OpenCV--SVM多分类问题

物体识别中经常遇到多分类器问题，svm是比较成熟和直接的想法。一般来说使用svm作为多分类器主要有以下思路：

1. 一对多(one-vs-all)。训练时依次将目标类别作为正样本，其余样本作为负样本，以此训练n个svm。这个在Andrew Ng的Machine leaning的课上介绍过。
   缺点：因为训练集是1：N的情况，存在较大的bias，不是特别实用。

2. 一对一(one-vs-one)。训练时，任意两类样本之间训练一个svm，则n类别，训练出(n-1)n/2个svm。在runtime时，对一个未知样本分类，则使用投票的法方法。libsvm即使用的该种方法。
   缺点：类别多的时候，(n-1)n/2个支持向量机，计算代价大。

3. 层次支持向量机。首先将所有类别分类为两个子类，再将子类进一步划分为两个子类，直到单独子类为止。好像一棵树耶。具体请参考：刘志刚, 李德仁, 秦前清, 等. 支持向量机在多类分类问题中的推广[J]. 2004.

4. DAG-SVMS。由Platt提出的决策导向的循环图DDAG导出的,是针对“一对一”SVMS存在误分、拒分现象提出的。请参考论文

简单示例

#include <opencv2/core.hpp>#include <opencv2/imgproc.hpp>#include "opencv2/imgcodecs.hpp"#include <opencv2/highgui.hpp>#include <opencv2/ml.hpp>using namespace cv;using namespace cv::ml;Vec3b getRandomColor(){    RNG rng(clock());    return Vec3b(rng.next() % 255, rng.next() % 255, rng.next() % 255);}int main(int, char**){    // Data for visual representation    int width = 512, height = 512;    Mat image = Mat::zeros(height, width, CV_8UC3);    // Set up training data    int labels[4] = {1, 2, 3, 4};    float trainingData[4][2] = { {100, 10}, {10, 500}, {500, 10}, {500, 500} };    Mat trainingDataMat(4, 2, CV_32FC1, trainingData);    Mat labelsMat(4, 1, CV_32SC1, labels);    // Train the SVM    //! [init]    Ptr<SVM> svm = SVM::create();    svm->setType(SVM::C_SVC);    svm->setKernel(SVM::POLY);    svm->setDegree(1.0);    svm->setTermCriteria(TermCriteria(TermCriteria::MAX_ITER, 100, 1e-6));    //! [init]    //! [train]//    svm->train(trainingDataMat, ROW_SAMPLE, labelsMat);//    Ptr<TrainData> auto_train_data = TrainData::create(trainingDataMat, ROW_SAMPLE, labelsMat);//    svm->trainAuto(auto_train_data);    svm->train(trainingDataMat, ROW_SAMPLE, labelsMat);    //! [train]    // Show the decision regions given by the SVM    //! [show]    Vec3b green(0,255,0), blue (255,0,0), red(0,0,255),yellow(0,255,255);    for (int i = 0; i < image.rows; ++i){        for (int j = 0; j < image.cols; ++j){            Mat sampleMat = (Mat_<float>(1,2) << j,i);            float response = svm->predict(sampleMat);            double ratio = 0.5;            if (response == 1)                image.at<Vec3b>(i,j)  = green*ratio;            else if (response == 2)                image.at<Vec3b>(i,j)  = blue*ratio;            else if(response == 3){                image.at<Vec3b>(i,j) = red*ratio;            }else if(response == 4){                image.at<Vec3b>(i,j) = yellow*ratio;            }        }    }    int thickness = -1;    int lineType = 8;    circle( image, Point(100, 10), 5, Scalar(  0,255,0), thickness, lineType );    circle( image, Point(10, 500), 5, Scalar(255,0,0), thickness, lineType );    circle( image, Point(500, 10), 5, Scalar(0,0,255), thickness, lineType );    circle( image, Point( 500, 500), 5, Scalar(0,255,255), thickness, lineType );    thickness = 2;    lineType  = 8;    Mat sv = svm->getSupportVectors();    std::cout << sv << std::endl;    for (int i = 0; i < sv.rows; ++i){        const float* v = sv.ptr<float>(i);        circle( image,  Point( (int) v[0], (int) v[1]), 6,  CV_RGB(128, 128, 128), 2);    }    imwrite("result.png", image);        // save the image    imshow("SVM Simple Example", image); // show it to the user    waitKey(0);}

效果：

注意点：
使用RBF核或者使用autotrain，参数选择十分重要。不行你试试哟！！！