自己写的demo记个笔记用
替换掉图片路径和保存路径svm训练的模型路径就可以跑
效果我觉的不行，目前也不知到如何优化、希望有大佬可以给点建议
流程
处理超像素
选择超像素
提取HOG、颜色直方图、LBP直方图特征
训练
预测

#include <iostream>
#include <opencv2/opencv.hpp>
#include <opencv2/ximgproc.hpp>
#include <opencv2/ml.hpp>using namespace cv;
using namespace std;
using namespace cv::ml;
Mat g_LabelSlic;  
Mat g_MaskSlic;
int g_NumSuperPixels;//目标图像和标签
multimap<int, Mat>g_mapImgSuperPixelsOfTarget;
multimap<int, Mat>g_mapImgSuperPixelsOfNonTarget;struct MouseCallbackData {Mat img;      // 原始图像Mat imgClone; // 克隆图像
};Mat GetSuperPixelsByLabel(const Mat& img, int superpixelID)
{int minX = img.cols, minY = img.rows, maxX = 0, maxY = 0;for (int y = 0; y < img.rows; y++) {for (int x = 0; x < img.cols; x++) {if (g_LabelSlic.at<int>(y, x) == superpixelID) {// 更新边界框坐标if (x < minX) minX = x;if (y < minY) minY = y;if (x > maxX) maxX = x;if (y > maxY) maxY = y;}}}// 确保边界框有效if (minX > maxX || minY > maxY) {cout << "未找到有效的超像素！" << endl;return Mat();  }Rect superPixelBoundingBox(minX, minY, maxX - minX + 1, maxY - minY + 1);Mat croppedRegion = img(superPixelBoundingBox).clone();  for (int y = 0; y < croppedRegion.rows; y++) {for (int x = 0; x < croppedRegion.cols; x++) {int origX = x + minX;int origY = y + minY;if (g_LabelSlic.at<int>(origY, origX) != superpixelID) {croppedRegion.at<Vec3b>(y, x) = Vec3b(0, 0, 0); }}}return croppedRegion; 
}vector<float> GetHOGDescriptor(Mat img)
{if (img.empty()) {std::cerr << "输入图像为空！" << std::endl;return {};}resize(img, img, Size(64, 64));Mat imgGray;cvtColor(img, imgGray, COLOR_BGR2GRAY);HOGDescriptor hog(Size(32, 32),  // 图像窗口大小Size(8, 8),   // 块大小Size(4, 4),     // 块步长Size(4, 4),     // cell 大小9               // 梯度方向 bins 数);vector<float> descriptors;hog.compute(imgGray, descriptors);return descriptors;
}// 提取 LBP 特征及其直方图
void ExtractLBPFeatures(Mat img, Mat& lbp, Mat& lbpHist)
{resize(img, img, Size(64, 64));// 转换为灰度图Mat gray;if (img.channels() == 3) {cvtColor(img, gray, COLOR_BGR2GRAY);}else {gray = img.clone();}// 初始化 LBP 特征矩阵lbp = Mat(gray.size(), CV_8UC1, Scalar(0));for (int y = 1; y < gray.rows - 1; y++) {for (int x = 1; x < gray.cols - 1; x++) {uchar center = gray.at<uchar>(y, x);uchar code = 0;code |= (gray.at<uchar>(y - 1, x - 1) > center) << 7; // 128code |= (gray.at<uchar>(y - 1, x) > center) << 6;     // 64code |= (gray.at<uchar>(y - 1, x + 1) > center) << 5; // 32code |= (gray.at<uchar>(y, x + 1) > center) << 4;     // 16code |= (gray.at<uchar>(y + 1, x + 1) > center) << 3; // 8code |= (gray.at<uchar>(y + 1, x) > center) << 2;     // 4code |= (gray.at<uchar>(y + 1, x - 1) > center) << 1; // 2code |= (gray.at<uchar>(y, x - 1) > center);           // 1lbp.at<uchar>(y, x) = code; // 将计算的 LBP 值存储}}// 计算 LBP 直方图const int histSize = 256; // LBP 特征值的范围const float range[] = { 0, 256 };const float* histRange = { range };calcHist(&lbp, 1, 0, Mat(), lbpHist, 1, &histSize, &histRange);normalize(lbpHist, lbpHist);
}Mat ExtractHSVHistogram(Mat img)
{resize(img, img, Size(64, 64));Mat hsv_img;cvtColor(img, hsv_img, cv::COLOR_BGR2HSV); vector<cv::Mat> hsv_planes;split(hsv_img, hsv_planes);  // 分割 H, S, V 通道int histSize = 256;  // 直方图分为 256 个区间float h_range[] = { 0, 180 };  // H 通道范围是 0-180float s_v_range[] = { 0, 256 };  // S 和 V 通道范围是 0-256const float* h_histRange = { h_range };const float* sv_histRange = { s_v_range };Mat h_hist, s_hist, v_hist;calcHist(&hsv_planes[0], 1, 0, cv::Mat(), h_hist, 1, &histSize, &h_histRange, true, false);  // H 通道直方图calcHist(&hsv_planes[1], 1, 0, cv::Mat(), s_hist, 1, &histSize, &sv_histRange, true, false);  // S 通道直方图calcHist(&hsv_planes[2], 1, 0, cv::Mat(), v_hist, 1, &histSize, &sv_histRange, true, false);  // V 通道直方图// 合并 H, S, V 直方图Mat hist;hconcat(h_hist, s_hist, hist);hconcat(hist, v_hist, hist);Mat normalizedHist;normalize(hist, normalizedHist);return normalizedHist;  // 返回归一化后的直方图
}void OnMouse(int event, int x, int y, int flags, void* param)
{MouseCallbackData* data = static_cast<MouseCallbackData*>(param);Mat& img = data->img;Mat& imgClone = data->imgClone;int mouseButtonClicked = 0;if (event == EVENT_LBUTTONDOWN) {mouseButtonClicked = 1;if (x >= 0 && x < img.cols && y >= 0 && y < img.rows) {int superpixelID = g_LabelSlic.at<int>(y, x);cout << "点击目标超像素ID: " << superpixelID << endl;g_mapImgSuperPixelsOfTarget.insert({ mouseButtonClicked, GetSuperPixelsByLabel(img, superpixelID) });circle(imgClone, Point(x, y), 3, Scalar(0, 255, 0), -1);imshow("Imageview", imgClone);}}else if (event == EVENT_RBUTTONDOWN) {mouseButtonClicked = 2;if (x >= 0 && x < img.cols && y >= 0 && y < img.rows) {int superpixelID = g_LabelSlic.at<int>(y, x);cout << "点击非目标超像素ID: " << superpixelID << endl;g_mapImgSuperPixelsOfNonTarget.insert({ mouseButtonClicked, GetSuperPixelsByLabel(img, superpixelID) });circle(imgClone, Point(x, y), 3, Scalar(0, 0, 255), -1);imshow("Imageview", imgClone);}}
}void SvmClassifier(multimap<int, vector<float>> HOGDescriptorOFTarget,multimap<int, Mat> lbpHistOFTarget,multimap<int, Mat> hsvHistOFTarget,multimap<int, vector<float>> HOGDescriptorOFNonTarget,multimap<int, Mat> lbpHistOFNonTarget,multimap<int, Mat> hsvHistOFNonTarget)
{cout << "star svm model train ..." << endl;Mat featureList;Mat labels;for (const auto& pair : HOGDescriptorOFTarget) {Mat hogMat(pair.second, CV_32F);normalize(hogMat, hogMat, 0, 1, NORM_MINMAX);hogMat = hogMat.reshape(1, 1);Mat lbpHist = lbpHistOFTarget.find(pair.first)->second;lbpHist.convertTo(lbpHist, CV_32F);lbpHist = lbpHist.reshape(1, 1);  // 展平Mat hsvHist = hsvHistOFTarget.find(pair.first)->second;hsvHist.convertTo(hsvHist, CV_32F);hsvHist = hsvHist.reshape(1, 1);  // 展平Mat combinedFeature;hconcat(hogMat, lbpHist, combinedFeature);  hconcat(combinedFeature, hsvHist, combinedFeature);  featureList.push_back(combinedFeature);labels.push_back(1); }for (const auto& pair : HOGDescriptorOFNonTarget) {Mat hogMat(pair.second, CV_32F);normalize(hogMat, hogMat, 0, 1, NORM_MINMAX); hogMat = hogMat.reshape(1, 1);Mat lbpHist = lbpHistOFNonTarget.find(pair.first)->second;lbpHist.convertTo(lbpHist, CV_32F);lbpHist = lbpHist.reshape(1, 1); Mat hsvHist = hsvHistOFNonTarget.find(pair.first)->second;hsvHist.convertTo(hsvHist, CV_32F);hsvHist = hsvHist.reshape(1, 1); Mat combinedFeature;hconcat(hogMat, lbpHist, combinedFeature);hconcat(combinedFeature, hsvHist, combinedFeature);featureList.push_back(combinedFeature);labels.push_back(0);  }Mat trainingData;vconcat(featureList, trainingData); //Mat labelsMat;// = Mat(labels).reshape(1, 1);Mat labelsMat = labels;trainingData.convertTo(trainingData, CV_32F);labelsMat.convertTo(labelsMat, CV_32S);Ptr<SVM> svm = SVM::create();svm->setKernel(SVM::RBF);svm->setType(SVM::C_SVC);svm->trainAuto(trainingData,ROW_SAMPLE,labelsMat,10);/*svm->setC(1.5);svm->setGamma(0.5);svm->setTermCriteria(cv::TermCriteria(cv::TermCriteria::MAX_ITER, 200, 1e-6));svm->train(trainingData, ROW_SAMPLE, labelsMat);*/svm->save("C:/Users/svs/Desktop/svm_model2.xml"); // 保存训练好的模型
}void DrawSuperpixelRegion(Mat& img, int targetLabel) {for (int y = 0; y < img.rows; y++) {for (int x = 0; x < img.cols; x++) {if (g_LabelSlic.at<int>(y, x) == targetLabel) {img.at<Vec3b>(y, x) = Vec3b(0, 255, 0); }}}
}int main() {Mat img = imread("C:/Users/svs/Desktop/test.jpeg");if (img.empty()) {cerr << "无法读取图片！" << endl;return -1;}//图片进行超像素分割int region_size = 50;float ruler = 20.0;int num_iterations = 100;Ptr<ximgproc::SuperpixelSLIC> slic = ximgproc::createSuperpixelSLIC(img, ximgproc::SLICO,region_size, ruler);slic->iterate(num_iterations);slic->getLabels(g_LabelSlic);slic->getLabelContourMask(g_MaskSlic);g_NumSuperPixels = slic->getNumberOfSuperpixels();Mat imgWithContours;img.copyTo(imgWithContours, ~g_MaskSlic);imshow("Superpixel Contours", imgWithContours);cout << "请点击选择一个超像素区域...\n";namedWindow("Imageview", WINDOW_AUTOSIZE);imshow("Image", img);Mat imgClone = img.clone();imshow("Imageview", imgClone);MouseCallbackData data;data.img = img;data.imgClone = imgClone;setMouseCallback("Imageview", OnMouse, &data);// 等待用户按 'q' 退出while (true) {const char key = waitKey(0);if (key == 'q') {break;}}/*提取获取图片的特征*/multimap<int, vector<float> >HOGDescriptorOFTarget;multimap<int, Mat >lbpHistOFTarget;multimap<int, Mat >hsvHistOFTarget;multimap<int, vector<float> >HOGDescriptorOFNonTarget;multimap<int, Mat>lbpHistOFNonTarget;multimap<int, Mat>hisHistOFNonTarget;for (const auto& pair : g_mapImgSuperPixelsOfTarget) {vector<float> vectorHOGDescriptor = GetHOGDescriptor(pair.second);HOGDescriptorOFTarget.insert({ pair.first, vectorHOGDescriptor});Mat lbpImg, lbpHist;ExtractLBPFeatures(pair.second, lbpImg, lbpHist);lbpHistOFTarget.insert({ pair.first, lbpHist });Mat hsvHist;hsvHist = ExtractHSVHistogram(pair.second);hsvHistOFTarget.insert({ pair.first,hsvHist });/*imshow("true", pair.second);waitKey(0);*/}for (const auto& pair : g_mapImgSuperPixelsOfNonTarget){vector<float> vectorHOGDescriptor = GetHOGDescriptor(pair.second);HOGDescriptorOFNonTarget.insert({ pair.first,vectorHOGDescriptor });Mat lbpImg, lbpHist;ExtractLBPFeatures(pair.second, lbpImg, lbpHist);lbpHistOFNonTarget.insert({ pair.first, lbpHist });Mat hsvHist;hsvHist = ExtractHSVHistogram(pair.second);hisHistOFNonTarget.insert({ pair.first,hsvHist});/* imshow("false", pair.second);waitKey(0);*/}SvmClassifier(HOGDescriptorOFTarget, lbpHistOFTarget, hsvHistOFTarget,HOGDescriptorOFNonTarget,lbpHistOFNonTarget, hisHistOFNonTarget);Ptr<SVM> svm = SVM::load("C:/Users/svs/Desktop/svm_model2.xml");if (svm.empty()) {std::cerr << "模型加载失败！\n";return -1;}//进行预测cout << "star predict ...\n";multimap<int, Mat> testSuperPixelsImg;vector<float> testHOGDescriptor;Mat testLbpHist;Mat testHsvHist;// 遍历每个超像素块for (int superpixelID = 0; superpixelID < g_NumSuperPixels; ++superpixelID) {Mat superpixelRegion = GetSuperPixelsByLabel(img, superpixelID);testSuperPixelsImg.insert({ superpixelID,  superpixelRegion });}for (const auto& pairs : testSuperPixelsImg) {/*imshow("test", pairs.second);waitKey(0);*/testHOGDescriptor = GetHOGDescriptor(pairs.second);Mat lbpImg;ExtractLBPFeatures(pairs.second, lbpImg,testLbpHist);testHsvHist = ExtractHSVHistogram(pairs.second);if (testHOGDescriptor.empty()){cerr << "HOG特征为空，超像素ID: \n";}if (testLbpHist.empty()){cerr << "LBP特征为空，超像素ID: \n";}if(testHsvHist.empty()){cerr << "Hsv直方图为空，超像素ID:\n";}Mat testHogMat(testHOGDescriptor, CV_32F);normalize(testHogMat, testHogMat, 0, 1, NORM_MINMAX);testHogMat = testHogMat.reshape(1, 1);testLbpHist.convertTo(testLbpHist, CV_32F);testLbpHist = testLbpHist.reshape(1, 1);  testHsvHist.convertTo(testHsvHist, CV_32F);testHsvHist = testHsvHist.reshape(1, 1);  Mat combinedFeature;hconcat(testHogMat, testLbpHist, combinedFeature);hconcat(combinedFeature, testHsvHist, combinedFeature);// 进行预测float response;response = svm->predict(combinedFeature);if (response == 1) {}else {DrawSuperpixelRegion(img, pairs.first);}}imshow("Image with Green Superpixel", img);cout << "predict successful\n";waitKey(0);destroyAllWindows();return 0;
}