CN104318575A - Multi-threshold image segmentation method based on comprehensive learning differential evolution algorithm - Google Patents

Abstract

The invention discloses a multi-threshold image segmentation method based on a comprehensive learning differential evolution algorithm. The method comprises the steps that in the mutation operation process of the differential evolution algorithm, a Binary tournament selection method is utilized to select an individual from species at random, a comprehensive individual is generated by the individual and an optimal individual, then the comprehensive individual serves as a basic individual, a mutation operation is carried out on the basic individual to generate a mutation individual, the searching speed is accelerated as fast as possible while the population diversity is kept, and then a crossover operation operator and a selection operation operator of a traditional differential evolution algorithm are carried out. Meanwhile, a zoom factor value and a crossover probability value are adjusted adaptively according to current search feedback information, so that the robustness of the algorithm is reinforced. The steps are repeatedly executed until a terminal condition is met, and the optimal individual obtained in the computation process is a final segmentation threshold of an image. By means of the multi-threshold image segmentation method based on the comprehensive learning differential evolution algorithm, the probability of local optimum can be reduced, the image segmentation accuracy is improved, the segmentation speed is accelerated, and the real time performance of the segmentation is improved.

Description

A kind of multi-threshold image segmentation method based on integrated learning Differential Evolution Algorithm

Technical field

The present invention relates to Digital Image Segmentation technology, especially a kind of multi-threshold image segmentation method based on integrated learning Differential Evolution Algorithm.

Background technology

Multi-threshold image segmentation is a kind of very important digital image processing method in Modern digital image process.In multi-threshold image segmentation process, usually according to the segmentation criterion preset, find multiple threshold value, identify the part interested in image, thus Iamge Segmentation is become several different part.In multi-threshold image segmentation process, how fast and effeciently the most key step is each threshold value of optimization.But tradition often all to there is search speed based on the multi-threshold image segmentation method of exhaustive search slow, the shortcoming that real-time is not high, especially when number of thresholds is larger, searching for consuming time being usually difficult in practical engineering application accepts.For this reason, intelligent optimization algorithm is applied in multi-threshold image segmentation by people, thus intelligence, solve segmentation threshold rapidly.Such as, wear Qionghai etc. and invent a kind of multi-threshold image segmentation method based on fitness random search behavior in 2010; Tong little Nian etc. proposed a kind of Dual-threshold image segmentation method based on quanta particle swarm optimization in 2010; Zhang Wei etc. proposed the fuzzy entropy coal dust Iamge Segmentation based on Modified particle swarm optimization in 2011.

Differential Evolution Algorithm is a kind of effective modern intelligence optimization algorithm of solving-optimizing problem.The structure of Differential Evolution Algorithm is very simple, easy to understand and realization, and has very strong self-organization, self study and adaptivity, and it to have become in solving-optimizing Study on Problems field a very active study hotspot.Differential Evolution Algorithm has successfully been used for solving multi-threshold image segmentation problem by people, but conventional differential evolution algorithmic often also exists easily be absorbed in local optimum when solving multi-threshold image segmentation problem, segmentation effect still needs to improve, the shortcoming that speed of convergence is slow and real-time is not strong.

Summary of the invention

The present invention is directed to also exist when conventional differential evolution algorithmic carries out multi-threshold image segmentation and be easily absorbed in local optimum, segmentation precision is not high, the shortcoming that splitting speed is slow and real-time is not strong, a kind of multi-threshold image segmentation method based on integrated learning Differential Evolution Algorithm is proposed, the method is in the mutation operation process of differential evolution, binary algorithm of tournament selection method is adopted from population, to select body one by one at random, and it and optimum individual are generated a comprehensive individuality, based on this comprehensive individuality, the individual mutation operation that performs generates variation individuality again, while maintenance population diversity, search speed is accelerated as far as possible with this, then the hybridization of conventional differential evolution algorithmic is performed, select operation operator, meanwhile, adjust the value of zoom factor and probability of crossover according to current search feedback information adaptively, strengthen the robustness of algorithm with this, repeat above-mentioned steps until meet end condition, the optimum individual obtained in computation process, be the segmentation threshold that image is final, compared with congenic method, the present invention can reduce the probability being absorbed in local optimum, improves the precision of Iamge Segmentation, accelerates the speed of segmentation, improves the real-time of segmentation.

Technical scheme of the present invention: a kind of multi-threshold image segmentation method based on integrated learning Differential Evolution Algorithm, comprises the following steps:

Step 1, user's initiation parameter, described initiation parameter comprises segmentation threshold quantity D, Population Size Popsize, maximum evaluation number of times MAX_FEs;

Step 2, current evolution algebraically t=0, and integrated learning rate Pr is set _i ^t=0.5, hybrid rate Cr _i ^t=0.9, zoom factor F _i ^t=0.5, wherein subscript i=1 ..., Popsize, Evaluation: Current number of times FEs=0;

Step 3, produces initial population at random wherein: subscript i=1 ..., Popsize, and for population P _tin i-th individuality, its random initializtion formula is:

$A_{i,j}^t=(j-1)\×255.0/D+\mathrm{rand}\left(\mathrm{0,1}\right)\·(255.0/D-1.0)$

Wherein subscript j=1 ..., D, and D is segmentation threshold quantity; for at population P _tin i-th individuality, store D segmentation threshold; Rand (0,1) obeys equally distributed random real number to produce function between [0,1];

Step 4, calculates population P _tin the adaptive value of each individuality, wherein adaptive value is larger, shows that individuality is more outstanding, individual for any one adaptive value calculate as follows:

$\mathrm{Fit}\left(A_i^t\right)=\underset{k=0}{\overset{D}{\mathrm{\Σ}}}{H}_k,$

Wherein H _kfor the entropy in a kth image intensity value interval, be calculated as follows:

$p_j=\frac{h\left(j\right)}{\underset{j=0}{\overset{255}{\mathrm{\Σ}}}h\left(j\right)},$

as k=0,

as subscript k=1,2 ..., during D-1

as k=D

Wherein h (j) is jth image intensity value sum of all pixels in the picture, p _jfor jth image intensity value probability in the picture; downward rounding operation symbol; w ₀for interval image intensity value accumulated probability and, H ₀for interval the entropy of image intensity value, w _kfor interval image intensity value accumulated probability and, subscript k=1,2 ... D-1, H _kfor interval the entropy of image intensity value, w _dfor interval image intensity value accumulated probability and, H _dfor interval the entropy of image intensity value; Then Evaluation: Current number of times FEs=FEs+Popsize, and preserve population P ^tthe maximum individuality of middle adaptive value is optimum individual Best ^t;

Step 5, makes counter i=1;

Step 6, if counter i is greater than Population Size Popsize, then forwards step 15 to, otherwise forwards step 7 to;

Step 7, calculates individual current composite learning rate NPr _i ^t, computing formula is as follows:

Wherein r1 is the random real number produced between [0,1];

Step 8, according to individuality current composite learning rate NPr _i ^t, to individuality produce an integrated learning individuality its step is as follows:

Step 8.1, makes counter j=1;

Step 8.2, if counter j is greater than D, then forwards step 9 to, otherwise forwards step 8.3 to;

Step 8.3, produces a random real number r2 between [0,1]; If r2 is less than individuality current composite learning rate NPr _i ^tthen forward step 8.7 to, otherwise forward step 8.4 to;

Step 8.4, random generation two unequal positive integer RI1, RI2 between [1, Popsize];

Step 8.5, if individual adaptive value be greater than individuality adaptive value, then otherwise $B_{i,j}^t=A_{\mathrm{RI}2,j}^t;$

Step 8.6, makes counter j=j+1, forwards step 8.2 to;

Step 8.7, make counter j=j+1, forward step 8.2 to;

Step 9, calculates individuality as follows current zoom factor NF _i ^twith current hybrid rate NCr _i ^t:

Wherein r3, r4 is [0,1] the random real number produced between, randc (0.5,0.3) be using 0.5 as location parameter, 0.3 produces the random real-number function of obeying Cauchy's distribution as scale parameter, rand (0,1) be produce to obey equally distributed random real-number function between [0,1];

Step 10, individual with integrated learning based on individual, and with NF _i ^tfor individuality current zoom factor, NCr _i ^tfor individuality current hybrid rate, produce individual test individual and it is individual to calculate test adaptive value its step is as follows:

Step 10.1, makes counter j=1;

Step 10.2, a random generation positive integer jRand between [1, D];

Step 10.3, random generation two unequal positive integer RI3, RI4 between [1, Popsize];

Step 10.4, if counter j is greater than D, then forwards step 10.9 to, otherwise forwards step 10.5 to;

Step 10.5, produces a random real number r5, if r5 is less than individuality between [0,1] current hybrid rate NCr _i ^tor jRand equals counter j, then forward step 10.6 to, otherwise forwards step 10.7 to;

Step 10.6, $U_{i,j}^t=B_{i,j}^t+{{\mathrm{NF}}_i}^t\·(A_{\mathrm{RI}3,j}^t-A_{\mathrm{RI}4,j}^t);$ Forward step 10.8 to;

Step 10.7, $U_{i,j}^t=A_{i,j}^t;$

Step 10.8, makes counter j=j+1, forwards step 10.4 to;

Step 10.9, calculates test individual adaptive value forward step 11 to;

Step 11, as follows at individuality individual with test between select individuality and enter population of future generation:

Step 12, upgrades individuality as follows integrated learning rate Pr _i ^t, zoom factor F _i ^t, hybrid rate Cr _i ^t:

Step 13, makes counter i=i+1;

Step 14, forwards step 6 to;

Step 15, Evaluation: Current number of times FEs=FEs+Popsize, preserves population P _tthe maximum individuality of middle adaptive value is optimum individual Best ^t; Current evolution algebraically t=t+1;

Step 16, repeats step 5 to step 15 until Evaluation: Current number of times FEs terminates after reaching MAX_FEs, the optimum individual Best obtained in implementation ^tbe D segmentation threshold of segmentation image, and with the D an obtained segmentation threshold to Image Segmentation Using.

Therefore, tool of the present invention has the following advantages: the present invention efficiently utilizes the advantageous combination information of random individual and optimum individual to coordinate the balance of global search and Local Search, search speed is accelerated as far as possible while maintenance population diversity, reduce the probability being absorbed in local optimum, and convergence speedup speed; On the other hand, the present invention adjusts the value of zoom factor and probability of crossover adaptively according to current search feedback information, enhances the robustness of algorithm; Compared with congenic method, the present invention has fully utilized the individual information of differential evolution population, can reduce the probability being absorbed in local optimum, improves the precision of Iamge Segmentation, accelerates the speed of segmentation, improves the real-time of segmentation.

Accompanying drawing explanation

Fig. 1 is Lena image;

Fig. 2 is the segmentation result of the present invention to Lena image.

Embodiment

Below by embodiment, and by reference to the accompanying drawings, technical scheme of the present invention is described in further detail.

Embodiment:

The present embodiment is based on the Lena Image Segmentation Using shown in Fig. 1, and specific embodiment of the invention step is as follows:

Step 1, user's initiation parameter, described initiation parameter comprises segmentation threshold quantity D=4, Population Size Popsize=100, maximum evaluation number of times MAX_FEs=60000;

Step 2, current evolution algebraically t=0, and integrated learning rate Pr is set _i ^t=0.5, hybrid rate Cr _i ^t=0.9, zoom factor F _i ^t=0.5, wherein subscript i=1 ..., Popsize, Evaluation: Current number of times FEs=0;

Step 3, produces initial population at random wherein: subscript i=1 ..., Popsize, and for population P _tin i-th individuality, its random initializtion formula is:

$A_{i,j}^t=(j-1)\×255.0/D+\mathrm{rand}\left(\mathrm{0,1}\right)\·(255.0/D-1.0)$

Wherein subscript j=1 ..., D, and D is segmentation threshold quantity; for at population P _tin i-th individuality, store D segmentation threshold; Rand (0,1) obeys equally distributed random real number to produce function between [0,1];

Step 4, calculates population P _tin the adaptive value of each individuality, wherein adaptive value is larger, shows that individuality is more outstanding, individual for any one adaptive value calculate as follows:

$\mathrm{Fit}\left(A_i^t\right)=\underset{k=0}{\overset{D}{\mathrm{\Σ}}}{H}_k,$

Wherein H _kfor the entropy in a kth image intensity value interval, be calculated as follows:

$p_j=\frac{h\left(j\right)}{\underset{j=0}{\overset{255}{\mathrm{\Σ}}}h\left(j\right)},$

as k=0,

as subscript k=1,2 ..., during D-1

as k=D

Wherein h (j) is the sum of all pixels of a jth image intensity value in Lena image, p _jfor the probability of a jth image intensity value in Lena image; for downward rounding operation accords with; w ₀for interval image intensity value accumulated probability and, H ₀for interval the entropy of image intensity value, w _kfor interval image intensity value accumulated probability and, subscript k=1,2 ... D-1, H _kfor interval the entropy of image intensity value, w _dfor interval image intensity value accumulated probability and, H _dfor interval the entropy of image intensity value; Then Evaluation: Current number of times FEs=FEs+Popsize, and preserve population P ^tthe maximum individuality of middle adaptive value is optimum individual Best ^t;

Step 5, makes counter i=1;

Step 6, if counter i is greater than Population Size Popsize, then forwards step 15 to, otherwise forwards step 7 to;

Step 7, calculates individual current composite learning rate NPr _i ^t, computing formula is as follows:

Wherein r1 is the random real number produced between [0,1];

Step 8, according to individuality current composite learning rate NPr _i ^t, to individuality produce an integrated learning individuality its step is as follows:

Step 8.1, makes counter j=1;

Step 8.2, if counter j is greater than D, then forwards step 9 to, otherwise forwards step 8.3 to;

Step 8.3, produces a random real number r2 between [0,1]; If r2 is less than individuality current composite learning rate NPr _i ^tthen forward step 8.7 to, otherwise forward step 8.4 to;

Step 8.4, random generation two unequal positive integer RI1, RI2 between [1, Popsize];

Step 8.5, if individual adaptive value be greater than individuality adaptive value, then otherwise $B_{i,j}^t=A_{\mathrm{RI}2,j}^t;$

Step 8.6, makes counter j=j+1, forwards step 8.2 to;

Step 8.7, make counter j=j+1, forward step 8.2 to;

Step 9, calculates individuality as follows current zoom factor NF _i ^twith current hybrid rate NCr _i ^t:

Wherein r3, r4 is [0,1] the random real number produced between, randc (0.5,0.3) be using 0.5 as location parameter, 0.3 produces the random real-number function of obeying Cauchy's distribution as scale parameter, rand (0,1) be produce to obey equally distributed random real-number function between [0,1];

Step 10, individual with integrated learning based on individual, and with NF _i ^tfor individuality current zoom factor, NCr _i ^tfor individuality current hybrid rate, produce individual test individual and it is individual to calculate test adaptive value its step is as follows:

Step 10.1, makes counter j=1;

Step 10.2, a random generation positive integer jRand between [1, D];

Step 10.3, random generation two unequal positive integer RI3, RI4 between [1, Popsize];

Step 10.4, if counter j is greater than D, then forwards step 10.9 to, otherwise forwards step 10.5 to;

Step 10.5, produces a random real number r5, if r5 is less than individuality between [0,1] current hybrid rate NCr _i ^tor jRand equals counter j, then forward step 10.6 to, otherwise forwards step 10.7 to;

Step 10.6, $U_{i,j}^t=B_{i,j}^t+{{\mathrm{NF}}_i}^t\·(A_{\mathrm{RI}3,j}^t-A_{\mathrm{RI}4,j}^t);$ Forward step 10.8 to;

Step 10.7, $U_{i,j}^t=A_{i,j}^t;$

Step 10.8, makes counter j=j+1, forwards step 10.4 to;

Step 10.9, calculates test individual adaptive value forward step 11 to;

Step 11, as follows at individuality individual with test between select individuality and enter population of future generation:

Step 12, upgrades individuality as follows integrated learning rate Pr _i ^t, zoom factor F _i ^t, hybrid rate Cr _i ^t:

Step 13, makes counter i=i+1;

Step 14, forwards step 6 to;

Step 15, Evaluation: Current number of times FEs=FEs+Popsize, preserves population P _tthe maximum individuality of middle adaptive value is optimum individual Best ^t; Current evolution algebraically t=t+1;

Step 16, repeats step 5 to step 15 until Evaluation: Current number of times FEs terminates after reaching MAX_FEs, the optimum individual Best obtained in implementation ^tbe D segmentation threshold of segmentation Lena image, and with the D an obtained segmentation threshold to Lena Image Segmentation Using.

Specific embodiment described herein is only to the explanation for example of the present invention's spirit.Those skilled in the art can make various amendment or supplement or adopt similar mode to substitute to described specific embodiment, but can't depart from spirit of the present invention or surmount the scope that appended claims defines.

Claims (1)

1. based on a multi-threshold image segmentation method for integrated learning Differential Evolution Algorithm, it is characterized in that, comprise the following steps:

Step 1, user's initiation parameter, described initiation parameter comprises segmentation threshold quantity D, Population Size Popsize, maximum evaluation number of times MAX_FEs;

Step 2, current evolution algebraically t=0, and integrated learning rate Pr is set _i ^t=0.5, hybrid rate Cr _i ^t=0.9, zoom factor F _i ^t=0.5, wherein subscript i=1 ..., Popsize, Evaluation: Current number of times FEs=0;

Step 3, produces initial population at random wherein: subscript i=1 ..., Popsize, and for population P _tin i-th individuality, its random initializtion formula is:

$A_{i,j}^t=(j-1)\×255.0/D+\mathrm{rand}\left(\mathrm{0,1}\right)\·(255.0/D-1.0)$

Wherein subscript j=1 ..., D, and D is segmentation threshold quantity; for at population P _tin i-th individuality, store D segmentation threshold; Rand (0,1) obeys equally distributed random real number to produce function between [0,1];

Step 4, calculates population P _tin the adaptive value of each individuality, wherein adaptive value is larger, shows that individuality is more outstanding, individual for any one adaptive value calculate as follows:

$\mathrm{Fit}\left(A_i^t\right)=\underset{k=0}{\overset{D}{\mathrm{\Σ}}}{H}_k,$

Wherein H _kfor the entropy in a kth image intensity value interval, be calculated as follows:

$p_j=\frac{h\left(j\right)}{\underset{j=0}{\overset{255}{\mathrm{\Σ}}}h\left(j\right)},$

as k=0,

as subscript k=1,2 ..., during D-1

as k=D

Wherein h (j) is jth image intensity value sum of all pixels in the picture, p _jfor jth image intensity value probability in the picture; for downward rounding operation accords with; w ₀for interval image intensity value accumulated probability and, H ₀for interval the entropy of image intensity value, w _kfor interval image intensity value accumulated probability and, subscript k=1,2 ... D-1, H _kfor interval the entropy of image intensity value, w _dfor interval image intensity value accumulated probability and, H _dfor interval the entropy of image intensity value; Then Evaluation: Current number of times FEs=FEs+Popsize, and preserve population P ^tthe maximum individuality of middle adaptive value is optimum individual Best ^t;

Step 5, makes counter i=1;

Step 6, if counter i is greater than Population Size Popsize, then forwards step 15 to, otherwise forwards step 7 to;

Step 7, calculates individual current composite learning rate NPr _i ^t, computing formula is as follows:

Wherein r1 is the random real number produced between [0,1];

Step 8, according to individuality current composite learning rate NPr _i ^t, to individuality produce an integrated learning individuality its step is as follows:

Step 8.1, makes counter j=1;

Step 8.2, if counter j is greater than D, then forwards step 9 to, otherwise forwards step 8.3 to;

Step 8.3, produces a random real number r2 between [0,1]; If r2 is less than individuality current composite learning rate NPr _i ^tthen forward step 8.7 to, otherwise forward step 8.4 to;

Step 8.4, random generation two unequal positive integer RI1, RI2 between [1, Popsize];

Step 8.5, if individual adaptive value be greater than individuality adaptive value, then otherwise $B_{i,j}^t=A_{\mathrm{RI}2,j}^t;$

Step 8.6, makes counter j=j+1, forwards step 8.2 to;

Step 8.7, make counter j=j+1, forward step 8.2 to;

Step 9, calculates individuality as follows current zoom factor NF _i ^twith current hybrid rate NCr _i ^t:

Wherein r3, r4 is [0,1] the random real number produced between, randc (0.5,0.3) be using 0.5 as location parameter, 0.3 produces the random real-number function of obeying Cauchy's distribution as scale parameter, rand (0,1) be produce to obey equally distributed random real-number function between [0,1];

Step 10, individual with integrated learning based on individual, and with NF _i ^tfor individuality current zoom factor, NCr _i ^tfor individuality current hybrid rate, produce individual test individual and it is individual to calculate test adaptive value its step is as follows:

Step 10.1, makes counter j=1;

Step 10.2, a random generation positive integer jRand between [1, D];

Step 10.3, random generation two unequal positive integer RI3, RI4 between [1, Popsize];

Step 10.4, if counter j is greater than D, then forwards step 10.9 to, otherwise forwards step 10.5 to;

Step 10.5, produces a random real number r5, if r5 is less than individuality between [0,1] current hybrid rate NCr _i ^tor jRand equals counter j, then forward step 10.6 to, otherwise forwards step 10.7 to;

Step 10.6, $U_{i,j}^t=B_{i,j}^t+{{\mathrm{NF}}_i}^t\·(A_{\mathrm{RI}3,j}^t-A_{\mathrm{RI}4,j}^t);$ Forward step 10.8 to;

Step 10.7, $U_{i,j}^t=A_{i,j}^t;$

Step 10.8, makes counter j=j+1, forwards step 10.4 to;

Step 10.9, calculates test individual adaptive value forward step 11 to;

Step 11, as follows at individuality individual with test between select individuality and enter population of future generation:

Step 12, upgrades individuality as follows integrated learning rate Pr _i ^t, zoom factor F _i ^t, hybrid rate Cr _i ^t:

Step 13, makes counter i=i+1;

Step 14, forwards step 6 to;

Step 15, Evaluation: Current number of times FEs=FEs+Popsize, preserves population P _tthe maximum individuality of middle adaptive value is optimum individual Best ^t; Current evolution algebraically t=t+1;

Step 16, repeats step 5 to step 15 until Evaluation: Current number of times FEs terminates after reaching MAX_FEs, the optimum individual Best obtained in implementation ^tbe D segmentation threshold of segmentation image, and with the D an obtained segmentation threshold to Image Segmentation Using.