CN104331893A - Complex image multi-threshold segmentation method - Google Patents

Abstract

The invention provides a hybrid PSOGSA with generalized opposition-based learning. Based on improved particle swarm optimization and gravitational search algorithm, the hybrid PSOGSA with generalized opposition-based learning comprises the steps of: initializing parameters of the PSO and GSA to generate initial positions of all particles randomly; deciding to generate reverse population based on a random number to calculate a fitness value or calculate a fitness value of the current population; mutating global optimal particles, taking the global optimal particles with a relatively big fitness value from the global optimal particles before and after mutation as new global optimal particles; outputting positions of the global optimal particles as image segmentation thresholds through iteration. According to the invention, a brand-new method for the complex image multi-threshold segmentation is provided, and multi-threshold segmentation with high precision and high stability is realized.

Description

A kind of complicated image multi-threshold segmentation method

Technical field

Patent of the present invention belongs to image processing algorithm design field, relates to a kind of complicated image multi-threshold segmentation method of Iamge Segmentation field.

Background technology

Carrying out image threshold segmentation splits simply effective, real-time feature because of it, receives and pays close attention to widely.Multi-threshold image segmentation, as the expansion of Threshold segmentation, have the advantage distinguishing background and multiple target, but shortcoming is calculation of complex, length consuming time.Increasing biological heuritic approach is applied in image segmentation algorithm in recent years, for the quick optimizing of image threshold.Wu one is congruent to 2014 and proposes minimum cross entropy reciprocal as thresholding selection rule, and using artificial ant colony algorithm is optimized calculating, has carried out single threshold segmentation to flame image.Shortcoming is that minimum cross entropy calculating formula reciprocal is complicated, and is only applicable to single threshold Iamge Segmentation, and is not suitable for the multi-threshold segmentation of complicated image.Chen Kai equals within 2014, to propose the multi thresholds calculating that glowworm swarm algorithm optimizes maximum two-dimentional Kapur entropy, achieves the multi-threshold image segmentation of complex target.Shortcoming is that the stability of algorithm is undesirable, and continuous run time result fluctuation is larger.Application number is the Dual-threshold image segmentation method that the Chinese invention patent application of CN201410040869.9 proposes based on bat algorithm optimization fuzzy entropy.Shortcoming is only applicable to dual threshold Iamge Segmentation, is not suitable for the Iamge Segmentation of more multi thresholds.

In image threshold searching process, single biological heuritic approach such as artificial bee colony algorithm, glowworm swarm algorithm, bat algorithm etc. generally have the defect that local search ability by force, is not easily absorbed in local optimum, this will cause the final segmentation threshold obtained not to be desirable image segmentation threshold, even greatly differ from each other from desired threshold, cause Iamge Segmentation inaccurate.Therefore design and be a kind ofly applicable to the high precision of complicated image, the multi-threshold segmentation method of high stability seems particularly important.

Summary of the invention

Goal of the invention:, complicated image not strong for conventional images dividing method local search ability splits inaccurate defect, the present invention proposes a kind of multi-threshold segmentation method of high precision for complicated image, high stability.

Technical scheme: for solving the problems of the technologies described above, the present invention is based on Modified particle swarm optimization and gravitation searches for hybrid algorithm, provide a kind of complicated image multi-threshold segmentation method (hybrid PSOGSA with generalized opposition-based learning, GOPSOGSA) to comprise the following steps:

(1) initialization particle group optimizing and gravitation search for hybrid algorithm parameters: total number of particles N, Studying factors c ₁, c ₂, inertia weight ω, reverse Probability p ₀, maximum iteration time MAXNGER, the initial position of all particles of stochastic generation;

(2) if several rand (0,1) of stochastic generation one 0 ~ 1 are less than p ₀, enter step (3), otherwise enter step (4);

(3) generate reverse population, calculate current population and the fitness value of oppositely planting particle, therefrom choose N number of optimal particle and form new population, enter step (5);

(4) fitness value of current population is calculated;

(5) upgrade global optimum particle according to fitness value, global optimum's particle made a variation, compare its fitness value with variation particle, get fitness value larger as new global optimum's particle;

(6) if current iteration number of times exceedes maximum iteration time, then stop iteration, export global optimum's particle position, as image segmentation threshold, multi-threshold segmentation is carried out to image, otherwise speed and the position of the formula Population Regeneration particle in hybrid algorithm is searched for according to particle group optimizing and gravitation, current iteration number of times adds 1, enters step (2).

Particularly, step (3) is further comprising the steps of:

(3.1) according to the Generalized Anti in multi-Level Threshold Image Segmentation to particle definition, generate the reverse population GOP of current population P; X _dfor a particle of population P, then its reverse particle X _d ^*by following formulae discovery:

$\left\{\begin{array}{c}{X_d}^{*}=k(a_d+b_d)-X_d\\ {X_d}^{*}=a_d,\mathrm{if}{X_d}^{*}<a_d,\\ {X_d}^{*}=b_d,\mathrm{if}{x_D}^{*}>b_d\end{array}\right.$ Wherein X _d∈ [a _d, b _d], k is the random number of [0,1].

(3.2) by the multi-Level Threshold Image Segmentation evaluation function preset as population's fitness calculating formula, calculate the fitness of current population P and reverse population GOP;

(3.3) sequence that 2N particle carries out fitness is amounted to current population P and reverse population GOP, therefrom select the highest particle of N number of fitness and form new population P, enter step (5).

Particularly, step (5) is further comprising the steps of:

(5.1) from current population P, the highest particle of fitness is picked out as global optimum's particle;

(5.2) global optimum's particle is made a variation according to normal mutation formula, generate variation particle;

Gbest _dthe component that global optimum's particle is tieed up at d, gbest _d ^*it is the optimal particle after variation.

$W_d\left(t\right)=\left(\underset{i=1}{\overset{\mathrm{PopSize}}{\mathrm{\&Sigma;}}}{V}_i^d\left(t\right)\right)/\mathrm{PopSize}$

gbest _d ^*(t)＝gbest _d(t)+W _d(t)·N(0,1)

In formula, PopSize represents the sum of particle, and N (0,1) is Standard Normal Distribution, and probability density function and distribution function are distinguished as follows:

$\mathrm{\&Phi;}\left(x\right)=\frac{1}{\sqrt{2\mathrm{\&pi;}}}{\&Integral;}_{-\&infin;}^x{e}^{{-t}^2/2}\mathrm{dt}$

(5.3) compare the fitness value of optimal particle and variation particle, get fitness in both larger as new global optimum's particle.

Particularly, step (6) is further comprising the steps of:

(6.1) if current iteration number of times exceedes maximum iteration time MAXNGER, then stop iteration, export global optimum's particle position, as image segmentation threshold, multi-threshold segmentation is carried out to image;

(6.2) if current iteration number of times does not exceed maximum iteration time MAXNGER, then search for speed and the position of the formula Population Regeneration particle in hybrid algorithm according to particle group optimizing and gravitation, current iteration number of times adds 1, enters step (2).

(6.2.1) in particle group optimizing and gravitation search algorithm, each particle represents a feasible solution, and there are oneself speed and position in each moment, if i-th particle is respectively X when the t time iteration in the position of d dimension and speed _i ^d(t) and V _i ^d(t), d=1,2 ..., D, D are search volume dimension, and group optimal solution is gbest,

More new formula is as follows for the speed of population particle:

V _i ^d(t+1)＝ωV _i ^d(t)+c′ ₁·rand ₁·a _i ^d(t)+c′ ₂·rand ₂·(gbest-X _i ^d(t))

In formula, ω is the inertia weight of particle; C ' ₁, c ' ₂for speedup factor; Rand ₁, rand ₂be respectively the random number of [0,1];

The location updating formula of population particle is as follows:

X _i ^d(t+1)＝X _i ^d(t)+V _i ^d(t+1)

(6.2.2) in the speed of step (6.2.1) population particle more in new formula, a _i ^dt () represents i-th particle acceleration in d dimension when the t time iteration, computing method are as follows:

a _i ^d(t+1)＝F _i ^d(t)/M _i ^d(t)

$F_i^d\left(t\right)=\underset{j=1,j\&NotEqual;i}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{rand}}_j\&CenterDot;F_{\mathrm{ij}}^d\left(t\right)$

F _i ^d(t), M _i ^dt () represents the size that i-th particle is made a concerted effort suffered by d dimension when the t time iteration and inertial mass respectively;

Under the computing formula of making a concerted effort:

$F_{\mathrm{ij}}^d\left(t\right)=\left[\frac{G\left(t\right)\&CenterDot;M_i\left(t\right)\&CenterDot;M_j\left(t\right)}{R_{\mathrm{ij}}\left(t\right)+\mathrm{\&epsiv;}}\right]\&CenterDot;[{X_j}^d\left(t\right)-{X_i}^d\left(t\right)]$

$F_i^d\left(t\right)=\underset{j=1,j\&NotEqual;i}{\overset{N}{\mathrm{\&Sigma;}}}{\mathrm{rand}}_j\&CenterDot;F_{\mathrm{ij}}^d\left(t\right)$

In formula, N is total number of particles; F _ij ^dt () represents that particle j is to the gravitation of particle i; Rand _jfor the random number of [0,1];

R _ij(t)=|| X _i(t), X _j(t) || ₂for the Euclidean distance of particle i and particle j; ε is the constant that a value is very little; G (t) is gravitational constant,

Calculating formula is as follows, wherein G ₀be constant with a, t is current iteration number of times, and maxt is maximum iteration time;

G(t)＝G ₀·exp(-a·t/maxt)

The inertial mass of particle can be tried to achieve by following formula:

$m_i\left(t\right)=\frac{{\mathrm{fit}}_i\left(t\right)-\mathrm{worst}\left(t\right)}{\mathrm{best}\left(t\right)-\mathrm{worst}\left(t\right)}$

$M_i\left(t\right)=m_i\left(t\right)/\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{m}_j\left(t\right)$

Fit in formula _it () represents the fitness value of i-th particle when the t time iteration;

For the multi-Level Threshold Image Segmentation problem of maximizing, best (t) and worst (t) is tried to achieve by following formula:

$\mathrm{best}\left(t\right)=\underset{j\&Element;\{\mathrm{1,2},...,N\}}{\max }{\mathrm{fit}}_j\left(t\right)$

$\mathrm{worst}\left(t\right)=\underset{j\&Element;\{\mathrm{1,2},...,N\}}{\min }{\mathrm{fit}}_j\left(t\right)$

Except technical matters, the technical characteristic forming technical scheme and the advantage brought by the technical characteristic of these technical schemes that the present invention recited above solves, the advantage that the other technologies feature comprised in the other technologies problem that complicated image multi-threshold segmentation method of the present invention can solve, technical scheme and these technical characteristics bring, will make further detailed description by reference to the accompanying drawings.

Accompanying drawing explanation

Fig. 1 is the basic flow sheet of complicated image multi-threshold segmentation method of the present invention;

The specific implementation process flow diagram that Fig. 2 is is multi-Level Threshold Image Segmentation evaluation function with multi thresholds Otsu algorithm;

Fig. 3 carries out two threshold values of the GOPSOGSA of the inventive method proposition, three threshold values, four threshold values and five Threshold segmentation design sketchs to Lena figure;

Fig. 4 carries out two threshold values of the GOPSOGSA of the inventive method proposition, three threshold values, four threshold values and five Threshold segmentation design sketchs to Gold hill figure;

Fig. 5 carries out two threshold values of the GOPSOGSA of the inventive method proposition, three threshold values, four threshold values and five Threshold segmentation design sketchs to Couple figure;

Fig. 6 carries out two threshold values of the GOPSOGSA of the inventive method proposition, three threshold values, four threshold values and five Threshold segmentation design sketchs to the base board defect figure of QFN packaged chip.

Embodiment

Embodiment:

As depicted in figs. 1 and 2, the multi thresholds Otsu algorithm of the present embodiment, also known as multi thresholds maximum between-cluster variance algorithm, is one of the most frequently used multi-Level Threshold Image Segmentation evaluation function.It is the gamma characteristic by image, image is divided into background and a few part of some targets.Inter-class variance between background and some targets is larger, illustrates that the difference of composing images part is larger, when partial target mistake is divided into background or part background mistake to be divided into target that part difference all can be caused to diminish.Therefore, the segmentation making inter-class variance maximum means that misclassification probability is minimum.Its computing formula is as follows, works as σ ²time maximum, corresponding solution is global optimum's segmentation threshold.

$\mathrm{Otsu}={\mathrm{\&sigma;}}^2=\underset{k=1}{\overset{j}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_k{({\mathrm{\&mu;}}_k-{\mathrm{\&mu;}}_T)}^2$

Its concrete implementation step is as follows:

(1) input picture.Initialization particle group optimizing and gravitation search for hybrid algorithm parameters: total number of particles N, Studying factors c ₁, c ₂, inertia weight ω, reverse Probability p ₀, maximum iteration time MAXNGER.The initial position of all particles of stochastic generation;

(2) if several rand (0,1) of stochastic generation one 0 ~ 1 are less than p ₀, enter step (3), otherwise enter step and turn (4);

(3) generate reverse population, calculate current population and the Otsu functional value of oppositely planting particle, the particle therefrom choosing N number of Otsu functional value larger forms new population, enters step (5);

(4) the Otsu functional value of current population is calculated;

(5) upgrade global optimum particle according to Otsu functional value, global optimum's particle made a variation, compare its Otsu functional value with variation particle, get Otsu functional value larger as new global optimum's particle;

(6) if current iteration number of times exceedes maximum iteration time, then stop iteration, export global optimum's particle position, as image segmentation threshold, multi-threshold segmentation is carried out to image, otherwise speed and the position of the formula Population Regeneration particle in hybrid algorithm is searched for according to particle group optimizing and gravitation, current iteration number of times adds 1, enters step (2).

The complicated image multi-threshold segmentation method that the present invention proposes, has obvious superiority compared with the conventional method in the stability of Iamge Segmentation precision and operation continuously.Below by way of one group of experiment, the superiority of the inventive method compared with gravitation search algorithm GSA, glowworm swarm algorithm FA is described.

As Fig. 3, shown in Fig. 4, Fig. 5 and Fig. 6, scheme using the Lena of classics respectively, Gold hill schemes, Couple figure as object, uses multi-threshold image segmentation algorithm of the present invention to test with the base board defect image of QFN (Quad Flat Non-Lead) packaged chip.Wherein Studying factors c ₁=0.5, c ₂=1.5, inertia weight ω=1.2, particle rapidity V ∈ [-5,5]; Experimental situation is CPU 2.4GHz, internal memory 8GB, MATLAB R2014a.The original image that Fig. 3 sets forth 4 width images and two threshold values of GOPSOGSA, three threshold values, four threshold values and the five Threshold segmentation results that adopt the inventive method to propose, wherein m is number of thresholds.

By the inventive method and based on gravitation search algorithm GSA, compare based on the Otsu multi-threshold image segmentation algorithm of glowworm swarm algorithm FA.Three kinds of algorithm particle/light of firefly borer populations are N=50, maximum iteration time MAXNGER=100.Initial Attraction Degree β in FA ₀=1, step factor a=0.5.The result of carrying out multi-threshold segmentation to four width images with three kinds of algorithms is as shown in table 1.

As shown in Table 1, for Complex multi-target image, when segmentation threshold quantity more (four, five threshold values), method of the present invention has embodied obvious advantage in segmentation accuracy.

Table 1

Adopt signal to noise ratio (S/N ratio) PSNR and the standard deviation sigma of running the Otsu functional value obtained for 30 times continuously to evaluate segmentation performance and the stability of three kinds of algorithms, result is as shown in table 2.Snr computation formula is wherein i with be respectively the image after the original image and segmentation being of a size of M × N.The standard deviation calculating formula of Otsu functional value is wherein k=30 is continuous number of run, f _iwith represent i-th Otsu functional value and k the continuous mean value running rear Otsu functional value respectively.

As shown in Table 2, when segmentation threshold quantity is more (four, five threshold values), the PSNR value of the inventive method is better than GSA and FA, the Otsu functional value standard deviation using the inventive method to obtain also will be significantly less than GSA and FA, decrease 98.5%, 90.6% respectively during five threshold values compared with GSA and FA, embody the inventive method continuous motion time cut quality higher, have more stability.

Table 2

In sum, the complicated image multi-threshold segmentation method that the present invention proposes is a kind of multi-Level Threshold Image Segmentation method of high precision, high stability.

Claims (5)

1. a complicated image multi-threshold segmentation method, is characterized in that comprising the following steps:

(1) initialization particle group optimizing and gravitation search for hybrid algorithm parameters: total number of particles N, Studying factors c ₁, c ₂, inertia weight ω, reverse Probability p ₀, maximum iteration time MAXNGER, the initial position of all particles of stochastic generation;

(2) if several rand (0,1) of stochastic generation one 0 ~ 1 are less than p ₀, enter step (3), otherwise enter step (4);

(3) generate reverse population, calculate current population and the fitness value of oppositely planting particle, therefrom choose N number of optimal particle and form new population, enter step (5);

(4) fitness value of current population is calculated;

(5) upgrade global optimum particle according to fitness value, global optimum's particle made a variation, compare its fitness value with variation particle, get fitness value larger as new global optimum's particle;

(6) if current iteration number of times exceedes maximum iteration time, then stop iteration, export global optimum's particle position, as image segmentation threshold, multi-threshold segmentation is carried out to image, otherwise speed and the position of the formula Population Regeneration particle in hybrid algorithm is searched for according to particle group optimizing and gravitation, current iteration number of times adds 1, enters step (2).

2. a kind of complicated image multi-threshold segmentation method according to claim 1, is characterized in that: described step (3) is further comprising the steps of:

(3.1) according to the Generalized Anti in multi-Level Threshold Image Segmentation to particle definition, generate the reverse population GOP of current population P;

(3.2) by the multi-Level Threshold Image Segmentation evaluation function preset as population's fitness calculating formula, calculate the fitness of current population P and reverse population GOP;

(3.3) sequence that 2N particle carries out fitness is amounted to current population P and reverse population GOP, therefrom select the highest particle of N number of fitness and form new population P, enter step (5).

3. a kind of complicated image multi-threshold segmentation method according to claim 1, is characterized in that: described step (5) is further comprising the steps of:

(5.1) from current population P, the highest particle of fitness is picked out as global optimum's particle;

(5.2) global optimum's particle is made a variation according to normal mutation formula, generate variation particle;

(5.3) compare the fitness value of optimal particle and variation particle, get fitness in both larger as new global optimum's particle.

4. a kind of complicated image multi-threshold segmentation method according to claim 1, is characterized in that: described step (6) is further comprising the steps of:

(6.1) if current iteration number of times exceedes maximum iteration time MAXNGER, then stop iteration, export global optimum's particle position, as image segmentation threshold, multi-threshold segmentation is carried out to image;

(6.2) if current iteration number of times does not exceed maximum iteration time MAXNGER, then search for speed and the position of the formula Population Regeneration particle in hybrid algorithm according to particle group optimizing and gravitation, current iteration number of times adds 1, enters step (2).

5. a kind of complicated image multi-threshold segmentation method according to claim 4, it is characterized in that: in described step (6.2), more new formula is as follows for the speed of population particle:

V _i ^d(t+1)＝ωV _i ^d(t)+c′ ₁·rand ₁·a _i ^d(t)+c′ ₂·rand ₂·(gbest-X _i ^d(t))

In formula, ω is the inertia weight of particle; C ' ₁, c ' ₂for speedup factor; Rand ₁, rand ₂be respectively the random number of [0,1], X _i ^d(t) and V _i ^d(t) be respectively i-th particle when the t time iteration d dimension position and speed, d=1,2 ..., D, D are search volume dimension, a _i ^dt () represents i-th particle acceleration in d dimension when the t time iteration;

The location updating formula of population particle is as follows:

X _i ^d(t+1)＝X _i ^d(t)+V _i ^d(t+1)。