在opencv3中利用SVM进行图像目标检测和分类

采用鼠标事件，手动选择样本点，包括目标样本和背景样本。组成训练数据进行训练

1、主函数

#include "stdafx.h"
#include "opencv2/opencv.hpp"
using namespace cv;
using namespace cv::ml;

Mat img,image;
Mat targetData, backData;
bool flag = true;
string wdname = "image";

void on_mouse(int event, int x, int y, int flags, void* ustc); //鼠标取样本点
void getTrainData(Mat &train_data, Mat &train_label);  //生成训练数据
void svm(); //svm分类

int main(int argc, char** argv)
{
    string path = "d:/peppers.png";
    img = imread(path);
    img.copyTo(image);
    if (img.empty())
    {
        cout << "Image load error";
        return ;
    }
    namedWindow(wdname);
    setMouseCallback(wdname, on_mouse, );

    for (;;)
    {
        imshow("image", img);

        int c = waitKey();
        if ((c & ) == )
        {
            cout << "Exiting ...\n";
            break;
        }
        if ((char)c == 'c')
        {
            flag = false;
        }
        if ((char)c == 'q')
        {
            destroyAllWindows();
            break;
        }
    }
    svm();
    return ;
}

首先输入图像，调用setMouseCallback函数进行鼠标取点

2、鼠标事件

//鼠标在图像上取样本点，按q键退出
void on_mouse(int event, int x, int y, int flags, void* ustc)
{
    if (event == CV_EVENT_LBUTTONDOWN)
    {
        Point pt = Point(x, y);
        Vec3b point = img.at<Vec3b>(y, x);  //取出该坐标处的像素值，注意x,y的顺序
        Mat tmp = (Mat_<float>(, ) << point[], point[], point[]);
        if (flag)
        {
            targetData.push_back(tmp); //加入正样本矩阵
            circle(img, pt, , Scalar(, , ), -, ); //画圆，在图上显示点击的点 

        }

        else
        {
            backData.push_back(tmp); //加入负样本矩阵
            circle(img, pt, , Scalar(, , ), -, ); 

        }
        imshow(wdname, img);
    }
}

用鼠标在图像上点击，取出当前点的红绿蓝像素值进行训练。先选择任意个目标样本，然后按"c“键后选择任意个背景样本。样本数可以自己随意决定。样本选择完后，按”q"键完成样本选择。

3、svm分类

void getTrainData(Mat &train_data, Mat &train_label)
{
    int m = targetData.rows;
    int n = backData.rows;
    cout << "正样本数：:" << m << endl;
    cout << "负样本数：" << n << endl;
    vconcat(targetData, backData, train_data); //合并所有的样本点，作为训练数据
    train_label = Mat(m + n, , CV_32S, Scalar::all()); //初始化标注
    for (int i = m; i < m + n; i++)
        train_label.at<int>(i, ) = -;
}

void svm()
{
    Mat train_data, train_label;
    getTrainData(train_data, train_label); //获取鼠标选择的样本训练数据

    // 设置参数
    Ptr<SVM> svm = SVM::create();
    svm->setType(SVM::C_SVC);
    svm->setKernel(SVM::LINEAR);

    // 训练分类器
    Ptr<TrainData> tData = TrainData::create(train_data, ROW_SAMPLE, train_label);
    svm->train(tData);

    Vec3b color(, , );
    // Show the decision regions given by the SVM
    for (int i = ; i < image.rows; ++i)
    for (int j = ; j < image.cols; ++j)
    {
        Vec3b point = img.at<Vec3b>(i, j);  //取出该坐标处的像素值
        Mat sampleMat = (Mat_<float>(, ) << point[], point[], point[]);
        float response = svm->predict(sampleMat);  //进行预测，返回1或-1,返回类型为float
        if ((int)response != )
            image.at<Vec3b>(i, j) = color;  //将背景点设为黑色
    }

    imshow("SVM Simple Example", image); // show it to the user
    waitKey();
}

将正负样本矩阵，用vconcat合并成一个矩阵，用作训练分类器，并对相应的样本进行标注。最后将识别出的目标保留，将背景部分调成黑色。

4、完整程序

// svm.cpp : 定义控制台应用程序的入口点。
//

#include "stdafx.h"
#include "opencv2/opencv.hpp"
using namespace cv;
using namespace cv::ml;

Mat img,image;
Mat targetData, backData;
bool flag = true;
string wdname = "image";

void on_mouse(int event, int x, int y, int flags, void* ustc); //鼠标取样本点
void getTrainData(Mat &train_data, Mat &train_label);  //生成训练数据
void svm(); //svm分类

int main(int argc, char** argv)
{
    string path = "d:/peppers.png";
    img = imread(path);
    img.copyTo(image);
    if (img.empty())
    {
        cout << "Image load error";
        return ;
    }
    namedWindow(wdname);
    setMouseCallback(wdname, on_mouse, );

    for (;;)
    {
        imshow("image", img);

        int c = waitKey();
        if ((c & ) == )
        {
            cout << "Exiting ...\n";
            break;
        }
        if ((char)c == 'c')
        {
            flag = false;
        }
        if ((char)c == 'q')
        {
            destroyAllWindows();
            break;
        }
    }
    svm();
    return ;
}

//鼠标在图像上取样本点，按q键退出
void on_mouse(int event, int x, int y, int flags, void* ustc)
{
    if (event == CV_EVENT_LBUTTONDOWN)
    {
        Point pt = Point(x, y);
        Vec3b point = img.at<Vec3b>(y, x);  //取出该坐标处的像素值，注意x,y的顺序
        Mat tmp = (Mat_<float>(, ) << point[], point[], point[]);
        if (flag)
        {
            targetData.push_back(tmp); //加入正样本矩阵
            circle(img, pt, , Scalar(, , ), -, ); //画出点击的点 

        }

        else
        {
            backData.push_back(tmp); //加入负样本矩阵
            circle(img, pt, , Scalar(, , ), -, ); 

        }
        imshow(wdname, img);
    }
}

void getTrainData(Mat &train_data, Mat &train_label)
{
    int m = targetData.rows;
    int n = backData.rows;
    cout << "正样本数：:" << m << endl;
    cout << "负样本数：" << n << endl;
    vconcat(targetData, backData, train_data); //合并所有的样本点，作为训练数据
    train_label = Mat(m + n, , CV_32S, Scalar::all()); //初始化标注
    for (int i = m; i < m + n; i++)
        train_label.at<int>(i, ) = -;
}

void svm()
{
    Mat train_data, train_label;
    getTrainData(train_data, train_label); //获取鼠标选择的样本训练数据

    // 设置参数
    Ptr<SVM> svm = SVM::create();
    svm->setType(SVM::C_SVC);
    svm->setKernel(SVM::LINEAR);

    // 训练分类器
    Ptr<TrainData> tData = TrainData::create(train_data, ROW_SAMPLE, train_label);
    svm->train(tData);

    Vec3b color(, , );
    // Show the decision regions given by the SVM
    for (int i = ; i < image.rows; ++i)
    for (int j = ; j < image.cols; ++j)
    {
        Vec3b point = img.at<Vec3b>(i, j);  //取出该坐标处的像素值
        Mat sampleMat = (Mat_<float>(, ) << point[], point[], point[]);
        float response = svm->predict(sampleMat);  //进行预测，返回1或-1,返回类型为float
        if ((int)response != )
            image.at<Vec3b>(i, j) = color;  //将背景设置为黑色
    }

    imshow("SVM Simple Example", image);
    waitKey();
}

输入原图像：

程序运行后显示：