1 #include "opencv2/core.hpp"
2 #include "opencv2/imgproc.hpp"
3 #include "opencv2/ml.hpp"
4 #include "opencv2/highgui.hpp"
5 #ifdef HAVE_OPENCV_OCL
6 #define _OCL_KNN_ 1 // select whether using ocl::KNN method or not, default is using
7 #define _OCL_SVM_ 1 // select whether using ocl::svm method or not, default is using
8 #include "opencv2/ocl/ocl.hpp"
9 #endif
10 
11 #include <stdio.h>
12 
13 using namespace std;
14 using namespace cv;
15 using namespace cv::ml;
16 
17 const Scalar WHITE_COLOR = Scalar(255,255,255);
18 const string winName = "points";
19 const int testStep = 5;
20 
21 Mat img, imgDst;
22 RNG rng;
23 
24 vector<Point>  trainedPoints;
25 vector<int>    trainedPointsMarkers;
26 const int MAX_CLASSES = 2;
27 vector<Vec3b>  classColors(MAX_CLASSES);
28 int currentClass = 0;
29 vector<int> classCounters(MAX_CLASSES);
30 
31 #define _NBC_ 1 // normal Bayessian classifier
32 #define _KNN_ 1 // k nearest neighbors classifier
33 #define _SVM_ 1 // support vectors machine
34 #define _DT_  1 // decision tree
35 #define _BT_  1 // ADA Boost
36 #define _GBT_ 0 // gradient boosted trees
37 #define _RF_  1 // random forest
38 #define _ANN_ 1 // artificial neural networks
39 #define _EM_  1 // expectation-maximization
40 
on_mouse(int event,int x,int y,int,void *)41 static void on_mouse( int event, int x, int y, int /*flags*/, void* )
42 {
43     if( img.empty() )
44         return;
45 
46     int updateFlag = 0;
47 
48     if( event == EVENT_LBUTTONUP )
49     {
50         trainedPoints.push_back( Point(x,y) );
51         trainedPointsMarkers.push_back( currentClass );
52         classCounters[currentClass]++;
53         updateFlag = true;
54     }
55 
56     //draw
57     if( updateFlag )
58     {
59         img = Scalar::all(0);
60 
61         // draw points
62         for( size_t i = 0; i < trainedPoints.size(); i++ )
63         {
64             Vec3b c = classColors[trainedPointsMarkers[i]];
65             circle( img, trainedPoints[i], 5, Scalar(c), -1 );
66         }
67 
68         imshow( winName, img );
69    }
70 }
71 
prepare_train_samples(const vector<Point> & pts)72 static Mat prepare_train_samples(const vector<Point>& pts)
73 {
74     Mat samples;
75     Mat(pts).reshape(1, (int)pts.size()).convertTo(samples, CV_32F);
76     return samples;
77 }
78 
prepare_train_data()79 static Ptr<TrainData> prepare_train_data()
80 {
81     Mat samples = prepare_train_samples(trainedPoints);
82     return TrainData::create(samples, ROW_SAMPLE, Mat(trainedPointsMarkers));
83 }
84 
predict_and_paint(const Ptr<StatModel> & model,Mat & dst)85 static void predict_and_paint(const Ptr<StatModel>& model, Mat& dst)
86 {
87     Mat testSample( 1, 2, CV_32FC1 );
88     for( int y = 0; y < img.rows; y += testStep )
89     {
90         for( int x = 0; x < img.cols; x += testStep )
91         {
92             testSample.at<float>(0) = (float)x;
93             testSample.at<float>(1) = (float)y;
94 
95             int response = (int)model->predict( testSample );
96             dst.at<Vec3b>(y, x) = classColors[response];
97         }
98     }
99 }
100 
101 #if _NBC_
find_decision_boundary_NBC()102 static void find_decision_boundary_NBC()
103 {
104     // learn classifier
105     Ptr<NormalBayesClassifier> normalBayesClassifier = StatModel::train<NormalBayesClassifier>(prepare_train_data());
106 
107     predict_and_paint(normalBayesClassifier, imgDst);
108 }
109 #endif
110 
111 
112 #if _KNN_
find_decision_boundary_KNN(int K)113 static void find_decision_boundary_KNN( int K )
114 {
115 
116     Ptr<KNearest> knn = KNearest::create();
117     knn->setDefaultK(K);
118     knn->setIsClassifier(true);
119     knn->train(prepare_train_data());
120     predict_and_paint(knn, imgDst);
121 }
122 #endif
123 
124 #if _SVM_
find_decision_boundary_SVM(double C)125 static void find_decision_boundary_SVM( double C )
126 {
127     Ptr<SVM> svm = SVM::create();
128     svm->setType(SVM::C_SVC);
129     svm->setKernel(SVM::POLY); //SVM::LINEAR;
130     svm->setDegree(0.5);
131     svm->setGamma(1);
132     svm->setCoef0(1);
133     svm->setNu(0.5);
134     svm->setP(0);
135     svm->setTermCriteria(TermCriteria(TermCriteria::MAX_ITER+TermCriteria::EPS, 1000, 0.01));
136     svm->setC(C);
137     svm->train(prepare_train_data());
138     predict_and_paint(svm, imgDst);
139 
140     Mat sv = svm->getSupportVectors();
141     for( int i = 0; i < sv.rows; i++ )
142     {
143         const float* supportVector = sv.ptr<float>(i);
144         circle( imgDst, Point(saturate_cast<int>(supportVector[0]),saturate_cast<int>(supportVector[1])), 5, Scalar(255,255,255), -1 );
145     }
146 }
147 #endif
148 
149 #if _DT_
find_decision_boundary_DT()150 static void find_decision_boundary_DT()
151 {
152     Ptr<DTrees> dtree = DTrees::create();
153     dtree->setMaxDepth(8);
154     dtree->setMinSampleCount(2);
155     dtree->setUseSurrogates(false);
156     dtree->setCVFolds(0); // the number of cross-validation folds
157     dtree->setUse1SERule(false);
158     dtree->setTruncatePrunedTree(false);
159     dtree->train(prepare_train_data());
160     predict_and_paint(dtree, imgDst);
161 }
162 #endif
163 
164 #if _BT_
find_decision_boundary_BT()165 static void find_decision_boundary_BT()
166 {
167     Ptr<Boost> boost = Boost::create();
168     boost->setBoostType(Boost::DISCRETE);
169     boost->setWeakCount(100);
170     boost->setWeightTrimRate(0.95);
171     boost->setMaxDepth(2);
172     boost->setUseSurrogates(false);
173     boost->setPriors(Mat());
174     boost->train(prepare_train_data());
175     predict_and_paint(boost, imgDst);
176 }
177 
178 #endif
179 
180 #if _GBT_
find_decision_boundary_GBT()181 static void find_decision_boundary_GBT()
182 {
183     GBTrees::Params params( GBTrees::DEVIANCE_LOSS, // loss_function_type
184                          100, // weak_count
185                          0.1f, // shrinkage
186                          1.0f, // subsample_portion
187                          2, // max_depth
188                          false // use_surrogates )
189                          );
190 
191     Ptr<GBTrees> gbtrees = StatModel::train<GBTrees>(prepare_train_data(), params);
192     predict_and_paint(gbtrees, imgDst);
193 }
194 #endif
195 
196 #if _RF_
find_decision_boundary_RF()197 static void find_decision_boundary_RF()
198 {
199     Ptr<RTrees> rtrees = RTrees::create();
200     rtrees->setMaxDepth(4);
201     rtrees->setMinSampleCount(2);
202     rtrees->setRegressionAccuracy(0.f);
203     rtrees->setUseSurrogates(false);
204     rtrees->setMaxCategories(16);
205     rtrees->setPriors(Mat());
206     rtrees->setCalculateVarImportance(false);
207     rtrees->setActiveVarCount(1);
208     rtrees->setTermCriteria(TermCriteria(TermCriteria::MAX_ITER, 5, 0));
209     rtrees->train(prepare_train_data());
210     predict_and_paint(rtrees, imgDst);
211 }
212 
213 #endif
214 
215 #if _ANN_
find_decision_boundary_ANN(const Mat & layer_sizes)216 static void find_decision_boundary_ANN( const Mat&  layer_sizes )
217 {
218     Mat trainClasses = Mat::zeros( (int)trainedPoints.size(), (int)classColors.size(), CV_32FC1 );
219     for( int i = 0; i < trainClasses.rows; i++ )
220     {
221         trainClasses.at<float>(i, trainedPointsMarkers[i]) = 1.f;
222     }
223 
224     Mat samples = prepare_train_samples(trainedPoints);
225     Ptr<TrainData> tdata = TrainData::create(samples, ROW_SAMPLE, trainClasses);
226 
227     Ptr<ANN_MLP> ann = ANN_MLP::create();
228     ann->setLayerSizes(layer_sizes);
229     ann->setActivationFunction(ANN_MLP::SIGMOID_SYM, 1, 1);
230     ann->setTermCriteria(TermCriteria(TermCriteria::MAX_ITER+TermCriteria::EPS, 300, FLT_EPSILON));
231     ann->setTrainMethod(ANN_MLP::BACKPROP, 0.001);
232     ann->train(tdata);
233     predict_and_paint(ann, imgDst);
234 }
235 #endif
236 
237 #if _EM_
find_decision_boundary_EM()238 static void find_decision_boundary_EM()
239 {
240     img.copyTo( imgDst );
241 
242     Mat samples = prepare_train_samples(trainedPoints);
243 
244     int i, j, nmodels = (int)classColors.size();
245     vector<Ptr<EM> > em_models(nmodels);
246     Mat modelSamples;
247 
248     for( i = 0; i < nmodels; i++ )
249     {
250         const int componentCount = 3;
251 
252         modelSamples.release();
253         for( j = 0; j < samples.rows; j++ )
254         {
255             if( trainedPointsMarkers[j] == i )
256                 modelSamples.push_back(samples.row(j));
257         }
258 
259         // learn models
260         if( !modelSamples.empty() )
261         {
262             Ptr<EM> em = EM::create();
263             em->setClustersNumber(componentCount);
264             em->setCovarianceMatrixType(EM::COV_MAT_DIAGONAL);
265             em->trainEM(modelSamples, noArray(), noArray(), noArray());
266             em_models[i] = em;
267         }
268     }
269 
270     // classify coordinate plane points using the bayes classifier, i.e.
271     // y(x) = arg max_i=1_modelsCount likelihoods_i(x)
272     Mat testSample(1, 2, CV_32FC1 );
273     Mat logLikelihoods(1, nmodels, CV_64FC1, Scalar(-DBL_MAX));
274 
275     for( int y = 0; y < img.rows; y += testStep )
276     {
277         for( int x = 0; x < img.cols; x += testStep )
278         {
279             testSample.at<float>(0) = (float)x;
280             testSample.at<float>(1) = (float)y;
281 
282             for( i = 0; i < nmodels; i++ )
283             {
284                 if( !em_models[i].empty() )
285                     logLikelihoods.at<double>(i) = em_models[i]->predict2(testSample, noArray())[0];
286             }
287             Point maxLoc;
288             minMaxLoc(logLikelihoods, 0, 0, 0, &maxLoc);
289             imgDst.at<Vec3b>(y, x) = classColors[maxLoc.x];
290         }
291     }
292 }
293 #endif
294 
main()295 int main()
296 {
297     cout << "Use:" << endl
298          << "  key '0' .. '1' - switch to class #n" << endl
299          << "  left mouse button - to add new point;" << endl
300          << "  key 'r' - to run the ML model;" << endl
301          << "  key 'i' - to init (clear) the data." << endl << endl;
302 
303     cv::namedWindow( "points", 1 );
304     img.create( 480, 640, CV_8UC3 );
305     imgDst.create( 480, 640, CV_8UC3 );
306 
307     imshow( "points", img );
308     setMouseCallback( "points", on_mouse );
309 
310     classColors[0] = Vec3b(0, 255, 0);
311     classColors[1] = Vec3b(0, 0, 255);
312 
313     for(;;)
314     {
315         uchar key = (uchar)waitKey();
316 
317         if( key == 27 ) break;
318 
319         if( key == 'i' ) // init
320         {
321             img = Scalar::all(0);
322 
323             trainedPoints.clear();
324             trainedPointsMarkers.clear();
325             classCounters.assign(MAX_CLASSES, 0);
326 
327             imshow( winName, img );
328         }
329 
330         if( key == '0' || key == '1' )
331         {
332             currentClass = key - '0';
333         }
334 
335         if( key == 'r' ) // run
336         {
337             double minVal = 0;
338             minMaxLoc(classCounters, &minVal, 0, 0, 0);
339             if( minVal == 0 )
340             {
341                 printf("each class should have at least 1 point\n");
342                 continue;
343             }
344             img.copyTo( imgDst );
345 #if _NBC_
346             find_decision_boundary_NBC();
347             imshow( "NormalBayesClassifier", imgDst );
348 #endif
349 #if _KNN_
350             find_decision_boundary_KNN( 3 );
351             imshow( "kNN", imgDst );
352 
353             find_decision_boundary_KNN( 15 );
354             imshow( "kNN2", imgDst );
355 #endif
356 
357 #if _SVM_
358             //(1)-(2)separable and not sets
359 
360             find_decision_boundary_SVM( 1 );
361             imshow( "classificationSVM1", imgDst );
362 
363             find_decision_boundary_SVM( 10 );
364             imshow( "classificationSVM2", imgDst );
365 #endif
366 
367 #if _DT_
368             find_decision_boundary_DT();
369             imshow( "DT", imgDst );
370 #endif
371 
372 #if _BT_
373             find_decision_boundary_BT();
374             imshow( "BT", imgDst);
375 #endif
376 
377 #if _GBT_
378             find_decision_boundary_GBT();
379             imshow( "GBT", imgDst);
380 #endif
381 
382 #if _RF_
383             find_decision_boundary_RF();
384             imshow( "RF", imgDst);
385 #endif
386 
387 #if _ANN_
388             Mat layer_sizes1( 1, 3, CV_32SC1 );
389             layer_sizes1.at<int>(0) = 2;
390             layer_sizes1.at<int>(1) = 5;
391             layer_sizes1.at<int>(2) = (int)classColors.size();
392             find_decision_boundary_ANN( layer_sizes1 );
393             imshow( "ANN", imgDst );
394 #endif
395 
396 #if _EM_
397             find_decision_boundary_EM();
398             imshow( "EM", imgDst );
399 #endif
400         }
401     }
402 
403     return 0;
404 }
405