CN101377850A - Method of multi-formwork image segmentation based on ant colony clustering - Google Patents

Abstract

The invention provides a multi-templet image separation method based on Ant Colony Clustering. The multi-templet separation method comprises the implementation steps as follows: step 1. image preprocessing; step 2. determining the attributes of each pixel with such optional templets as Laplacian templet, Canny templet, Sobel templet, Roberts templet and so on; step 3. calculating an initial clustering center and the functional value of an initial optimum value; step 4. determining a search point set and initializing pheromone concentration and relevant parameters according to the intial clustering center; step 5. arranging M ants at random positions, clustering each ant at the search point set and updating the overall optimum value; step 6. updating the pheromone concentration; step 7. repeating step 5 and step 6 until the predetermined algorithm is completed NCmax times; step 8. finishing the algorithm and outputting the optimal results. The multi-templet image separation method effectively improve the separation speed and the zone integrality, which is an efficient path for solving the problem of image separation.

Description

A kind of multi-template image partition method based on ant colony clustering

(1) technical field

A kind of multi-template image partition method based on ant colony clustering (Ant Colony Clustering) of the present invention belongs to the computer vision information processing and handles the field.

(2) background technology

Image segmentation is a major issue of image processing field, and it is the basis of many Flame Image Process problems.Be widely used in a lot of fields, comprised image co-registration, pattern-recognition, computer vision, aircraft navigation, virtual reality, industrial detection, traffic administration, digital photogrammetry, medical image analysis etc.But because the complicacy of picture background, influences such as the diversity of target signature and noise make image segmentation become the difficult point of image processing techniques.

As the most basic in a computer vision field of information processing problem, image segmentation has attracted the research worker of many different field, comprise fields such as artificial intelligence, Aeronautics and Astronautics, navigation, robot, it is a research focus of present image processing field.Traditional image Segmentation Technology mainly contains two kinds, and a kind of is the method (cluster just) that the zone generates, and one is based on the method for Image Edge-Detection.These two kinds of methods have all obtained good effect at different images respectively, thereby all are the methods of using at present more widely.But at different condition, requirement and application target, classic method has shown significant limitation again.Present traditional clustering method poor effect under target and the unconspicuous situation of background, and calculated amount is big; Edge detection operator also all can only just can reach good effect under the certain conditions separately, seem powerless in complexity with changeable environment is next, and have the discontinuous or inaccurate problem in border in border, simultaneously, be subjected to the sensitivity that the threshold value selection causes to noise.

Ant group optimization (Ant Colony Optimization) algorithm is the bionical optimized Algorithm of ant colony foraging behavior in a kind of simulation Bugdom of recent development, this algorithm has adopted the positive feedback self-catalysis mechanism that walks abreast, have stronger robustness, good Distributed Calculation mechanism, be easy to and advantages such as additive method combines, showing its excellent performance and huge development potentiality aspect many complicated optimum problem solving.

Ant colony optimization algorithm is the optimized Algorithm of being come by the evolution of ant foraging behavior, is to carry out the information transmission by a kind of material that is referred to as pheromones (Pheromone) between the ant individuality, thereby can cooperates mutually, finishes complicated task.Ant in motion process, it can stay a certain amount of pheromones on the path of process, the intensity of pheromones is relevant with path.And ant can the perception path in motion process on the existence and the intensity thereof of pheromones, and instruct the selection of oneself to the path with this, ant tends to move towards the higher direction of pheromones intensity.Therefore, the ant group's who is made up of a large amount of ants collective behavior just shows a kind of information positive feedback phenomenon: the ant of passing by on a certain path is many more, and then the late comer selects the probability in this path just big more.It between the ant individuality purpose that reaches search food by the interchange of this information.Ant group algorithm has adopted the positive feedback self-catalysis mechanism that walks abreast, this algorithm has stronger robustness, good Distributed Calculation mechanism, is easy to and advantages such as additive method combines, and is also showing excellent performance and huge development potentiality aspect other many complicated optimum problem solving.

Occurring in nature, this class social animal of picture ant, ability of single ant and intelligence are very simple, no matter but they nest, look for food, migrate, clean complex behaviors such as ant cave by what mutual coordination, the division of labor, cooperation finished that worker ant or queen all can not have enough abilities to command to finish.The food source of ant always random scatter can find as long as we just examine that around ant nest after after a while, ant can be found a shortest path from the ant nest to the food source.Scientist once studied ant group's foraging behavior by " doube bridge experiment ".Discovery is except finding the shortest path between nest and the food source, and the ant group has extremely strong adaptive faculty to environment.For example when original shortest path became infeasible owing to the appearance of a new barrier, the ant group energy found a new shortest path rapidly.Therefore, in actual life, we always can observe a large amount of ants and form the path that is close to straight line between nest and food source, rather than curve or circle wait other shape, shown in Fig. 1 (a).Ant colony can not only be finished complicated task, and the variation that can also conform, when barrier occurring suddenly on ant group moving line, each ant distribution is uniform at the beginning, canal path length whether not, ant is always earlier by select each paths with equiprobability, shown in Fig. 1 (b).Ant can stay pheromones on the path of its process in motion process, and the existence and the intensity thereof of this material of energy perception, and instructs own travel direction with this, and ant tends to the high direction of pheromone concentration and moves.Just leave over often than the quantity of information on the short path in equal time, then select also to increase, thereupon shown in Fig. 1 (c) than the ant of short path.Be not difficult to find out, because the ant cluster behavior that a large amount of ants are formed has shown a kind of information positive feedback phenomenon, be that the ant of passing by on a certain path is many more, then the late comer selects the probability in this path just big more, search for food by this information interchange mechanism exactly between the ant individuality, and finally advance along shortest path, shown in Fig. 1 (d).

How does the ant group finish these complex tasks? the bionicist passes through a large amount of observations, discovers, ant is when seeking food, can on the path of its process, discharge the distinctive pheromones of a kind of ant, make other ants in the certain limit can feel this material, and tend to move towards the high direction of this material intensity.Therefore, ant group's collective behavior shows as a kind of information positive feedback phenomenon: the ant number of process is many more on certain paths, the pheromones that stays on it is also just many more (certainly, passing meeting is in time evaporated gradually), ant selected the probability in this path also high more afterwards, thereby had more increased the intensity of pheromones on this path.As time goes on, whole ant group finally can converge on the shortest traverse path.

Ant group algorithm is to be used to solve traveling salesman problem at first, the simple pictute of traveling salesman problem is: a given n city, a travelling salesman is arranged from a certain city, visit each city and once and only return the former city of setting out after once, require to find out a touring path the shortest.

As a kind of emerging heuristic bionic intelligence optimized Algorithm, people have been penetrated into a plurality of applications by single originally traveling salesman problem field to the research of ant colony optimization algorithm at present, develop into solution multidimensional dynamic combined optimization problem by solving one dimension static optimization problem, be extended to research in the continuous domain scope gradually by the research in the discrete domain scope, and in the hardware realization of ant colony optimization algorithm, also obtained a lot of breakthroughs, thereby make this emerging bionical optimized Algorithm show vitality and vast potential for future development.

(3) summary of the invention

The present invention proposes a kind of multi-template image partition method based on ant colony clustering, its objective is provides a kind of effective way that solves image segmentation problem, also can be applicable to other complicated intelligent optimization problem.

This method is regarded image as the set of the pixel with different gradient features, the utilization ant group optimization carries out the cluster at edge to it, thereby realize cutting apart of image, edge extracting result according to different masterplates sets initial cluster center again, algorithm is improved, and image is carried out cluster analysis with the algorithm after improving, extract edge of image.

Ant group algorithm is to be used to solve traveling salesman problem (Traveling Salesman Problem at first, TSP), the simple pictute of traveling salesman problem is: a given n city, a travelling salesman is arranged from a certain city, visit each city and once and only return the former city of setting out after once, require to find out a touring path the shortest.

The mathematical model of basic ant group algorithm is as follows:

If b _i(t) expression t is positioned at the ant number of element i, τ constantly _Ij(t) be t constantly the path (n represents the TSP scale for i, the j) quantity of information on, i.e. city total number, m is the total number of ant among the ant group, then $m=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{b}_i\left(t\right);$ $\mathrm{\Γ}=\left\{{\mathrm{\τ}}_{\mathrm{ij}}\left(t\right)\right|c_i,c_j\⋐C\}$ Be that t gathers element among the C (city) constantly and connects l in twos _IjThe set of last residual risk amount.Quantity of information equates on each paths of initial time, and to establish the initial information amount be τ _Ij(0)=const.

Ant k (k=1,2 ... .. m) in motion process, determines its shift direction according to the quantity of information on each paths.Here with taboo table tabu _k(k=1,2 ...., m) write down the current city of passing by of ant k, set tabu _kAlong with evolutionary process is done dynamically to adjust.

In the search procedure, ant comes the computing mode transition probability according to the heuristic information in quantity of information on each paths and path.

Be illustrated in t moment ant k is transferred to element (city) j by element (city) i state transition probability

$p_{\mathrm{ij}}^k\left(t\right)=\left\{\begin{array}{cc}\frac{{\left[{\mathrm{\τ}}_{\mathrm{ij}}\left(t\right)\right]}^{\mathrm{\α}}{\left[{\mathrm{\η}}_{\mathrm{ij}}\right]}^{\mathrm{\β}}}{\underset{k\∈{\mathrm{allowed}}_k}{\mathrm{\Σ}}{\left[{\mathrm{\τ}}_{\mathrm{ik}}\left(t\right)\right]}^{\mathrm{\α}}{\left[{\mathrm{\η}}_{\mathrm{ik}}\right]}^{\mathrm{\β}}}& \mathrm{if\; j}\∈{\mathrm{allowed}}_k\\ 0& \mathrm{otherwise}\end{array}\right.---\left(1\right)$

In the formula, allowed _k={ C-tabu _kNext step allows the city of selection to represent ant k.

α is the heuristic factor of information, the relative importance of expression track has reflected information role when ant moves that ant is accumulated in motion process, its value is big more, then this ant tends to select the path of other ant process more, and collaborative is strong more between the ant;

β is the heuristic factor of expectation, and the relative importance of expression visibility has reflected that ant heuristic information in motion process selects the attention degree that is subjected in the path ant, and its value is big more, and then to approach greed more regular for this state transition probability.

η _Ij(t) be heuristic function, its expression formula is as follows

${\mathrm{\η}}_{\mathrm{ij}}\left(t\right)=\frac{1}{d_{\mathrm{ij}}}---\left(2\right)$

In the formula, d _IjRepresent the distance between adjacent two cities.For ant k, d _IjMore little, η then _IjIt is (t) big more, Also just big more.Obviously, this heuristic function represents that ant transfers to the expected degree of element (city) j from element (city) i.

Cause too much that for fear of the residual risk element residual risk floods heuristic information, after every ant is covered a traversal (also i.e. loop ends) that goes on foot or finish all n city, upgrade processing residual risk.This update strategy has imitated the characteristics of human brain memory, when fresh information constantly deposits brain in, is stored in the As time goes on desalination gradually of old information in the brain, even forgets.

Thus, (i, j) quantity of information on can be adjusted according to the following rules to be engraved in the path during t+n

τ _ij(t+n)＝(1-ρ)·τ _ij(t)+Δτ _ij(t) (3)

$\mathrm{\Δ}{\mathrm{\τ}}_{\mathrm{ij}}\left(t\right)=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}\mathrm{\Δ}{\mathrm{\τ}}_{\mathrm{ij}}^k\left(t\right)---\left(4\right)$

In the formula, ρ represents the pheromones volatility coefficient, and then 1-ρ represents the residual factor of pheromones, and in order to prevent the unlimited accumulation of information, the span of ρ is: $\mathrm{\ρ}\⋐\left[\mathrm{0,1}\right);$

Δ τ _Ij(t) represent path (i, j) the pheromones increment on, initial time in this circulation $\mathrm{\Δ}{\mathrm{\τ}}_{\mathrm{ij}}^k\left(0\right)=0,$

Represent k ant this circulation in stay path (i, j) quantity of information on.

In the Ant-Cycle model:

(5)

In the formula, Q represents pheromones intensity, and it influences convergence of algorithm speed to a certain extent;

L _kThe total length of representing k ant path of passing by in this circulation.

A kind of multi-template image partition method of the present invention based on ant colony clustering, its specific implementation step following (can referring to Fig. 2):

Step 1: image pre-service

Exist each species diversity owing to carry out picture format, size and the feature of image segmentation, many times all be not suitable for image is directly cut apart.Therefore, before image segmentation, image is carried out pre-service such as sharpening, can make its feature more outstanding, make that final extraction effect is better.

These pre-treatment step mainly comprise:

(1) image reads and is converted to gray level image

All being converted to earlier gray level image for the image of different-format carries out the edge again and cuts apart.Make disposal route unified like this, be convenient to carry out programmed.

(2) adjustment of image size

For the image that collects under the different situations, its size specification has nothing in common with each other, and existing it is adjusted to a suitable size carry out subsequent treatment again, saves time and improves effect.

(3) sharpening of medium filtering denoising and image

Image denoising sound, The noise in the time of can alleviating image segmentation, thus can improve segmentation precision and accuracy greatly; Image sharpening can make the edge of image feature more outstanding, thereby make final effect better.

Through after the pre-service, picture size is suitable, and noise is less, and the edge feature distinctness is suitable for follow-up image segmentation work.

Step 2: the attribute of determining each pixel

Owing to be to extract edge of image, so with the attribute of the individual pixel of Grad conduct of the calculating gained of multiple masterplate (as Laplacian masterplate, Canny masterplate etc.).Template is selected to select to be best suited for two kinds of templates of this image according to the different image and the different templates scope of application.

After choosing two kinds of templates (template A, template B), the template A gradient of each pixel of computed image and template B gradient are as two attributes of each point, the foundation of ant group hunting.

According to two kinds of gradients, calculate marginal point set separately respectively.Extracting the edge of image point set by a kind of masterplate wherein is a, and extracting the edge of image point set by another kind of masterplate is b.

Selectable template has: Laplacian masterplate, Canny masterplate, Sobel masterplate, Roberts masterplate etc.

Step 3: calculate initial cluster center and initial optimization degree functional value F ₀

As the initial edge point set, all belong to two attributes mean value separately of pixel of a ∩ b as initial cluster center (Cen with the common factor a ∩ b of a and b _a, Cen _b).All two property values that belong to the pixel of a ∩ b constitute two arrays, are designated as ab ₁₁, ab ₁₂

F ₀＝var(ab ₁)+var(ab ₂) (6—1)

Wherein, var is the function that calculates variance.

In the iteration each time afterwards, all can obtain an edge point set, concentrate two attributes of pixel to constitute two arrays this marginal point, be designated as ab _I1, ab _I2(i is an iterations)

F＝var(ab _i1)+var(ab _i2) (6—2)

Optimization degree function F is to characterize the function that edge extracting is optimized degree, and F is more little, shows the property variance of edge point set and more little, and the edge extracting result is good more.

Step 4: determine search point set and plain concentration of initialization information and correlation parameter according to initial cluster center

The each point that will belong to Laplacian or Canny border directly keeps, and asks the distance of remaining each point to initial cluster center.Suppose that certain some attribute is for (x, y), it to distances of clustering centers is so

$d=\sqrt{({|x-{\mathrm{Cen}}_a|}^2+{|y-{\mathrm{Cen}}_b|}^2)}---\left(7\right)$

Generally speaking, keep central proper from the 12% nearest left and right sides of cluster centre, the possibility that other point is a frontier point very little (almost nil), just can directly cast out, no longer consider them in the search afterwards, can reduce a lot of calculated amount like this, thereby improve cluster seeking efficient greatly.

The point (A_and_B) that will belong to a ∩ b gives initial information plain concentration 0.95, the point (A_or_B) that belongs to a or b but do not belong to a ∩ b gives initial information plain concentration 0.5, in 12% the point that keeps, give initial information plain concentration 0.3 from the nearest first half of cluster centre (Center-front), latter half (Center-back) gives initial information plain concentration 0.2, thereby obtains pheromone concentration matrix τ.

Correlation parameter initialization: N is set _Cmax, M, ρ, ζ, wherein: Nc_max is this algorithm maximum cycle; M is the ant total number; ρ is the pheromones residual coefficients, i.e. pheromones is residual after each generation; ζ is a coefficient, determines its value size according to the actual conditions of split image.

Step 5: M ant is placed at random position, and every ant carries out cluster and upgrades global optimum the search point set respectively

According to pheromone concentration, select formula according to probability

Wherein, rand () is a random number; Determine whether certain point is grouped in the point set of border, M ant can obtain M kind result like this.The quantity M of ant is set by experience by the size of actual needs and image.

Calculate the edge optimization degree F of M kind result in the iteration each time respectively.Get the F of that result of F minimum as this generation optimum _DbestCompare F _DbestWith the F of global optimum _GbestIf, F _DbestLess than F _Gbest, so with this generation optimal result be updated to global optimum, this generation F _DbestGive F _Gbest

All can produce this generation optimal edge pixel point set L each time in the iteration _Best, finish N _{Cmax time}After the iteration, produce the L of global optimum _Gbest

Step 6: the plain concentration of lastest imformation

After iteration finishes each time, carry out pheromones and upgrade, its update rule is as follows

τ(t+1)＝ρ·τ(t)+Δτ(t) (9)

Wherein, ρ is the pheromones residual coefficients, i.e. pheromones is residual after each generation.Δ τ is the pheromone concentration Increment Matrix, and its value is calculated with following formula.

$\mathrm{\Δ\τ}=\frac{\mathrm{\ζ}\·F_0}{F_{\mathrm{dbest}}}\×L_{\mathrm{Best}}---\left(10\right)$

Wherein, F _DbestEdge optimization degree for this generation optimum;

L _BestBe this generation optimal edge pixel point set;

ζ is a coefficient, determines its value size according to the actual conditions of split image.

Step 7: step 5 that repeats to return and step 6, up to finishing predetermined algorithm cycle index N _Cmax

Step 8: algorithm finishes, and the output optimal result.The edge pixel point set L of global optimum _Gbest

Wherein, the two kinds of template A and the B that are used for Edge extraction select the Laplacian masterplate, Canny masterplate, Sobel masterplate, any two kinds combination in the Roberts masterplate.

Wherein, two kinds of template A and the B that is used for Edge extraction selects Laplacian masterplate and Canny masterplate.

A kind of multi-template image partition method of the present invention based on ant colony clustering, its advantage and effect are: this method is for solving the effective way of image segmentation problem in the Vision information processing field, application is extensive, relates to the field of Image Information Processing as Aeronautics and Astronautics, robot, commercial production etc.

(4) description of drawings

The ant group seeks the process of food in Fig. 1 reality

Fig. 2 is based on the multi-template image segmentation flow process of ant colony clustering

Fig. 3 ant colony clustering evolution curve

Number in the figure and symbol description are as follows:

F ₀---initial optimization degree functional value

M---the ant number that is calculating

M---ant total number

F _Dbest---this generation optimum solution

F _Gbest---globally optimal solution

Nc---algorithm cycle index

Nc_max---algorithm maximum cycle

Y---(being) satisfies condition

N---(denying) do not satisfy condition

(5) embodiment

Verify the performance of the present invention's put forward based on the multi-template image partition method of ant colony clustering below by a specific embodiment, what adopted is that the image of jpg form of a width of cloth 570*447 is as identifying object.Experimental situation is P43.06Ghz, the 1G internal memory, and the MATLAB7.1 version, its specific implementation step is as follows:

Step 1: image pre-service

(1) image reads and is converted to gray level image

Identifying object is converted to gray level image carries out the edge again and cut apart, make disposal route unified, be convenient to carry out programmed.

(2) adjustment of image size

It is adjusted to a suitable size carry out subsequent treatment again, save time and improve effect.

(3) sharpening of medium filtering denoising and image

Image denoising sound, The noise when alleviating image segmentation; Image sharpening makes the edge of image feature more outstanding.

Through after the pre-service, picture size is suitable, and noise is less, and the edge feature distinctness is suitable for follow-up image segmentation work.

Step 1: parameter initialization: N _Cmax=100, M=8, ρ=0.9, ζ=0.15, the plain concentration of A_and_B initial information gets 0.95, and the plain concentration of A and B initial information all gets 0.5, the plain concentration of Center-front initial information gets 0.3, and the plain concentration of Center-back initial information gets 0.2, masterplate be chosen as Canny and Laplacian.

Step 2: the attribute of determining each pixel

Owing to be to extract edge of image, so with the attribute of the individual pixel of Grad conduct of the calculating gained of various masterplates.The Laplacian gradient of each pixel of computed image and Canny gradient are as two attributes of each point, the foundation of ant group hunting.According to two kinds of gradients, calculate marginal point set separately respectively.Extracting the edge of image point set by the Laplacian masterplate is a, and extracting the edge of image point set by the Canny masterplate is b.

Step 3: calculate initial cluster center and initial optimization degree functional value F ₀

As the initial edge point set, all belong to two attributes mean value separately of pixel of a ∩ b as initial cluster center (0.3,0.2) with the common factor a ∩ b of a and b.

F＝var(ab ₁)+var(ab ₂)

Step 4: determine search point set and plain concentration of initialization information and correlation parameter according to initial cluster center

The each point that will belong to Laplacian or Canny border directly keeps, and asks the distance of remaining each point to initial cluster center.Suppose that certain some attribute is for (x, y), it to distances of clustering centers is so

$d=\sqrt{({|x-0.3|}^2+{|y-0.2|}^2)}$

Keep central from cluster centre nearest 12%, the point (A_and_B) that will belong to a ∩ b gives initial information plain concentration 0.95, belongs to a or b but the point (A_or_B) that do not belong to a ∩ b gives initial information plain concentration 0.5.

Correlation parameter initialization: N _Cmax=100, M=8, ρ=0.9, ζ=0.15, the plain concentration of A_and_B initial information gets 0.95, and the plain concentration of A and B initial information all gets 0.5, the plain concentration of Center-front initial information gets 0.3, and the plain concentration of Center-back initial information gets 0.2, masterplate be chosen as Canny and Laplacian.

Step 5: every ant carries out cluster and upgrades global optimum the search point set respectively

According to pheromone concentration, select formula according to probability

(3)

Determine whether certain point is grouped in the point set of border, such 8 ants can obtain 8 kinds of results.

Calculate the edge optimization degree F of 8 kinds of results in the iteration each time respectively.Get the F of that result of F minimum as this generation optimum _DbestCompare F _DbestWith the F of global optimum _GbestIf, F _DbestLess than F _Gbest, so with this generation optimal result be updated to global optimum, this generation F _DbestGive F _Gbest

All can produce this generation optimal edge pixel point set L each time in the iteration _Best, finish 100 iteration after, produce the L of global optimum _Gbest

Step 6: the plain concentration of lastest imformation

After iteration finishes each time, carry out pheromones and upgrade, its update rule is as follows

τ(t+1)＝0.9·τ(t)+Δτ(t)

$\mathrm{\Δ\τ}=\frac{0.15\·F_0}{F_{\mathrm{dbest}}}\×L_{\mathrm{Best}}$

Step 7: step 5 that repeats to return and step 6, up to finishing predetermined algorithm cycle index N _Cmax=100.

Step 8: algorithm finishes, and the output optimal result.The edge pixel point set L of global optimum _Gbest

By the result of experiment operation as seen, the effect on the border that the Canny masterplate extracts is fine, but the details of image is too much, and a lot of lines are arranged, and does not have but well to reflect that the lines that each relatively independent part only is the border pile up.The image that the Laplacian masterplate extracts has same problem, and well rim detection is not come out for some darker zones; And use this paper based on the method for ant group optimization to image segmentation, by experimental result as seen, more secretly can detect with fine, and can not seem that details is too much, whole split image stereovision is more intense, lines are the various piece in the composing images each other, and are not only that lines are piled up, and segmentation result is more accurate.Experiment shows that dividing method proposed by the invention has improved splitting speed and regional integrality effectively, shows the segmentation effect that this method can obtain.The given evolutionary process of Fig. 3 is tending towards an optimal value comparatively reposefully, reaches the stable state convergence at last.

Claims (3)

1, a kind of multi-template image partition method based on ant colony clustering, it is characterized in that: the performing step of this method is as follows:

Step 1: the image pre-service mainly comprises:

(1) image reads and is converted to gray level image

All being converted to earlier gray level image for the image of different-format carries out the edge again and cuts apart;

(2) adjustment of image size

For the image that collects under the different situations, its size specification has nothing in common with each other, and now it is adjusted to a suitable size and carries out subsequent treatment again;

(3) sharpening of medium filtering denoising and image

Image denoising sound, The noise when alleviating image segmentation; Image sharpening makes the edge of image feature more outstanding;

Step 2: the attribute of determining each pixel

According to the different image and the different templates scope of application, select to be best suited for two kinds of templates of this Edge extraction, calculate the attribute of the Grad of gained as pixel;

After choosing two kinds of templates (template A, template B), the template A gradient of each pixel of computed image and template B gradient are as two attributes of each point, the foundation of ant group hunting;

According to two kinds of gradients, calculate marginal point set separately respectively; Extracting the edge of image point set by a kind of masterplate wherein is a, and extracting the edge of image point set by another kind of masterplate is b;

Step 3: calculate initial cluster center and initial optimization degree functional value F ₀

As the initial edge point set, all belong to two attributes mean value separately of pixel of a ∩ b as initial cluster center (Cen with the common factor a ∩ b of a and b _a, Cen _b); All two property values that belong to the pixel of a ∩ b constitute two arrays, are designated as ab _L1, ab _L2

F ₀＝var(ab ₁)+var(ab ₂) (6—1)

Wherein, var is the function that calculates variance;

In the iteration each time afterwards, all can obtain an edge point set, concentrate two attributes of pixel to constitute two arrays this marginal point, be designated as ab _I1, ab _I2(i is an iterations)

F＝var(ab _i1)+var(ab _i2) (6—2)

Optimization degree function F is to characterize the function that edge extracting is optimized degree, and F is more little, shows the property variance of edge point set and more little, and the edge extracting result is good more;

Step 4: determine search point set and plain concentration of initialization information and correlation parameter according to initial cluster center

The each point that will belong to Laplacian or Canny border directly keeps, and asks the distance of remaining each point to initial cluster center; Suppose that certain some attribute is for (x, y), it to distances of clustering centers is so

$d=\sqrt{({|x-{\mathrm{Cen}}_a|}^2+{|y-{\mathrm{Cen}}_b|}^2)}---\left(7\right)$

Keep central properly from the 12% nearest left and right sides of cluster centre, other point is that frontier point is directly cast out, and no longer considers them in the search afterwards;

The point (A_and_B) that will belong to a ∩ b gives initial information plain concentration 0.95, the point (A_or_B) that belongs to a or b but do not belong to a ∩ b gives initial information plain concentration 0.5, in 12% the point that keeps, give initial information plain concentration 0.3 from the nearest first half of cluster centre (Center-front), latter half (Center-back) gives initial information plain concentration 0.2, thereby obtains pheromone concentration matrix τ;

Correlation parameter initialization: N is set _Cmax, M, ρ, ζ, wherein: Nc_max is this algorithm maximum cycle; M is the ant total number; ρ is the pheromones residual coefficients, i.e. pheromones is residual after each generation; ζ is a coefficient, determines its value size according to the actual conditions of split image;

Step 5: M ant is placed at random position, and every ant carries out cluster and upgrades global optimum the search point set respectively

According to pheromone concentration, select formula according to probability

Wherein, rand () is a random number; Determine whether certain point is grouped in the point set of border, M ant can obtain M kind result like this; The quantity M of ant is set by experience by the size of actual needs and image;

Calculate the edge optimization degree F of M kind result in the iteration each time respectively; Get the F of that result of F minimum as this generation optimum _DbestCompare F _DbestWith the F of global optimum _GbestIf, F _DbestLess than F _Gbest, so with this generation optimal result be updated to global optimum, this generation F _DbestGive F _Gbest

All can produce this generation optimal edge pixel point set L each time in the iteration _Best, finish N _{Cmax time}After the iteration, produce the L of global optimum _Gbest

Step 6: the plain concentration of lastest imformation

After iteration finishes each time, carry out pheromones and upgrade, its update rule is as follows

τ(t+1)＝ρ·τ(t)+Δτ(t) (9)

Wherein, ρ is the pheromones residual coefficients, i.e. pheromones is residual after each generation; Δ τ is the pheromone concentration Increment Matrix, and its value is calculated with following formula;

$\mathrm{\Δ\τ}=\frac{\mathrm{\ζ}\·F_0}{F_{\mathrm{dbest}}}\×L_{\mathrm{Best}}---\left(10\right)$

Wherein, F _DbestEdge optimization degree for this generation optimum;

L _BestBe this generation optimal edge pixel point set;

ζ is a coefficient, determines its value size according to the actual conditions of split image;

Step 7: step 5 that repeats to return and step 6, up to finishing predetermined algorithm cycle index N _Cmax

Step 8: algorithm finishes, and output optimal result, i.e. global optimum's edge pixel point set L _Gbest

2, a kind of multi-template image partition method as claimed in claim 1 based on ant colony clustering, it is characterized in that: the two kinds of template A and the B that are used for Edge extraction select the Laplacian masterplate, the Canny masterplate, Sobel masterplate, any two kinds combination in the Roberts masterplate.

3, a kind of multi-template image partition method based on ant colony clustering as claimed in claim 2 is characterized in that: the two kinds of template A and the B that are used for Edge extraction select Laplacian masterplate and Canny masterplate.

Images