SVM+Surf+K-means花朵分类(Matlab)

详细算法、步骤介绍见python版：
https://editor.csdn.net/md/?articleId=103732613

训练文件

clc
clear all
train_path = "g:/flowers/train/";
test_path = "g:/flowers/test/";
K = 100;
build_centers(train_path,K);
%构建训练集特征向量
[x_train,y_train] = cal_vec(train_path,K);
%将特征向量和标签输入到SVM分类器中
Train(x_train,y_train);
%计算测试集的正确率
[predicted_label, accuracy, decision_values] = Test(test_path,K);

function f1 = calcSurfFeature(img)
%将图像转化为灰度图像
gray = rgb2gray(img);
%计算surf特征点
points = detectSURFFeatures(gray);
%计算描述向量
[f1, vpts1] = extractFeatures(gray, points);
end

%%计算词袋
function centers = learnVocabulary(features,K)
%设置词袋数量
% use k-means to cluster a bag of features
%[聚类编号，点的位置] = [输入，聚类个数，初始中心点，聚类重复次数]
opts = statset('Display','final','MaxIter',1000);
[idx ,centers] = kmeans(features ,K,'Replicates',10, 'Options',opts);
end

%%根据字典计算图片的特征向量
function featVec = calcFeatVec(features, centers,K)
featVec = zeros(1,K);
%m表示特征点的个数，n表示特征点的维度
[m,n] = size(features);
for i = 1:m
    fi = features(i,:);
    %%特征是1*64，而字典是K*64，需要求该特征与哪个词袋最近
    res = zeros(K,64);
    for j = 1:K
        res(j,:) = fi;
    end
    %将该特征纵向复制变成K*64
    %求出距离
    distance = sum((res - centers).^2,2);
    distance = sqrt(distance);
    %将距离排序，求得最小的距离
    [x,y] = min(distance);
    featVec(y) = featVec(y)+1;
end
end

%%%建立字典
function build_centers(path,K)
%求得路径下子目录，将子目录路径保存到cate中
sub = dir(path);
cate = [];
features = [];
for i = 1:size(sub)
    if sub(i).name ~= "." && sub(i).name ~= ".." && sub(i).isdir
        s = strcat(strcat(path ,sub(i).name),"/");
        cate = [cate s];
    end
end
for i = 1:length(cate)
    imgDir  = dir( strcat(cate(i),'*.jpg'));
    for j = 1:length(imgDir)
        img = imread(strcat(cate(i),imgDir(j).name));
        %计算surf特征
        img_f = calcSurfFeature(img);
        %特征点加入特征集合
        features = [features;img_f];
    end
end
[m,n] = size(features);
fprintf("训练集特征点集合:[ %d , %d ]\n",m,n);
centers = learnVocabulary(features,K);
filename = "g:/flowers/svm/svm_centers.mat";
save(filename,'centers');
end

%%根据centers计算图片特征向量
function [data_vec,labels] = cal_vec(path,K)
load("g:/flowers/svm/svm_centers.mat");
%存放图片特征向量
data_vec = [];
%存放图片标签（所属类别）
labels = [];
%得到当前目录下的子目录
sub = dir(path);
cate = [];
for i = 1:size(sub)
    if sub(i).name ~= "." && sub(i).name ~= ".." && sub(i).isdir
        s = strcat(strcat(path ,sub(i).name),"/");
        cate = [cate s];
    end
end
for i = 1:length(cate)
    imgDir  = dir( strcat(cate(i),'*.jpg'));
    for j = 1:length(imgDir)
        img = imread(strcat(cate(i),imgDir(j).name));
        %计算surf特征
        img_f = calcSurfFeature(img);
        img_vec = calcFeatVec(img_f, centers,K);
        data_vec = [data_vec;img_vec];
        labels = [labels,i];
    end
end
end

%训练SVM分类器
function Train(data_vec,labels)
modle = libsvm_train(labels',data_vec,'-t 0');
save("g:/flowers/svm/svm_model.mat",'modle');
fprintf("Training Done!\n");
end

%SVM测试测试集正确率
function [predicted_label, accuracy, decision_values] = Test(path,K)
%读取SVM模型
load("g:/flowers/svm/svm_model.mat");
%读取字典
load("g:/flowers/svm/svm_centers.mat");
%计算每张图片的特征向量
[data_vec,labels] = cal_vec(path,K);
[num_test , y] = size(data_vec);
[predicted_label, accuracy, decision_values] = libsvm_predict(labels',data_vec,modle);
fprintf("Testing Done!\n");
end

预测文件

clc
clear all
img_path = "e:\minist\predict\";
class_flower = ["0","1","2","3","4","5","6","7","8","9"];
K = 100;
for i = 1:5
    path = strcat(strcat(img_path,num2str(i)),".jpg");
    img = imread(path);
    figure(i);
    imshow(img);
    res = predict(img,K);
    xlabel(["预测结果：",class_flower(res)]);
end

%预测单张图片
function res = predict(img,K)
load("g:/flowers/svm/svm_centers.mat");
load("g:/flowers/svm/svm_model.mat");
features = calcSurfFeature(img);
featVec = calcFeatVec(features, centers,K);
res = libsvm_predict([1],featVec,modle);
end

function f1 = calcSurfFeature(img)
%将图像转化为灰度图像
gray = rgb2gray(img);
%计算suft特征点
points = detectSURFFeatures(gray);
%计算描述向量
[f1, vpts1] = extractFeatures(gray, points);
end

%%根据字典计算图片的特征向量
function featVec = calcFeatVec(features, centers,K)
featVec = zeros(1,K);
%m表示特征点的个数，n表示特征点的维度
[m,n] = size(features);
for i = 1:m
    fi = features(i,:);
    %%特征是1*64，而字典是50*64，需要求该特征与哪个词袋最近
    res = zeros(K,64);
    for j = 1:K
        res(j,:) = fi;
    end
    %将该特征纵向复制变成K*64
    %求出距离
    distance = sum((res - centers).^2,2);
    distance = sqrt(distance);
    %将距离排序，求得最小的距离
    [x,y] = min(distance);
    featVec(y) = featVec(y)+1;
end
end