⛄一、和声搜索算法的图像阈值寻优算法简介

苹果图像识别是指将苹果果实从枝叶、土壤、天空等背景中分离出来, 即图像分割。最大类间方差法 (OTSU算法)是由日本学者大津展之 (OTSU) 提出的全局阈值选取方法。该方法存在缺乏自适应性、易造成噪声干扰和过分割现象、运算需要大量的时间等问题, 需要进一步改进。

和声搜索 (Harmony search, HS) 算法是通过类比音乐和最优化问题相似性提出的启发式智能进化算法。算法是对音乐演奏中乐师们凭借自己的记忆, 通过反复调整乐队中各乐器的音调, 最终达到一个美妙和声状态过程的模拟。HS算法概念简单, 收敛速度快, 容易实现, 解的产生方式新颖, 且只有少数参数需要调整, 在有关问题上展示了较遗传算法、模拟退火算法和禁忌搜索更好的性能。本文采用和声搜索算法与经典阈值分割Otsu算法结合的方法, 对苹果图像进行阈值寻优。

1 和声搜索算法思想
随机产生HMS个初始解 (和声) 放入和声记忆库 (Harmony Memory, HM) 内, HMS为和声记忆库的大小, HMCR为和声记忆库取值概率, PAR为音调微调概率, BW为音调微调带宽, Tmax为算法创作的次数。且r1, r2, r∈[0, 1]。

通过式 (2) 产生的新和声音调xi’还需进行音调微调:

对和声记忆库按以下更新策略进行更新:

上述过程不断重复, 直至创作 (迭代) 次数达到Tmax为止。

采用和声搜索算法, 将其应用于图像阈值分割算法寻取最优阈值, 提出一种基于和声搜索算法的图像阈值寻优算法 (Image Threshold Optimization Algorithm based on Harmony Search Algorithm, HS-OTSU) 。

2 和声搜索算法与OTSU算法结合原理
2.1 基于和声搜索算法的图像阈值寻优算法思想
HS-OTSU算法是将苹果图像编码处理后, 将图像作为目标函数进行阈值寻优, 选取图像的类间方差作为和声搜索算法的目标函数, 通过和声搜索算法寻找最大的分割阈值。

HS-OTSU算法思想:

Step1:读取苹果图像, 对原始图像进行预处理。

Step2:统计图像目标和背景的灰度级像素, 计算图像目标和背景两类间方差。

Step3:随机生成初始和声群体, 进行8位二进制编码, HMS=256为和声记忆库的大小。第i个和声记为xi= (x1, x2, Λ, xN) , 其中N=8为和声音调个数。

Step4:选取图像类间方差作为HS算法的目标函数f (x i) , 按式 (1) 初始化和声记忆库HM。

Step5:随机生成r1, r2, r∈[0 1, ], 如果r1<HCMR, 则新和声音调xi’按 (2) 式计算, 且如果r2<P AR, 则刚计算出的新和声音调xi’再按式 (3) 计算;如果r1≥≥H C M R, 则按式 (4) 计算新和声。按式 (5) 对和声记忆库进行更新, 如此反复迭代直至Tmax为止。

Step6:输出最优阈值。

2.2 基于和声搜索算法的图像阈值寻优算法流程
HS-OTSU算法流程图如图1所示。

⛄二、部分源代码

close all
clear all

% Se carga la imagen RGB o escala de grises
I1 = imread(‘Picture 148710088.jpg’);
I=rgb2gray(I1);
level = 3;

% Se obtienen los histogramas si la imagen es RGB uno por cada canal si es
% en escala de grises solamente un historgrama.
if size(I,3) == 1 %grayscale image
[n_countR, x_valueR] = imhist(I(:,:,1));
elseif size(I,3) == 3 %RGB image
%histograma para cada canal RGB
[n_countR, x_valueR] = imhist(I(:,:,1));
[n_countG, x_valueG] = imhist(I(:,:,2));
[n_countB, x_valueB] = imhist(I(:,:,3));
end
Nt = size(I,1) * size(I,2); %Cantidad total de pixeles en la imagen RENG X COL
%Lmax niveles de color a segmentar 0 - 256
Lmax = 256; %256 different maximum levels are considered in an image (i.e., 0 to 255)

% Distribucion de probabilidades de cada nivel de intensidad del histograma 0 - 256
for i = 1:Lmax
if size(I,3) == 1
%grayscale image
probR(i) = n_countR(i) / Nt;
elseif size(I,3) == 3
%RGB image
probR(i) = n_countR(i) / Nt;
probG(i) = n_countG(i) / Nt;
probB(i) = n_countB(i) / Nt;
end
end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Parametros del problema de segmentacion
N_PAR = level; %number of thresholds (number of levels-1) (dimensiones)
ndim = N_PAR;

%Parametros Harmony Search %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
MaxAttempt = 25000; % Max number of Attempt
% Initial parameter setting
HS_size = 50; %Length of solution vector
HMacceptRate = 0.95; %HM Accepting Rate
PArate = 0.5; %Pitch Adjusting rate

if size(I,3) == 1
%Imagen escala de grises
range = ones(ndim,2);
range(:,2) = range(:,2) * Lmax;

%initializa harmony memory
HM = zeros(HS_size,ndim);

%Imagen RGB
range = ones(ndim,2);
range(:,2) = range(:,2) * Lmax;
%IR
xR = zeros(HS_size,ndim);
%IG
xG = zeros(HS_size,ndim);
%IB
xB = zeros(HS_size,ndim);

end

C_Func = 0;
tic
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Generating Initial Solution Vector

%evalua x en la funcion objetivo
 %[HMbest, fitBestR] = fitnessIMG(I, HS_size, Lmax, level, HM, probR);

% C_Func = length(HMbest);

HMbest = Kapur(HS_size,level,HM,probR);

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Starting the Harmony Search
for count = 1:MaxAttempt,

    else
        % Harmony Memory Accepting Rate
        x(j) = HM(fix(HS_size * rand) + 1,j); %<--- prob HMCR
        if (rand <= PArate)
            % Pitch Adjusting in a given range
            pa = (range(j,2) - range(j,1)) / pa_range(j);
            x(j) = x(j) + pa * (rand - 0.5);
          end %% for j
% Evaluate the new solution
%evalua x en la funcion objetivo
x = fix(sort(x));
%evalua x en la funcion objetivo
%[fbest, fitBestR] = fitnessIMG(I, 1, Lmax, level, x, probR);
fbest = Kapur(1,level,x,probR);
C_Func = C_Func + 1;


% Find the best in the HS solution vector   
[HStemp, ii] = sort(HMbest, 'descend'); %Maximiza

% Updating the current solution if better
if fbest > HMbest(HS_size), %maximiza
    HM(HS_size, :) = x;
    HMbest(HS_size) = fbest;
end
solution = x;   % Record the solution
%Obtiene los mejores valores de cada attempt y los alamacena
[mm,ii] = max(HMbest); %maximiza
Fit_bests(count) = mm; %Mejores Fitness
HS_elem(count,:) = HM(ii,1:ndim-1); %Mejores Elementos de HM
HS_bestit = HM(ii,1:ndim-1); %Guarda el mejor HS
HS_bestF = mm; %Guarda el mejor fitness

% Output the results to screen
str=strcat(‘Best estimates: =’,num2str(HS_bestit));
str=strcat(str,’ fmin='); str=strcat(str,num2str(HS_bestF));

%Save the best values that will be chek in the stop criterion
if count == 1 || HS_bestF > HS_ant
HS_ant = HS_bestF;
cc = 0;
elseif HS_bestF == HS_ant
cc = cc + 1;
end

if cc > (MaxAttempt * 0.10)
break;
end

end %% for count (harmony search)
toc
%plot fitness
plot(Fit_bests)

%Prepare results to be show
gBestR = sort(HS_bestit);
Iout = imageGRAY(I,gBestR);
Iout2 = mat2gray(Iout);
%Show results

MEANR = mean(Fit_bests)     %Mean of fitness
PSNRV = PSNR(I, Iout)       %PSNR between original image I and the segmented image Iout
Fit_bests(count)            %Best fitness
%Show results on images
figure
subplot(121)
   imshow(I);title('原图')
subplot(122)
imshow(Iout);title('分割图')


%Plot the threshold values over the histogram
figure 
plot(probR)
hold on
vmax = max(probR);
for i = 1:ndim-1
    line([intensity(i), intensity(i)],[0 vmax],[1 1],'Color','r','Marker','.','LineStyle','-')
    %plot(lineas(i,:))
    hold on
end
 hold off

⛄三、运行结果

⛄四、matlab版本及参考文献

1 matlab版本
2014a

2 参考文献
[1] 刘立群,火久元.基于和声搜索算法的图像阈值寻优算法[J].洛阳理工学院学报(自然科学版). 2014,24(03)

3 备注
简介此部分摘自互联网，仅供参考，若侵权，联系删除