic diagram based on fuzzy reasoning.

Fig. 3 is two-dimensional gray histogram.

Fig. 4 is Gray Level-Gradient Co-occurrence Matrix quadrantal diagram.

Fig. 5 is genetic algorithm selected threshold process flow diagram.

Embodiment

Below in conjunction with accompanying drawing, the invention will be further described.

With reference to Fig. 1 ~ Fig. 5, a kind of moving target detecting method based on improving minimum cross entropy, described object detection method comprises the following steps:

1), image difference:

With reference to Fig. 2, image difference is used for obtaining difference image, because applied environment background of the present invention is relatively simple, therefore adopt the background subtraction based on fuzzy reasoning, its main flow figure as shown in Figure 2, asks for an interview list of references about the background subtraction based on fuzzy reasoning: Chen Yong, Xiao Gang in detail, Chen Jiujun etc. based on the object detection method [J] of fuzzy reasoning background segment. computer engineering, o. 11th the 36th volume 190-194 page.

2), binaryzation:

2.1), the calculating of gray scale and gradient and normalization.

Suppose to be expressed as f (x, y) at the pixel value at width gray level image (x, a y) place, so normalized image gray-scale value:

F(x,y)=INT(f(x,y)×L/f _max)+1 （1）

In formula, INT represents rounding operation, and L represents the maximal value of Normalized Grey Level value, f _maxrepresent the maximal value of gray scale in original image.

In the present invention, the Grad of each pixel of image adopts Laplace operator to calculate, and formula is as follows:

g(m,n)=4×f(m,n)-f(m+1,n)-f(m-1,n)-f(m,n+1)-f(m,n-1) （2）

In like manner, normalized gradient is as follows:

G(x,y)=INT(g(x,y)×L'/g _max)+1 （3）

In formula, g _maxrepresent the maximal value of gradient in image, L' represents the maximal value of normalized gradient value.

2.2), Gray Level-Gradient Co-occurrence Matrix and probability calculation thereof.

For the Normalized Grey Level image F (x obtained in above-mentioned calculating, y) with gradient image G (x, y), calculate and meet F (x simultaneously, y)=i and G (x, the total number of the pixel of y)=j, its result is (i, j) the individual element value c in Gray Level-Gradient Co-occurrence Matrix _ij, wherein i represents gray-scale value, and j represents Grad.

So, the new probability formula of gray-scale value to be i and Grad the be co-occurrence matrix correspondence position of j is expressed as follows:

$p_{\mathrm{ij}}=c_{\mathrm{ij}}/\left(\underset{i=0}{\mathrm{\Σ}}\underset{j=0}{\mathrm{\Σ}}{c}_{\mathrm{ij}}\right)---\left(4\right)$

2.3), the calculating of 2-D minimum cross entropy.

For the two dimensional image histogram be made up of gray scale-field gray scale, as shown in figure 3-1.In figure, L represents maximum gray scale, and X-coordinate axle represents gray-scale value, and Y-coordinate axle represents field gray-scale value.The informational probability of edge and noise is less, can ignore region C and D and disregard, and therefore, if when threshold value is located at (s, t), then has:

$P_A=\underset{i=0}{\overset{s}{\mathrm{\Σ}}}\underset{j=0}{\overset{t}{\mathrm{\Σ}}}{p}_{\mathrm{ij}},$ $P_B=\underset{i=s+1}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=t+1}{\overset{L-1}{\mathrm{\Σ}}}{p}_{\mathrm{ij}}---\left(5\right)$

If μ _a(s, t), μ _baverage in the class of the target and background after (s, t) expression segmentation, then have:

${\mathrm{\μ}}_A(s,t)=\frac{1}{P_A}\underset{i=0}{\overset{s}{\mathrm{\Σ}}}\underset{j=0}{\overset{t}{\mathrm{\Σ}}}{\mathrm{ijp}}_{\mathrm{ij}}---\left(6\right)$

${\mathrm{\μ}}_B(s,t)=\frac{1}{P_B}\underset{i=s+1}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=t+1}{\overset{L-1}{\mathrm{\Σ}}}{\mathrm{ijp}}_{\mathrm{ij}}---\left(7\right)$

Therefore definable two dimension cross entropy is:

$D(A,B,s,t)=\underset{i=0}{\overset{s}{\mathrm{\Σ}}}\underset{j=0}{\overset{t}{\mathrm{\Σ}}}({\mathrm{ijp}}_{\mathrm{ij}}\ln \frac{\mathrm{ij}}{{\mathrm{\μ}}_A(s,t)}+{\mathrm{\μ}}_A(s,t)p_{\mathrm{ij}}\ln \frac{{\mathrm{\μ}}_A(s,t)}{\mathrm{ij}})+$ （8）

$\underset{i=s+1}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=t+1}{\overset{L-1}{\mathrm{\Σ}}}({\mathrm{ijp}}_{\mathrm{ij}}\ln \frac{\mathrm{ij}}{{\mathrm{\μ}}_B(s,t)}+{\mathrm{\μ}}_B(s,t)p_{\mathrm{ij}}\ln \frac{{\mathrm{\μ}}_B(s,t)}{\mathrm{ij}})$

The optimal threshold vector chosen should meet:

D(A,B,s ^*,t ^*)=max{1/D(A,B,s,t)} （9）

The above-mentioned definition procedure being 2-D minimum cross entropy.

From above-mentioned steps, the Gray Level-Gradient Co-occurrence Matrix of image is the matrix [c of L × L' dimension _ij] _{l × L'}, be the quadrantal diagram of Gray Level-Gradient Co-occurrence Matrix as shown in Figure 4, in figure, X-coordinate axle represents Grad, and Y-coordinate axle represents gray-scale value, and the initial point of matrix, in X, Y-axis intersection, represents that Grad increases to the right, represents that gray-scale value increases downwards.

Suppose in upper figure, threshold value is at (s, t) place and the gray-scale value of target is on the low side, otherwise the gray-scale value of background then.Because the grey level distribution of the target and background inside in image is very even, thus the Grad of this part is zero or very little; Grad for marginal portion is often larger.Accordingly, Gray Level-Gradient Co-occurrence Matrix is divided into 4 quadrants, quadrant A represents target information, and quadrant D represents background information, and quadrant B and C all represents edge and noise information.From Fig. 3-4, quadrant B is the marginal information of target area to background area transition, and quadrant C is the marginal information of background area to object area transition.Because the information edge informational influence that quadrant A and D comprises is very little, namely probability is less, so negligible.By B and C quadrant probability calculation formula in formula (5) known co-occurrence matrix be:

$P_B=\underset{i=0}{\mathrm{\Σ}}\underset{j=t+1}{\mathrm{\Σ}}{p}_{\mathrm{ij}},$ $P_C=\underset{i=s+1}{\mathrm{\Σ}}\underset{j=t+1}{\mathrm{\Σ}}{p}_{\mathrm{ij}}---\left(10\right)$

If μ _b(s, t), μ _caverage in the class of the target and background after (s, t) expression segmentation, then have:

${\mathrm{\μ}}_B(s,t)=\frac{1}{P_B}\underset{i=0}{\overset{s}{\mathrm{\Σ}}}\underset{j=t+1}{\overset{L^{\′}-1}{\mathrm{\Σ}}}{\mathrm{ijp}}_{\mathrm{ij}}---\left(11\right)$

${\mathrm{\μ}}_C(s,t)=\frac{1}{P_C}\underset{i=s+1}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=t+1}{\overset{L^{\′}-1}{\mathrm{\Σ}}}{\mathrm{ijp}}_{\mathrm{ij}}---\left(12\right)$

The cross entropy formula that can obtain image is:

$D(B,C,s,t)=\underset{i=0}{\overset{s}{\mathrm{\Σ}}}\underset{j=t+1}{\overset{L^{\′}-1}{\mathrm{\Σ}}}({\mathrm{ijp}}_{\mathrm{ij}}\ln \frac{\mathrm{ij}}{{\mathrm{\μ}}_B(s,t)}+{\mathrm{\μ}}_B(s,t)p_{\mathrm{ij}}\ln \frac{{\mathrm{\μ}}_B(s,t)}{\mathrm{ij}})+$

$\underset{i=s+1}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=t+1}{\overset{L^{\′}-1}{\mathrm{\Σ}}}({\mathrm{ijp}}_{\mathrm{ij}}\ln \frac{\mathrm{ij}}{{\mathrm{\μ}}_C(s,t)}+{\mathrm{\μ}}_C(s,t)p_{\mathrm{ij}}\ln \frac{{\mathrm{\μ}}_C(s,t)}{\mathrm{ij}})---\left(13\right)$

For the quantity of information obtaining target and background and edge in image is maximum, optimal threshold vector (s ^*, t ^*) should meet:

$(s^{*},t^{*})=\arg \underset{0<t<L^{\&prime;}-1}{\underset{0<s<L-1}{\max }}\{1/D(B,C,s,t)\}---\left(14\right)$

2.4) optimal threshold, based on genetic algorithm is chosen.

In above-mentioned optimal threshold vector form calculates, if exhaustive formula (9) all solutions, then draw optimal threshold, the time that its computation process expends is very long.The people such as the congruent people of Wu one and Long Jianwu propose to use quick recursion formula to solve for two and three dimensions entropy respectively, but its computing velocity is still very consuming time.Genetic algorithm (GeneticAlgorithm, GA) be a kind of evolution laws using for reference organic sphere to find the method for optimum solution, the adaptability that can realize each gene string improves.For improving solution efficiency further, the present invention's employing is carried out conditional entropy optimal threshold based on genetic algorithm and is chosen.

Genetic algorithm is combined with Minimum Cross-Entropy Algorithm, needs to consider two problems: the solution of (1) how encoded question; (2) the every bar chromosome of tolerance how is constructed to the fitness function of the adaptedness of problem.The object that these two kinds of algorithms combine is that the cross entropy making image is minimum, optimum segmentation threshold (s can be obtained ^*, t ^*).The present invention, in order to improve the execution efficiency of algorithm, directly encodes to Two Dimensional Thresholding vector, and the threshold value of so every bar gene string encoding representative can make the inverse of cross entropy after segmentation maximum, thus makes its fitness the highest.

In the present invention, the basic procedure of genetic algorithm as shown in Figure 5, wherein, population number gets 30-50, iterations is got 10-30 time, mating rate is chosen for (0.25,0.75), and aberration rate is chosen for (0.01,0.2), the too large too little performance that all can affect algorithm of these parameter choose.

With reference to Fig. 4, the key step of genetic algorithm selected threshold is as follows:

A) set population number POPSIZE and iterations MAXGENS, binary coding is carried out to bivector, and random generation initialization population.Population number generally gets 30-50, and iterations is generally got 10-30 time, and too little meeting makes the performance of genetic algorithm become very poor; Then can increase calculated amount too greatly, make convergence time elongated.

B) according to the actual requirements, the variation range of variable in fitness function is determined.Initialized population is decoded, and according to the fitness of formulae discovery every bar gene string.

C) adopt wheel robin unconditionally to copy in new population of future generation by gene string large for fitness, then to genetic operator computings such as the selection of parent population, mating and variations, thus obtain the gene string of new population of future generation.Mating and variation are the genetic operators producing new population, if mating rate is too large, are destroyed very soon by the gene string structure making fitness high, too little, search are stagnated, so generally get (0.25,0.75).If aberration rate is too large, genetic algorithm will be made to become random search, too little, new individuality can not be produced, generally get (0.01,0.2).

D) multiply according to the iterations of setting, reach end condition and then return best gene string, and it can be used as two-dimentional segmentation threshold; Otherwise return and c) continue follow-on procreation.

2.5), based on the segmentation quality assessment that homogeneity is estimated.

In order to verify the validity of this algorithm further, the homogeneity that algorithm adopts the people such as Lei Bo to mention is estimated as evaluation index.Homogeneity estimates one of index being used to evaluate image partition method performance, is to describe the homogeneity of same class pixel value distribution in image.In image, the homogeneity in a region and the variance in this region are inversely proportional to, and region internal variance is less, represent that area uniformity is better.

Suppose given image-region Ω, the gray-scale value f (x, y) that the pixel in region is corresponding represents, uses represent the gray average in region, variance is used represent, the area (i.e. the sum of picture point) in region uses A _Ωrepresent.For the threshold value T searched out, gray average computing formula is as follows:

$f_{\mathrm{\&Omega;}}=\underset{(x,y)\&Element;\mathrm{\&Omega;}}{\mathrm{\&Sigma;}}f(x,y)/A_{\mathrm{\&Omega;}}---\left(15\right)$

Can formula of variance be obtained as follows by mean value formula:

${\mathrm{\&sigma;}}_{\mathrm{\&Omega;}}^2=\underset{(x,y)\&Element;\mathrm{\&Omega;}}{\mathrm{\&Sigma;}}(f(x,y)-f_{\mathrm{\&Omega;}})/A_{\mathrm{\&Omega;}}---\left(16\right)$

So the homogeneity of image is estimated U computing formula and can be expressed as follows:

$U=1-\mathrm{\&Sigma;}{\mathrm{\&sigma;}}_{\mathrm{\&Omega;}}^2/C---\left(17\right)$

In formula, C represents region normalized parameter, is used for limiting the value estimated of homogeneity in (0,1), and its computing formula is as follows:

C=(f _max-f _min) ²/2 （18）

In formula, f _max, f _minrepresent minimum and maximum gray-scale value in image respectively.

Homogeneity estimates the main internal information considering image object and background, can evaluate the threshold value chosen like this and whether have good segmentation quality.According to the optimal threshold s solved ^*, can calculate homogeneity by formula (15) to (18) and estimate U, its numerical value larger expression segmentation quality is better; Otherwise then quality is poorer.

Given threshold T _uif homogeneity is estimated U value and is greater than threshold value T _u, then next step is entered; If homogeneity estimates U be less than or equal to threshold value T _u, then need to forward step 2.4 to), recalculate suboptimum threshold value, until evaluation quality is greater than threshold value T _u.

3), morphological operation.

Morphological operation is exactly carry out dilation and erosion operation to the image after binaryzation, and these two kinds of operations are bases of Morphological scale-space, and expansive working is by crack bridge joint, and corrosion is from binary image, eliminate incoherent details.The present invention is first expanded to binary image and is carried out corrosion treatment again.

4), connected component analysis.

Connected component analysis method is just used to the connected domain obtaining each object in image, also referred to as connected component labeling (Connected Component Labeling, CCL).The present invention adopts a kind of connected component analysis method based on region growth method and wire tag method.

5), visual inspection identification

After the process of above four steps, in binary image from top left corner pixel point, if current pixel point is 1, then adopting greedy algorithm to search for its 8 field pixel value is the pixel of 1.Until search is the pixel of 1 less than pixel or has traveled through all picture positions.While search, set up object chain, the positional information of each target and relevant information all will be kept in object chain, after target identifying terminates, namely be the target information detected in object chain.

Claims (4)

1., based on the moving target detecting method improving minimum cross entropy, it is characterized in that: described object detection method comprises the following steps:

1) image difference: utilize image difference to be used for obtaining difference image;

2) binaryzation: binarization can regard image segmentation process as again, comprises following process:

2.1), the gray-scale value of computed image and Grad, and gray-scale value and Grad to be normalized:

Suppose to be expressed as f (x, y) at the pixel value at width gray level image (x, a y) place, so normalized image gray-scale value:

F(x,y)＝INT(f(x,y)×L/f _max)+1 (1)

In formula, INT represents rounding operation, and L represents the maximal value of Normalized Grey Level value, f _maxrepresent the maximal value of gray scale in original image;

The Grad of each pixel of image adopts Laplace operator to calculate, and formula is as follows:

g(m,n)＝4×f(m,n)-f(m+1,n)-f(m-1,n)-f(m,n+1)-f(m,n-1) (2)

In like manner, normalized gradient is as follows:

G(x,y)＝INT(g(x,y)×L'/g _max)+1 (3)

In formula, g _maxrepresent the maximal value of gradient in image, L' represents the maximal value of normalized gradient value;

2.2), calculate each element value of Gray Level-Gradient Co-occurrence Matrix, and calculate the probability matrix of its correspondence;

For the Normalized Grey Level image F (x obtained in above-mentioned calculating, y) with gradient image G (x, y), calculate and meet F (x simultaneously, y)=i and G (x, the total number of the pixel of y)=j, its result is (i, j) the individual element value c in Gray Level-Gradient Co-occurrence Matrix _ij, wherein i represents gray-scale value, and j represents Grad;

So, the new probability formula of gray-scale value to be i and Grad the be co-occurrence matrix correspondence position of j is expressed as follows:

$p_{\mathrm{ij}}=c_{\mathrm{ij}}/\left(\underset{i=0}{\mathrm{\&Sigma;}}\underset{j=0}{\mathrm{\&Sigma;}}{c}_{\mathrm{ij}}\right)---\left(4\right)$

2.3), according to step 2.2) Gray Level-Gradient Co-occurrence Matrix that obtains and probability matrix thereof carry out the 2-D minimum cross entropy of computed image;

L represents the maximal value of Normalized Grey Level value, X-coordinate axle represents gray-scale value, Y-coordinate axle represents field gray-scale value, Gray Level-Gradient Co-occurrence Matrix is divided into 4 quadrants, quadrant A represents target information, quadrant D represents background information, and quadrant B is the marginal information of target area to background area transition, and quadrant C is the marginal information of background area to object area transition; The informational probability of edge and noise is less, can ignore region C and D and disregard, and therefore, if when threshold value is located at (s, t), then has:

$P_A=\underset{i=0}{\overset{s}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{t}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}},P_B=\underset{i=s+1}{\overset{L-1}{\mathrm{\&Sigma;}}}\underset{j=t+1}{\overset{L-1}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}}---\left(5\right)$

If μ _a(s, t), μ _baverage in the class of the target and background after (s, t) expression segmentation, then have:

${\mathrm{\&mu;}}_A(s,t)=\frac{1}{P_A}\underset{i=0}{\overset{s}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{t}{\mathrm{\&Sigma;}}}\mathrm{ij}{p}_{\mathrm{ij}}---\left(6\right)$

${\mathrm{\&mu;}}_B(s,t)=\frac{1}{P_B}\underset{i=s+1}{\overset{L-1}{\mathrm{\&Sigma;}}}\underset{j=t+1}{\overset{L-1}{\mathrm{\&Sigma;}}}\mathrm{ij}{p}_{\mathrm{ij}}---\left(7\right)$

Therefore definable two dimension cross entropy is:

$\begin{array}{c}D(A,B,s,t)=\underset{i=0}{\overset{s}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{t}{\mathrm{\&Sigma;}}}({\mathrm{ijp}}_{\mathrm{ij}}\ln \frac{\mathrm{ij}}{{\mathrm{\&mu;}}_A(s,t)}+{\mathrm{\&mu;}}_A(s,t)p_{\mathrm{ij}}\ln \frac{{\mathrm{\&mu;}}_A(s,t)}{\mathrm{ij}})+\\ \underset{i=s+1}{\overset{L-1}{\mathrm{\&Sigma;}}}\underset{j=t+1}{\overset{L-1}{\mathrm{\&Sigma;}}}({\mathrm{ijp}}_{\mathrm{ij}}\ln \frac{\mathrm{ij}}{{\mathrm{\&mu;}}_B(s,t)}+{\mathrm{\&mu;}}_B(s,t)p_{\mathrm{ij}}\ln \frac{{\mathrm{\&mu;}}_B(s,t)}{\mathrm{ij}})\end{array}---\left(8\right)$

The optimal threshold vector chosen should meet:

D(A,B,s ^*,t ^*)＝max{1/D(A,B,s,t)} (9)

The above-mentioned definition procedure being 2-D minimum cross entropy;

By B and C quadrant probability calculation formula in formula (5) known co-occurrence matrix be:

$P_B=\underset{i=0}{\mathrm{\&Sigma;}}\underset{j=t+1}{\mathrm{\&Sigma;}}{p}_{\mathrm{ij}},P_C=\underset{i=s+1}{\mathrm{\&Sigma;}}\underset{j=t+1}{\mathrm{\&Sigma;}}{p}_{\mathrm{ij}}---\left(10\right)$

If μ _b(s, t), μ _caverage in the class of the target and background after (s, t) expression segmentation, then have:

${\mathrm{\&mu;}}_B(s,t)=\frac{1}{P_B}\underset{i=0}{\overset{s}{\mathrm{\&Sigma;}}}\underset{j=t+1}{\overset{L^{\&prime;}-1}{\mathrm{\&Sigma;}}}\mathrm{ij}{p}_{\mathrm{ij}}---\left(11\right)$

${\mathrm{\&mu;}}_C(s,t)=\frac{1}{P_C}\underset{i=s+1}{\overset{L-1}{\mathrm{\&Sigma;}}}\underset{j=t+1}{\overset{L^{\&prime;}-1}{\mathrm{\&Sigma;}}}{\mathrm{ijp}}_{\mathrm{ij}}---\left(12\right)$

The cross entropy formula obtaining image is:

$\begin{array}{c}D(B,C,s,t)=\underset{i=0}{\overset{s}{\mathrm{\&Sigma;}}}\underset{j=t+1}{\overset{L^{\&prime;}-1}{\mathrm{\&Sigma;}}}({\mathrm{ijp}}_{\mathrm{ij}}\ln \frac{\mathrm{ij}}{{\mathrm{\&mu;}}_B(s,t)}+{\mathrm{\&mu;}}_B(s,t)p_{\mathrm{ij}}\ln \frac{{\mathrm{\&mu;}}_B(s,t)}{\mathrm{ij}})+\\ \underset{i=s+1}{\overset{L-1}{\mathrm{\&Sigma;}}}\underset{j=t+1}{\overset{L^{\&prime;}-1}{\mathrm{\&Sigma;}}}({\mathrm{ijp}}_{\mathrm{ij}}\ln \frac{\mathrm{ij}}{{\mathrm{\&mu;}}_C(s,t)}+{\mathrm{\&mu;}}_C(s,t)p_{\mathrm{ij}}\ln \frac{{\mathrm{\&mu;}}_C(s,t)}{\mathrm{ij}})\end{array}---\left(13\right)$

For the quantity of information obtaining target and background and edge in image is maximum, optimal threshold vector (s ^*, t ^*) should meet:

$(s^{*},t^{*})=\arg \underset{0<t<L^{\&prime;}-1}{\underset{0<s<L-1}{\max }}\{1/D(B,C,s,t)\}---\left(14\right)$

2.4), utilize genetic algorithm and minimum cross entropy formula to choose the optimal threshold of image, namely meet the segmentation threshold that cross entropy in image is minimum;

The key step of genetic algorithm selected threshold is as follows:

A) set population number POPSIZE and iterations MAXGENS, binary coding is carried out to bivector, and random generation initialization population, population number gets 30-50, and iterations is got 10-30 time;

B) according to the actual requirements, determine the variation range of variable in fitness function, initialized population is decoded, and according to the fitness of formulae discovery every bar gene string;

C) adopt wheel robin unconditionally to copy in new population of future generation by gene string large for fitness, then parent population selected, the genetic operator computing of mating and variation, thus obtain the gene string of new population of future generation; Mating and variation are the genetic operators producing new population, and mating rate is got (0.25,0.75), and aberration rate is got (0.01,0.2);

D) multiply according to the iterations of setting, reach end condition and then return best gene string, and it can be used as two-dimentional segmentation threshold; Otherwise return and c) continue follow-on procreation;

3) morphological operation: first binary image is expanded and carry out corrosion treatment again;

4) connected component analysis;

5) visual inspection identification: in binary image from top left corner pixel point, if current pixel point is 1, then adopting greedy algorithm to search for its 8 field pixel value is the pixel of 1; Until search is the pixel of 1 less than pixel or has traveled through all picture positions; While search, set up object chain, the positional information of each target and relevant information all will be kept in object chain, after target identifying terminates, namely be the target information detected in object chain.

2. a kind of based on improving the moving target detecting method of minimum cross entropy as claimed in claim 1, it is characterized in that: described step 1) in, with the background subtraction based on fuzzy reasoning.

3. a kind of based on improving the moving target detecting method of minimum cross entropy as claimed in claim 1 or 2, it is characterized in that: described binarization also comprises: 2.5), for the segmentation threshold chosen, utilize homogeneity to estimate and evaluate.

4. a kind of based on improving the moving target detecting method of minimum cross entropy as claimed in claim 1 or 2, it is characterized in that: described step 3) in, adopt the regional analysis based on region growth method and wire tag method.