1 #include <iostream> 2 #include <math.h> 3 #include <string> 4 #include "cv.h" 5 #include "ml.h" 6 #include "highgui.h" 7 8 using namespace cv; 9 using namespace std; 10 11 bool plotSupportVectors=true; 12 int numTrainingPoints=200; 13 int numTestPoints=2000; 14 int size=200; 15 int eq=0; 16 17 // accuracy 18 float evaluate(cv::Mat& predicted, cv::Mat& actual) { 19 assert(predicted.rows == actual.rows); 20 int t = 0; 21 int f = 0; 22 for(int i = 0; i < actual.rows; i++) { 23 float p = predicted.at<float>(i,0); 24 float a = actual.at<float>(i,0); 25 if((p >= 0.0 && a >= 0.0) || (p <= 0.0 && a <= 0.0)) { 26 t++; 27 } else { 28 f++; 29 } 30 } 31 return (t * 1.0) / (t + f); 32 } 33 34 // plot data and class 35 void plot_binary(cv::Mat& data, cv::Mat& classes, string name) { 36 cv::Mat plot(size, size, CV_8UC3); 37 plot.setTo(cv::Scalar(255.0,255.0,255.0)); 38 for(int i = 0; i < data.rows; i++) { 39 40 float x = data.at<float>(i,0) * size; 41 float y = data.at<float>(i,1) * size; 42 43 if(classes.at<float>(i, 0) > 0) { 44 cv::circle(plot, Point(x,y), 2, CV_RGB(255,0,0),1); 45 } else { 46 cv::circle(plot, Point(x,y), 2, CV_RGB(0,255,0),1); 47 } 48 } 49 cv::imshow(name, plot); 50 } 51 52 // function to learn 53 int f(float x, float y, int equation) { 54 switch(equation) { 55 case 0: 56 return y > sin(x*10) ? -1 : 1; 57 break; 58 case 1: 59 return y > cos(x * 10) ? -1 : 1; 60 break; 61 case 2: 62 return y > 2*x ? -1 : 1; 63 break; 64 case 3: 65 return y > tan(x*10) ? -1 : 1; 66 break; 67 default: 68 return y > cos(x*10) ? -1 : 1; 69 } 70 } 71 72 // label data with equation 73 cv::Mat labelData(cv::Mat points, int equation) { 74 cv::Mat labels(points.rows, 1, CV_32FC1); 75 for(int i = 0; i < points.rows; i++) { 76 float x = points.at<float>(i,0); 77 float y = points.at<float>(i,1); 78 labels.at<float>(i, 0) = f(x, y, equation); 79 } 80 return labels; 81 } 82 83 void svm(cv::Mat& trainingData, cv::Mat& trainingClasses, cv::Mat& testData, cv::Mat& testClasses) { 84 CvSVMParams param = CvSVMParams(); 85 86 param.svm_type = CvSVM::C_SVC; 87 param.kernel_type = CvSVM::RBF; //CvSVM::RBF, CvSVM::LINEAR ... 88 param.degree = 0; // for poly 89 param.gamma = 20; // for poly/rbf/sigmoid 90 param.coef0 = 0; // for poly/sigmoid 91 92 param.C = 7; // for CV_SVM_C_SVC, CV_SVM_EPS_SVR and CV_SVM_NU_SVR 93 param.nu = 0.0; // for CV_SVM_NU_SVC, CV_SVM_ONE_CLASS, and CV_SVM_NU_SVR 94 param.p = 0.0; // for CV_SVM_EPS_SVR 95 96 param.class_weights = NULL; // for CV_SVM_C_SVC 97 param.term_crit.type = CV_TERMCRIT_ITER +CV_TERMCRIT_EPS; 98 param.term_crit.max_iter = 1000; 99 param.term_crit.epsilon = 1e-6; 100 101 // SVM training (use train auto for OpenCV>=2.0) 102 CvSVM svm(trainingData, trainingClasses, cv::Mat(), cv::Mat(), param); 103 104 cv::Mat predicted(testClasses.rows, 1, CV_32F); 105 106 for(int i = 0; i < testData.rows; i++) { 107 cv::Mat sample = testData.row(i); 108 109 float x = sample.at<float>(0,0); 110 float y = sample.at<float>(0,1); 111 112 predicted.at<float>(i, 0) = svm.predict(sample); 113 } 114 115 cout << "Accuracy_{SVM} = " << evaluate(predicted, testClasses) << endl; 116 plot_binary(testData, predicted, "Predictions SVM"); 117 118 // plot support vectors 119 if(plotSupportVectors) { 120 cv::Mat plot_sv(size, size, CV_8UC3); 121 plot_sv.setTo(cv::Scalar(255.0,255.0,255.0)); 122 123 int svec_count = svm.get_support_vector_count(); 124 for(int vecNum = 0; vecNum < svec_count; vecNum++) { 125 const float* vec = svm.get_support_vector(vecNum); 126 cv::circle(plot_sv, Point(vec[0]*size, vec[1]*size), 3 , CV_RGB(0, 0, 0)); 127 } 128 cv::imshow("Support Vectors", plot_sv); 129 } 130 } 131 132 void mlp(cv::Mat& trainingData, cv::Mat& trainingClasses, cv::Mat& testData, cv::Mat& testClasses) { 133 134 cv::Mat layers = cv::Mat(4, 1, CV_32SC1); 135 136 layers.row(0) = cv::Scalar(2); 137 layers.row(1) = cv::Scalar(10); 138 layers.row(2) = cv::Scalar(15); 139 layers.row(3) = cv::Scalar(1); 140 141 CvANN_MLP mlp; 142 CvANN_MLP_TrainParams params; 143 CvTermCriteria criteria; 144 criteria.max_iter = 100; 145 criteria.epsilon = 0.00001f; 146 criteria.type = CV_TERMCRIT_ITER | CV_TERMCRIT_EPS; 147 params.train_method = CvANN_MLP_TrainParams::BACKPROP; 148 params.bp_dw_scale = 0.05f; 149 params.bp_moment_scale = 0.05f; 150 params.term_crit = criteria; 151 152 mlp.create(layers); 153 154 // train 155 mlp.train(trainingData, trainingClasses, cv::Mat(), cv::Mat(), params); 156 157 cv::Mat response(1, 1, CV_32FC1); 158 cv::Mat predicted(testClasses.rows, 1, CV_32F); 159 for(int i = 0; i < testData.rows; i++) { 160 cv::Mat response(1, 1, CV_32FC1); 161 cv::Mat sample = testData.row(i); 162 163 mlp.predict(sample, response); 164 predicted.at<float>(i,0) = response.at<float>(0,0); 165 166 } 167 168 cout << "Accuracy_{MLP} = " << evaluate(predicted, testClasses) << endl; 169 plot_binary(testData, predicted, "Predictions Backpropagation"); 170 } 171 172 void knn(cv::Mat& trainingData, cv::Mat& trainingClasses, cv::Mat& testData, cv::Mat& testClasses, int K) { 173 174 CvKNearest knn(trainingData, trainingClasses, cv::Mat(), false, K); 175 cv::Mat predicted(testClasses.rows, 1, CV_32F); 176 for(int i = 0; i < testData.rows; i++) { 177 const cv::Mat sample = testData.row(i); 178 predicted.at<float>(i,0) = knn.find_nearest(sample, K); 179 } 180 181 cout << "Accuracy_{KNN} = " << evaluate(predicted, testClasses) << endl; 182 plot_binary(testData, predicted, "Predictions KNN"); 183 184 } 185 186 void bayes(cv::Mat& trainingData, cv::Mat& trainingClasses, cv::Mat& testData, cv::Mat& testClasses) { 187 188 CvNormalBayesClassifier bayes(trainingData, trainingClasses); 189 cv::Mat predicted(testClasses.rows, 1, CV_32F); 190 for (int i = 0; i < testData.rows; i++) { 191 const cv::Mat sample = testData.row(i); 192 predicted.at<float> (i, 0) = bayes.predict(sample); 193 } 194 195 cout << "Accuracy_{BAYES} = " << evaluate(predicted, testClasses) << endl; 196 plot_binary(testData, predicted, "Predictions Bayes"); 197 198 } 199 200 void decisiontree(cv::Mat& trainingData, cv::Mat& trainingClasses, cv::Mat& testData, cv::Mat& testClasses) { 201 202 CvDTree dtree; 203 cv::Mat var_type(3, 1, CV_8U); 204 205 // define attributes as numerical 206 var_type.at<unsigned int>(0,0) = CV_VAR_NUMERICAL; 207 var_type.at<unsigned int>(0,1) = CV_VAR_NUMERICAL; 208 // define output node as numerical 209 var_type.at<unsigned int>(0,2) = CV_VAR_NUMERICAL; 210 211 dtree.train(trainingData,CV_ROW_SAMPLE, trainingClasses, cv::Mat(), cv::Mat(), var_type, cv::Mat(), CvDTreeParams()); 212 cv::Mat predicted(testClasses.rows, 1, CV_32F); 213 for (int i = 0; i < testData.rows; i++) { 214 const cv::Mat sample = testData.row(i); 215 CvDTreeNode* prediction = dtree.predict(sample); 216 predicted.at<float> (i, 0) = prediction->value; 217 } 218 219 cout << "Accuracy_{TREE} = " << evaluate(predicted, testClasses) << endl; 220 plot_binary(testData, predicted, "Predictions tree"); 221 222 } 223 224 225 int main() { 226 227 cv::Mat trainingData(numTrainingPoints, 2, CV_32FC1); 228 cv::Mat testData(numTestPoints, 2, CV_32FC1); 229 230 cv::randu(trainingData,0,1); 231 cv::randu(testData,0,1); 232 233 cv::Mat trainingClasses = labelData(trainingData, eq); 234 cv::Mat testClasses = labelData(testData, eq); 235 236 plot_binary(trainingData, trainingClasses, "Training Data"); 237 plot_binary(testData, testClasses, "Test Data"); 238 239 svm(trainingData, trainingClasses, testData, testClasses); 240 mlp(trainingData, trainingClasses, testData, testClasses); 241 knn(trainingData, trainingClasses, testData, testClasses, 3); 242 bayes(trainingData, trainingClasses, testData, testClasses); 243 decisiontree(trainingData, trainingClasses, testData, testClasses); 244 245 cv::waitKey(); 246 247 return 0; 248 }
图像分类结果：
