CN106097351A - A kind of based on multiobject adaptive threshold image partition method - Google Patents

Abstract

The invention discloses a kind of based on multiobject adaptive threshold image partition method, comprise the following steps: 1. input image to be split, if coloured image, gray level image need to be converted to；2. the parameter of multi-target evolution self adaptation mesh threshold Image Segmentation is set, max-thresholds is split number and is set to 5；3. the view data of pair input carries out the multiple target threshold Image Segmentation of 1～5 threshold values respectively；4. the Parato optimal solution that will be obtained under 1～5 threshold values by step 2, utilizes F function to try to achieve the optimal solution under 1～5 threshold values respectively, 5. by comparing the optimal segmenting threshold as image of the difference most suitable solution of selection between F functional value；6. the optimal threshold by choosing carries out category division to original image and obtains final segmentation result.The present invention is capable of the segmentation of adaptive threshold image, and segmentation result is accurate, and algorithm realizes simple.

Description

A kind of based on multiobject adaptive threshold image partition method

Technical field

The invention belongs to image processing field, be specifically related to a kind of based on multiobject adaptive threshold image segmentation side Method.

Background technology

Image segmentation be that in image procossing, the bottom is also a most important step, its objective is according to the gray scale in image, Color and Texture eigenvalue divide an image into several mutually disjoint regions, and make the same area have similar spy Levy, there is between zones of different obvious feature difference.

Image partition method based on threshold value is usually, according to some threshold value criterion, image is carried out Threshold segmentation, because of This image obtained is optimum or close to optimum under this criterion.But in actual applications, image segmentation is one According to people's actual demand and actual application environment, carrying out splitting according to concrete problem, the most single threshold value criterion has Concrete demand may be can not meet, it would be desirable to go to consider a problem from multiple angles.Additionally, in the selection of threshold number Aspect the most first determines threshold number and splits, and judgment threshold that thus must be artificial splits number.

Summary of the invention

It is an object of the invention to provide a kind of based on multiobject adaptive threshold image partition method, above-mentioned to overcome The defect that prior art exists, the present invention is capable of the segmentation of adaptive threshold image, and segmentation result is accurate, and algorithm realizes letter Single.

For reaching above-mentioned purpose, the present invention adopts the following technical scheme that

A kind of based on multiobject adaptive threshold image partition method, comprise the following steps:

Step one: input image to be split, if image to be split is coloured image, is first converted into gray level image；

Step 2: carry out the multi-target evolution of some different threshold number for the gray level image obtained by step one respectively Threshold segmentation, obtains some different Parato disaggregation；

Step 3: use function F_kObtain each Parato and solve the optimal threshold concentrated；

Step 4: by the F under relatively different threshold number_kDifference identification Optimal-threshold segmentation number；

Step 5: by Optimal-threshold segmentation number, original image is carried out category division and obtain final segmentation result.

Further, in step 2 multi-target evolution threshold segmentation method particularly as follows:

2a) arranging the parameter of multi-target evolution Threshold segmentation: max-thresholds number is 5, population scale is 200, maximum something lost Passage number is 300, and crossover probability is 0.9, and mutation probability is 0.1, and genovariation scope is 5～15, gene code scope 0～ 255；

2b) initialization of population, randomly generates 200 individualities, it is assumed that have k the most each chromosome of threshold value to have k gene position, Each gene position value is 0～255, genetic algebra g=1；

2c) calculate two target function values of each individuality in population, two target function values are respectively added to dyeing K+1, k+2 gene position of body；

2d) utilizing step 2c) population carries out the middle target function value calculated non-dominated ranking and crowding distance calculates, And individual sequence value and crowding distance are respectively added to k+3, k+4 gene position of chromosome；

2e) start to evolve, use tournament method to select half quantity from population according to sequence value and crowding distance size Individual as parent population；

2f) parent population is intersected and mutation operation, produce progeny population；

2g) current population merged with progeny population and carry out elitist selection, it is thus achieved that identical with initial population size is new Generation population；

If 2h) g > 300, then perform step 3, otherwise, g=g+1, jump to step 2c).

Further, step 2c) in two object functions be respectively embody inter-class variance function f₁And maximum possible Retain image original information and the entropy function f of edge contour information₂。

Further, calculate two target function values of each individuality in population particularly as follows:

Assume that the number of pixels comprised in piece image is N, the grey level range in image be [0 ..., L], gray level is The number of pixels of i is n_i, then the probability that gray level i occurs is:

$P_i=\frac{n_i}{N}$

Then object function f₁For:

f₁(t₁,t₂... t_k)=w₁(u₁-u_T)²+w₂(u₂-u_T)²+…+w_k+1(u_k+1-u_T)²

Wherein:

$w_1=\underset{i=0}{\overset{t_1-1}{\Σ}}{p}_i,...,w_n=\underset{i=t_{n-1}}{\overset{t_n-t_{n-1}}{\Σ}}{p}_i,...,w_{k+1}=\underset{i=t_k}{\overset{L}{\Σ}}{p}_i;$

$u_1=\underset{i=0}{\overset{t_1-1}{\Σ}}\frac{i*p_i}{w_1},...,u_n=\underset{i=t_{n-1}}{\overset{t_n-t_{n-1}}{\Σ}}\frac{i*p_i}{w_n},...,u_{k+1}=\underset{i=t_k}{\overset{L}{\Σ}}\frac{i*p_i}{w_{k+1}};$

$u_T=\underset{i=0}{\overset{L}{\Σ}}i*p_i$

Wherein, (t₁,t₂... t_k) it is the threshold value of image, the threshold number that k is split by image, w_n(1≤n≤k+1) is The gray scale probability of occurrence of the n-th class pixel and, u_n(1≤n≤k+1) is the average gray of the n-th class pixel, u_TIt it is entire image Average gray value；

Object function f₂For:

f₂(t₁,t₂... t_k)=H₁+H₂+…+H_k+1

Wherein:

$H_1=-\underset{i=0}{\overset{t_1-1}{\Σ}}\frac{p_i}{w_1}ln\frac{p_i}{w_1},w_1=\underset{i=0}{\overset{t_1-1}{\Σ}}{p}_i$

$H_n=-\underset{i=t_{n-1}}{\overset{t_n-t_{n-1}}{\Σ}}\frac{p_i}{w_n}ln\frac{p_i}{w_n},w_n=\underset{i=t_{n-1}}{\overset{t_n-t_{n-1}}{\Σ}}{p}_i$

$H_{k+1}=-\underset{i=t_k}{\overset{L}{\Σ}}\frac{p_i}{w_{k+1}}ln\frac{p_i}{w_{k+1}},w_{k+1}=\underset{i=t_k}{\overset{L}{\Σ}}{p}_i$

In formula, (t₁,t₂... t_k) it is the threshold value of image, the threshold number that k is split by image, H_n(1≤n≤k+1) is The comentropy sum of the n-th class pixel, w_n(1≤n≤k+1) be the n-th class pixel gray scale probability of occurrence and, p_iOccur for pixel i Probability.

Further, function F in step 3_kParticularly as follows:

$F_k=\frac{N*\left(m\right(k+1)-m\left(1\right))}{k*\underset{j=1}{\overset{k+1}{\Σ}}\underset{i\∈C_j}{\Σ}{(g_i-m_j)}^2}$

Wherein, N is the number of pixels of entire image, the threshold number that k is split by image, and m (k+1) is kth+1 class Average gray, m (1) is the average gray of the first kind, m_jFor the average gray of jth class, C_jIt is the collection of pixels of jth class, g_iIt it is the gray value of pixel i.

Further, by the F under relatively different threshold number in step 4_kDifference identification Optimal-threshold segmentation number, Particularly as follows:

4a) calculate the F under present threshold value number_Δ=F_k-F_k-1If this value is negative value, is set to 0；Under first threshold F_Δ=F₁；

4b) compare the F under different threshold number_ΔValue, selects F_ΔThreshold number when value is maximum is as Optimal-threshold segmentation Number.

Compared with prior art, the present invention has a following useful technique effect:

The present invention is split first with multi-target evolution threshold segmentation method.With single Threshold segmentation criterion algorithm Comparing, multi-target evolution Threshold Segmentation Algorithm has the most bimodal no longer sensitivity to image background and target are the brightest, and to the greatest extent may be used The original information retaining image that can be many and edge contour information, the result of Threshold segmentation more preferably meets the demand of reality.Finally Employ F_kFunction can select suitable threshold value from last group Pareto optimal solution, moreover it is possible to from different threshold number Obtain optimal threshold value, and obtain ideal segmentation result.

Accompanying drawing explanation

Fig. 1 is the overall flow figure of the inventive method；

Fig. 2 is the shot1 image segmentation result comparison diagram of the inventive method；

Fig. 3 is the Berkeley image segmentation result comparison diagram used in emulation experiment of the present invention；

Fig. 4 is the Berkeley image segmentation result comparison diagram used in l-G simulation test of the present invention；

Wherein, (a) is image to be split；B () is Otsu method segmentation result under given threshold number；C () is given threshold Maximum entropy method (MEM) segmentation result under value number；(d) be the present invention under not giving threshold number, adaptive segmentation result.

Detailed description of the invention

Below in conjunction with the accompanying drawings the present invention is described in further detail:

A kind of based on multiobject adaptive threshold image partition method, comprise the following steps:

Step one: input image to be split, if image to be split is coloured image, is first converted into gray level image；

Step 2: carry out the multi-target evolution of some different threshold number for the gray level image obtained by step one respectively Threshold segmentation, obtains some different Parato disaggregation；Concretely comprise the following steps:

2a) arranging the parameter of multi-target evolution Threshold segmentation: max-thresholds number is 5, population scale is 200, maximum something lost Passage number is 300, and crossover probability is 0.9, and mutation probability is 0.1, and genovariation scope is 5～15, gene code scope 0～ 255；

2b) initialization of population, randomly generates 200 individualities, it is assumed that have k the most each chromosome of threshold value to have k gene position, Each gene position value is 0～255, genetic algebra g=1；

2c) calculate two target function values of each individuality in population, two target function values are respectively added to dyeing K+1, k+2 gene position of body；

Two object functions are respectively and embody inter-class variance function f₁And maximum possible retain image original information and The entropy function f of edge contour information₂, for calculating target function, it is assumed that the number of pixels comprised in piece image is N, image In grey level range be [0 ..., L], gray level be the number of pixels of i be n_i, then the probability that gray level i occurs is:

$p_i=\frac{n_i}{N}$

Then object function f₁For:

f₁(t₁,t₂... t_k)=w₁(u₁-u_T)²+w₂(u₂-u_T)²+…+w_k+1(u_k+1-u_T)²

Wherein:

$w_1=\underset{i=0}{\overset{t_1-1}{\Σ}}{p}_i,...,w_n=\underset{i=t_{n-1}}{\overset{t_n-t_{n-1}}{\Σ}}{p}_i,...,w_{k+1}=\underset{i=t_k}{\overset{L}{\Σ}}{p}_i;$

$u_1=\underset{i=0}{\overset{t_1-1}{\Σ}}\frac{i*p_i}{w_1},...,u_n=\underset{i=t_{n-1}}{\overset{t_n-t_{n-1}}{\Σ}}\frac{i*p_i}{w_n},...,u_{k+1}=\underset{i=t_k}{\overset{L}{\Σ}}\frac{i*p_i}{w_{k+1}};$

$u_T=\underset{i=0}{\overset{L}{\Σ}}i*p_i$

(t in above formula₁,t₂... t_k) it is the threshold value of image, the threshold number that k is split by image, w_n(1≤n≤k+1) Be the n-th class pixel gray scale probability of occurrence and, u_n(1≤n≤k+1) is the average gray of the n-th class pixel, u_TIt it is entire image Average gray value.

Object function f₂For:

f₂(t₁,t₂... t_k)=H₁+H₂+…+H_k+1

Wherein:

$H_1=-\underset{i=0}{\overset{t_1-1}{\Σ}}\frac{p_i}{w_1}ln\frac{p_i}{w_1},w_1=\underset{i=0}{\overset{t_1-1}{\Σ}}{p}_i$

$H_n=-\underset{i=t_{n-1}}{\overset{t_n-t_{n-1}}{\Σ}}\frac{p_i}{w_n}ln\frac{p_i}{w_n},w_n=\underset{i=t_{n-1}}{\overset{t_n-t_{n-1}}{\Σ}}{p}_i$

$H_{k+1}=-\underset{i=t_k}{\overset{L}{\Σ}}\frac{p_i}{w_{k+1}}ln\frac{p_i}{w_{k+1}},w_{k+1}=\underset{i=t_k}{\overset{L}{\Σ}}{p}_i$

(t in above formula₁,t₂... t_k) it is the threshold value of image, the threshold number that k is split by image, H_n(1≤n≤k+1) It is the comentropy sum of the n-th class pixel, w_n(1≤n≤k+1) be the n-th class pixel gray scale probability of occurrence and, p_iOccur for pixel i Probability.

In order to obtain optimal threshold, function f need to be maximized simultaneously₁And f₂.Due in multi-objective Evolutionary Algorithm, the most excellent Change functional minimum value and be easy to the calculating of algorithm, 1/f will be minimized the most in the method simultaneously₁、1/f₂Two object functions come Ask for optimal threshold.

2d) utilizing step 2c) population carries out the middle target function value calculated non-dominated ranking and crowding distance calculates, And individual sequence value and crowding distance are respectively added to k+3, k+4 gene position of chromosome；

2e) start to evolve, use tournament method to select half quantity from population according to sequence value and crowding distance size Individual as parent population；

2f) parent population is intersected and mutation operation, produce progeny population；

2g) current population merged with progeny population and carry out elitist selection, it is thus achieved that identical with initial population size is new Generation population；

If 2h) g > 300, then perform step 3, otherwise, g=g+1, jump to step 2c).

Step 3: use function F_kObtain each Parato and solve the optimal threshold concentrated, function F_kIt is worth the biggest, then splits knot Fruit is the best；Particularly as follows:

$F_k=\frac{N*\left(m\right(k+1)-m\left(1\right))}{k*\underset{j=1}{\overset{k+1}{\Σ}}\underset{i\∈C_j}{\Σ}{(g_i-m_j)}^2}$

Wherein, N is the number of pixels of entire image, the threshold number that k is split by image, and m (k+1) is kth+1 class Average gray, m (1) is the average gray of the first kind, m_jFor the average gray of jth class, C_jIt is the collection of pixels of jth class, g_iIt it is the gray value of pixel i.

Step 4: by the F under relatively different threshold number_kDifference identification Optimal-threshold segmentation number；Particularly as follows:

4a) calculate the F under present threshold value number_Δ=F_k-F_k-1If this value is negative value, is set to 0；Under first threshold F_Δ=F₁；

4b) compare the F under different threshold number_ΔValue, selects F_ΔThreshold number when value is maximum is as Optimal-threshold segmentation Number.

Step 5: by Optimal-threshold segmentation number, original image is carried out category division and obtain final segmentation result.

In order to verify effectiveness of the invention, experiment is chosen a width shot image (in Fig. 2 (a) shown in) and 2 width images in Berkeley image data base (as shown in (a) in (a) and Fig. 4 in Fig. 3), (b) in Fig. 2-Fig. 4 be to Determining the segmentation result of Otsu method under threshold number (i.e. maximum variance between clusters), (c) in Fig. 2-Fig. 4 is under given threshold number The segmentation result of maximum entropy method (MEM), (d) in Fig. 2-Fig. 4 is the segmentation result of the present invention.

Simulated effect is analyzed: the present invention utilizes multi-objective Evolutionary Algorithm to split threshold binary image, last optimum In the selection solved, use F_kFunction adaptive can be selected optimal threshold and obtain the final result of threshold Image Segmentation, phase Solving final threshold value for traditional given threshold number and have relatively much progress, segmentation result also increases.

Claims (6)

1. one kind based on multiobject adaptive threshold image partition method, it is characterised in that comprise the following steps:

Step one: input image to be split, if image to be split is coloured image, is first converted into gray level image；

Step 2: carry out the multi-target evolution threshold value of some different threshold number for the gray level image obtained by step one respectively Segmentation, obtains some different Parato disaggregation；

Step 3: use function F_kObtain each Parato and solve the optimal threshold concentrated；

Step 4: by the F under relatively different threshold number_kDifference identification Optimal-threshold segmentation number；

Step 5: by Optimal-threshold segmentation number, original image is carried out category division and obtain final segmentation result.

One the most according to claim 1 is based on multiobject adaptive threshold image partition method, it is characterised in that step Multi-target evolution threshold segmentation method in rapid two particularly as follows:

2a) arranging the parameter of multi-target evolution Threshold segmentation: max-thresholds number is 5, population scale is 200, maximum heredity generation Number is 300, and crossover probability is 0.9, and mutation probability is 0.1, and genovariation scope is 5～15, gene code scope 0～255；

2b) initialization of population, randomly generates 200 individualities, it is assumed that have k the most each chromosome of threshold value to have k gene position, each Gene position value is 0～255, genetic algebra g=1；

2c) calculate two target function values of each individuality in population, two target function values are respectively added to the k of chromosome + 1, k+2 gene position；

2d) utilizing step 2c) population carries out the middle target function value calculated non-dominated ranking and crowding distance calculates, and will Individual sequence value and crowding distance are respectively added to k+3, k+4 gene position of chromosome；

2e) start to evolve, use tournament method to select the individuality of half quantity from population according to sequence value and crowding distance size As parent population；

2f) parent population is intersected and mutation operation, produce progeny population；

2g) current population merged with progeny population and carry out elitist selection, it is thus achieved that a new generation identical with initial population size Population；

If 2h) g > 300, then perform step 3, otherwise, g=g+1, jump to step 2c).

One the most according to claim 2 is based on multiobject adaptive threshold image partition method, it is characterised in that step Rapid 2c) in two object functions be respectively and embody inter-class variance function f₁And maximum possible retain image original information and The entropy function f of edge contour information₂。

One the most according to claim 3 is based on multiobject adaptive threshold image partition method, it is characterised in that meter Calculate two target function values of each individuality in population particularly as follows:

Assume that the number of pixels comprised in piece image is N, the grey level range in image be [0 ..., L], gray level is i's Number of pixels is n_i, then the probability that gray level i occurs is:

$P_i=\frac{n_i}{N}$

Then object function f₁For:

f₁(t₁,t₂... t_k)=w₁(u₁-u_T)²+w₂(u₂-u_T)²+…+w_k+1(u_k+1-u_T)²

Wherein:

$w_1=\underset{i=0}{\overset{t_1-1}{\Σ}}{p}_i,...,w_n=\underset{i=t_{n-1}}{\overset{t_n-t_{n-1}}{\Σ}}{p}_i,...,w_{k+1}=\underset{i=t_k}{\overset{L}{\Σ}}{p}_i;$

$u_1=\underset{i=0}{\overset{t_1-1}{\Σ}}\frac{i*p_i}{w_1},...,u_n=\underset{i=t_{n-1}}{\overset{t_n-t_{n-1}}{\Σ}}\frac{i*p_i}{w_n},...,u_{k+1}=\underset{i=t_k}{\overset{L}{\Σ}}\frac{i*p_i}{w_{k+1}};$

$u_T=\underset{i=0}{\overset{L}{\Σ}}i*p_i$

Wherein, (t₁,t₂... t_k) it is the threshold value of image, the threshold number that k is split by image, w_n(1≤n≤k+1) is the n-th class The gray scale probability of occurrence of pixel and, u_n(1≤n≤k+1) is the average gray of the n-th class pixel, u_TIt it is the average ash of entire image Angle value；

Object function f₂For:

f₂(t₁,t₂... t_k)=H₁+H₂+…+H_k+1

Wherein:

$H_1=-\underset{i=0}{\overset{t_1-1}{\mathrm{\Σ}}}\frac{p_i}{w_1}ln\frac{p_i}{w_1},w_1=\underset{i=0}{\overset{t_1-1}{\mathrm{\Σ}}}{p}_i$

$H_n=-\underset{i=t_{n-1}}{\overset{t_n-t_{n-1}}{\mathrm{\Σ}}}\frac{p_i}{w_n}ln\frac{p_i}{w_n},w_n=\underset{i=t_{n-1}}{\overset{t_n-t_{n-1}}{\mathrm{\Σ}}}{p}_i$

$H_{k+1}=-\underset{i=t_k}{\overset{L}{\Σ}}\frac{p_i}{w_{k+1}}ln\frac{p_i}{w_{k+1}},w_{k+1}=\underset{i=t_k}{\overset{L}{\Σ}}{p}_i$

In formula, (t₁,t₂... t_k) it is the threshold value of image, the threshold number that k is split by image, H_n(1≤n≤k+1) is the n-th class The comentropy sum of pixel, w_n(1≤n≤k+1) be the n-th class pixel gray scale probability of occurrence and, p_iThe probability occurred for pixel i.

One the most according to claim 1 is based on multiobject adaptive threshold image partition method, it is characterised in that step Function F in rapid three_kParticularly as follows:

$F_k=\frac{N*\left(m\right(k+1)-m\left(1\right))}{k*\underset{j=1}{\overset{k+1}{\Σ}}\underset{i\∈C_j}{\Σ}{(g_i-m_j)}^2}$

Wherein, N is the number of pixels of entire image, the threshold number that k is split by image, and m (k+1) is the gray scale of kth+1 class Meansigma methods, m (1) is the average gray of the first kind, m_jFor the average gray of jth class, C_jIt is the collection of pixels of jth class, g_iIt is The gray value of pixel i.

One the most according to claim 1 is based on multiobject adaptive threshold image partition method, it is characterised in that step By the F under relatively different threshold number in rapid four_kDifference identification Optimal-threshold segmentation number, particularly as follows:

4a) calculate the F under present threshold value number_Δ=F_k-F_k-1If this value is negative value, is set to 0；F under first threshold_Δ =F₁；

4b) compare the F under different threshold number_ΔValue, selects F_ΔThreshold number when value is maximum is as Optimal-threshold segmentation number.