day38:opencv与传统机器学习

1、K均值

K均值是最简单的聚类方法之一，是一种无监督学习。K均值聚类的原理是通过指定种类数目对数据进行聚类，例如根据颜色将围棋棋盘上的棋子分成两类.

K 均值聚类算法主要分为以下4个步骤。

第一步：指定将数据聚类成看K 类，并随机生成K 个中心点;

第二步:  遍历所有数据，根据数据与中心的位置关系将每个数据归类到不同的中心;

第三步: 计算每个聚类的平均值，并将均值作为新的中心点;

第四步: 重复第二步和第三步，直到每个聚类中心点的坐标收敛 ，输出聚类结果。

void visionagin:: MyKmeanpoints()
{
	//颜色表
	vectorcolor_lut;
	color_lut.push_back(Scalar(0, 0, 255));
	color_lut.push_back(Scalar(0, 255,0));
	color_lut.push_back(Scalar(255,0,0));
	RNG rng(10086);
	Mat img(Size(500, 500), CV_8UC3,Scalar::all(255));//500*500白色画布
	int k_num = 3;
	int pt1_num = rng.uniform(20, 60);
	int pt2_num = rng.uniform(20, 60);
	int pt3_num = rng.uniform(20, 60);
	int pointsnum = pt1_num + pt2_num + pt3_num;
	Mat points(pointsnum, 1, CV_32FC2);
	//生成pointsnum个随机点
	int i = 0;
	for (; i < pt1_num; ++i)
	{
		Point2f p;
		p.x = rng.uniform(0, 100);
		p.y = rng.uniform(100, 200);
		points.at(i, 0) = p;
	}
	for (; i < pt1_num+pt2_num; ++i)
	{
		Point2f p;
		p.x = rng.uniform(0, 100);
		p.y = rng.uniform(400, 500);
		points.at(i, 0) = p;
	}
	for (; i < pt1_num+pt2_num+pt3_num; ++i)
	{
		Point2f p;
		p.x = rng.uniform(300, 400);
		p.y = rng.uniform(300, 400);
		points.at(i, 0) = p;
	}
	Mat labels;
	TermCriteria T = TermCriteria(TermCriteria::COUNT | TermCriteria::EPS, 10, 0.1);
	Mat centers;
	//kmean分类
	kmeans(points, k_num, labels, T, 3, KMEANS_PP_CENTERS,centers);
	//绘制随机点
	for (int j = 0; j (j, 0);
		Point2f temp = points.at(j, 0);
		circle(img, temp, 2, color_lut[index], -1);

	}
	//以各类中心点为圆心画圆
	for (int m = 0; m (m, 0);
		ce.y = centers.at(m, 1);

		circle(img, ce, 50, Scalar(0,200,0), 1);
	}
	imshow("result", img);
}

根据k 均值聚类可以实现基于像素值的 图像分割。 与坐标点聚类 相似， 图像分割时的聚类数 据是每个像素的像素值.

在下列程序中， 首先将每个像素的像素值整理成符合Kmeans() 函数处理的 行数据形式 之后根据需求 选择聚类种类数目.在聚类完成后，将不同类中的像素表示成不同的颜色， 最后在图像窗口中显示.

在该程序中 分别将原图像分割成3 类和5类， 通过结果可以看出， 在利用k 均值进行图像分割时，合适的聚类种类数目是一 项重要的数据参数.

void visionagin:: MyKmeanimg()
{
	Mat img = imread("C:\\Users\\86176\\Downloads\\visionimage\\women.jfif");
	if (!img.data)
	{
		cout << "read failed!" << endl;
	}
	//将图像转化成符合Kmeans输入的数据格式
	int pointscount = img.rows * img.cols;
	Mat imgchanged=img.reshape(3, pointscount);
	Mat data;
	imgchanged.convertTo(data, CV_32F);
	//进行kmeans分类
	Mat	labels;
	int k_num = 3;
	TermCriteria T = TermCriteria(TermCriteria::COUNT | TermCriteria::EPS, 10, 0.1);
	kmeans(data, k_num, labels, T, 3, KMEANS_PP_CENTERS);
	//着色分类结果
	vector color_lut;
	color_lut.push_back(Vec3b(0, 0, 255));
	color_lut.push_back(Vec3b( 0, 255,0));
	color_lut.push_back(Vec3b(255,0,0));
	color_lut.push_back(Vec3b(255, 2555, 0));
	Mat result = Mat::zeros(img.size(), img.type());
	for (int i = 0; i < img.rows; ++i)
	{
		for (int j = 0; j < img.cols; ++j)
		{
			int index = labels.at(i * img.cols + j,0);
			result.at(i, j) =color_lut[index];
		}
	}
	imshow("result", result);
}

 2.k近邻

 

 

 

 数据进行训练并保存：

void visionagin:: Myknearesttrain()
{
	Mat img = imread("C:\\Users\\86176\\Downloads\\visionimage\\digits.png");//该图片50行，每5行一个数字
	Mat gray;
	cvtColor(img, gray, COLOR_BGR2GRAY);
	//将图像分割成5000行 20*20
	Mat image = Mat::zeros(5000, 400, CV_8UC1);
	Mat labels = Mat::zeros(5000, 1, CV_8UC1);
	//Rect 为将每个单幅图数据拷贝给image的位置
	Rect singleimg;
	singleimg.x = 0;
	singleimg.height = 1;
	singleimg.width = 400;
	int index = 0;//第几个单幅图像的数据
	for (int i = 0; i < 50; i++)
	{
		int label = i / 5;
		int datay = i * 20;
		for (int j = 0; j < 100; j++)
		{
			int datax = j * 20;
			Mat singlenum = Mat::zeros(20, 20, CV_8UC1);
			//单个图片20*20 像素
			for (int m = 0; m < 20; m++)
			{
				for (int n = 0; n < 20; n++)
				{
					singlenum.at(m, n) = gray.at(datay + m, datax + n);
				}
			}
			//将单个图片转化为行数据
			Mat rowdata = singlenum.reshape(1, 1);//1*400
			singleimg.y = index;
			cout << "提取第" << index +1<< "个数据" << endl;
			//将每一幅图的数据 放到image
			rowdata.copyTo(image(singleimg));
			//记录每副图像的label;
			labels.at(index,0) = label;
			index++;
		}
	}
	imwrite("C:\\Users\\86176\\Downloads\\visionimage\\5000幅数字按行排列结果.png", image);
	imwrite("C:\\Users\\86176\\Downloads\\visionimage\\5000幅数字图像标签.png", labels);


	//加载训练模型,转换数据类型
	image.convertTo(image, CV_32F);
	labels.convertTo(labels, CV_32F);
	//构造训练类
	Ptr tdata = cv::ml::TrainData::create(image, ml::ROW_SAMPLE, labels);
	//进行训练
	// 
	//创建K近邻类
	Ptr knn = ml::KNearest::create();
	knn->setDefaultK(5);//每个类别拿出5个数据
	knn->setIsClassifier(true);//进行分类
	//开始训练
	knn->train(tdata);
	//保存训练结果
	knn->save("C:\\Users\\86176\\Downloads\\visionimage\\digits_train_model.yml");
	cout << "已完成训练！" << endl;

}

加载数据

void visionagin:: Myknearesttest()
{
	Mat data = imread("C:\\Users\\86176\\Downloads\\visionimage\\5000幅数字按行排列结果.png",IMREAD_ANYDEPTH);
	Mat labels = imread("C:\\Users\\86176\\Downloads\\visionimage\\5000幅数字图像标签.png",IMREAD_ANYDEPTH);
	data.convertTo(data, CV_32F);
	labels.convertTo(labels, CV_32S);
	//加载模型
	Ptr knn = Algorithm::load("C:\\Users\\86176\\Downloads\\visionimage\\digits_train_model.yml");
	//预测
	Mat result;
	knn->findNearest(data, 5, result);
	//统计预测结果与实际情况相同的数目
	int count = 0;
	for (int i = 0; i < result.rows; ++i)
	{
		int predict = result.at(i, 0);
		if (predict == labels.at(i, 0))
		{
			count++;
		}
	}
	float d = static_cast(count )/ result.rows;
	cout << "count is :" << count << endl;
	cout << "正确性：" << d << endl;
	//实例进行预测
	Mat testimg1 = imread("C:\\Users\\86176\\Downloads\\visionimage\\handWrite01.png", IMREAD_ANYDEPTH);
	Mat testimg2 = imread("C:\\Users\\86176\\Downloads\\visionimage\\handWrite02.png",IMREAD_ANYDEPTH);
	imshow("testimg1", testimg1);
	imshow("testimg2", testimg2);
	//建立一个存放test数据的矩阵
	Mat testdata = Mat::zeros(2, 400, CV_8UC1);
	resize(testimg1,testimg1,Size(20,20));
	resize(testimg2,testimg2,Size(20, 20));
	Rect roi;
	roi.x = 0;
	roi.height = 1;
	roi.width = 400;
	roi.y = 0;
	Mat onedata=testimg1.reshape(1, 1);
	Mat twodata=testimg2.reshape(1, 1);
	onedata.copyTo(testdata(roi));
	roi.y = 1;
	twodata.copyTo(testdata(roi));
	testdata.convertTo(testdata, CV_32F);
	
	Mat result2;
	knn->findNearest(testdata, 5, result2);
	cout << "handWrite01预测结果：" << result2.at(0,0) << endl;
	cout << "handWrite02预测结果：" << result2.at(1, 0) << endl;
}

 3.决策树：

void visionagin:: MyDtrees()
{
	Mat data = imread("C:\\Users\\86176\\Downloads\\visionimage\\5000幅数字按行排列结果.png", IMREAD_ANYDEPTH);
	Mat labels = imread("C:\\Users\\86176\\Downloads\\visionimage\\5000幅数字图像标签.png", IMREAD_ANYDEPTH);
	data.convertTo(data, CV_32F);
	labels.convertTo(labels, CV_32S);
	//构建决策树
	Ptr DTmodel = ml::DTrees::create();
	DTmodel->setMaxDepth(8);
	DTmodel->setCVFolds(0);
	//构建训练数据集
	Ptr dtraindata = ml::TrainData::create(data, ml::ROW_SAMPLE, labels);
	//开始训练
	DTmodel->train(dtraindata);
	//保存训练模型
	DTmodel->save("C:\\Users\\86176\\Downloads\\visionimage\\决策树训练模型.yml");
	//测试模型精度
	Mat result;
	DTmodel->predict(data, result);//进行预测
	int count = 0;
	for (int i = 0; i < result.rows; ++i)
	{
		int predict = result.at(i, 0);
		if (predict == labels.at(i, 0))
		{
			count++;
		}
	}
	float rate = 0;
	rate = static_cast(count) / result.rows;
	cout << "决策树模型的精度：" << rate << endl;

	//实例测试模型
	Mat testimg1 = imread("C:\\Users\\86176\\Downloads\\visionimage\\handWrite01.png", IMREAD_ANYDEPTH);
	Mat testimg2 = imread("C:\\Users\\86176\\Downloads\\visionimage\\handWrite02.png", IMREAD_ANYDEPTH);
	imshow("handWrite01", testimg1);
	imshow("handWrite02", testimg2);
	//缩小尺寸
	resize(testimg1, testimg1, Size(20, 20));
	resize(testimg2, testimg2, Size(20, 20));
	//构建输入矩阵
	Mat testdata = Mat::zeros(2, 400, CV_8UC1);
	Rect roi;
	roi.x = 0;
	roi.y = 0;
	roi.height = 1;
	roi.width = 400;
	Mat onedata = testimg1.reshape(1, 1);
	Mat twodata = testimg2.reshape(1, 1);
	onedata.copyTo(testdata(roi));
	roi.y = 1;
	twodata.copyTo(testdata(roi));
	testdata.convertTo(testdata,CV_32FC1);
	//预测
	Mat result2;
	DTmodel->predict(testdata, result2);
	cout << "handWrite01的预测结果是：" << result2.at(0, 0) << endl;
	cout << "handWrite02的预测结果是：" << result2.at(1, 0) << endl;
}

 4.随机森林

void visionagin:: MyRtrees()
{
	Mat data = imread("C:\\Users\\86176\\Downloads\\visionimage\\5000幅数字按行排列结果.png", IMREAD_ANYDEPTH);
	Mat labels = imread("C:\\Users\\86176\\Downloads\\visionimage\\5000幅数字图像标签.png", IMREAD_ANYDEPTH);
	data.convertTo(data, CV_32FC1);
	labels.convertTo(labels, CV_32SC1);
	Ptr rtreemodel = ml::RTrees::create();
	rtreemodel->setTermCriteria(TermCriteria(TermCriteria::MAX_ITER + TermCriteria::EPS, 100, 0.1));
	//构建训练集
	Ptr traindata = ml::TrainData::create(data, ml::ROW_SAMPLE, labels);
	rtreemodel->train(traindata);
	//训练模型保存
	rtreemodel->save("C:\\Users\\86176\\Downloads\\visionimage\\随机数树训练模型.yml");
	Mat result;
	rtreemodel->predict(data, result);
	//计算精度
	int count=0;
	for (int i = 0; i < result.rows; i++)
	{
		int predict = result.at(i, 0);
		if (predict == labels.at(i, 0))
		{
			count++;
		}
	}
	double rate = 0;
	rate = static_cast(count) / result.rows;
	cout << "rtrees的精度：" << rate << endl;
	//实例测试随机数模型
	Mat testimg1 = imread("C:\\Users\\86176\\Downloads\\visionimage\\handWrite01.png", IMREAD_ANYDEPTH);
	Mat testimg2 = imread("C:\\Users\\86176\\Downloads\\visionimage\\handWrite02.png", IMREAD_ANYDEPTH);
	imshow("handWrite01", testimg1);
	imshow("handWrite02", testimg2);
	//变换尺寸
	resize(testimg1, testimg1, Size(20, 20));
	resize(testimg2, testimg2, Size(20, 20));
	//建立输入矩阵
	Mat testdata = Mat::zeros(2, 400, CV_8UC1);
	Mat onedata = testimg1.reshape(1, 1);
	Mat twodata = testimg2.reshape(1, 1);
	Rect roi;
	roi.x = 0;
	roi.y = 0;
	roi.height = 1;
	roi.width = 400;
	onedata.copyTo(testdata(roi));
	roi.y = 1;
	twodata.copyTo(testdata(roi));
	testdata.convertTo(testdata, CV_32F);
	//预测
	Mat result2;
	rtreemodel->predict(testdata, result2);
	cout << "handWrite01的预测结果：" << result2.at(0, 0) << endl;
	cout << "handWrite02的预测结果：" << result2.at(1, 0) << endl;
}

5.支持向量机

 

 

void visionagin:: MySVM()
{
	FileStorage fread("C:\\Users\\86176\\Downloads\\visionimage\\point.yml", FileStorage::READ);
	Mat data,labels;
	fread["data"] >> data;
	fread["labls"] >> labels;
	fread.release();
	vectorcolor;
	color.push_back(Vec3b(0, 0, 255));
	color.push_back(Vec3b(0, 255, 0));
	Mat img = Mat(Size(600, 600), CV_8UC3, Scalar::all(255));
	Mat img2;
	img.copyTo(img2);
	for (int i = 0; i < labels.rows; ++i)
	{
		Point2f p;
		p.x = data.at(i, 0);
		p.y = data.at(i, 1);
		circle(img, p, 3, color[labels.at(i, 0)], -1);
		circle(img2, p, 3, color[labels.at(i, 0)], -1);
	}
	imshow("原数据", img);
	//data.convertTo(data, CV_32FC1);
	//labels.convertTo(labels, CV_32SC1);
	//构建svm模型
	Ptr svmmodel = ml::SVM::create();
	svmmodel->setKernel(ml::SVM::INTER);
	svmmodel->setType(ml::SVM::C_SVC);
	svmmodel->setTermCriteria(TermCriteria(TermCriteria::MAX_ITER + TermCriteria::EPS, 100, 0.01));
	//构建输入的数据
	Ptr traindata = ml::TrainData::create(data, ml::ROW_SAMPLE, labels);
	//训练
	svmmodel->train(traindata);
	//对每个点进行预测
	Mat temp = Mat::zeros(1, 2, CV_32FC1);
	for (int i = 0; i < img.rows; i+=2)
	{
		for (int j = 0; j < img.cols; j+=2)
		{
			temp.at(0) = (float)j;//x
			temp.at(1) = (float)i;//y
			int result;//单个数据结果
			result=svmmodel->predict(temp);
			img2.at(i,j) = color[result];
		}
	}
	imshow("result", img2);
}