OpenCV3 SVM ANN Adaboost KNN 随机森林等机器学习方法对OCR分类

转摘自http://www.cnblogs.com/denny402/p/5032839.html
opencv3中的ml类与opencv2中发生了变化，下面列举opencv3的机器学习类方法实例：
用途是opencv自带的ocr样本的分类功能，其中神经网络和adaboost训练速度很慢，效果还是knn的最好；
  1 #include <opencv2/opencv.hpp>
  2 #include <iostream>
  3 using namespace std;
  4 using namespace cv;
  5 using namespace cv::ml;
  6 
  7 // 读取文件数据
  8 bool read_num_class_data(const string& filename, int var_count, Mat* _data, Mat* _responses)
  9 {
 10     const int M = 1024;
 11     char buf[M + 2];
 12 
 13     Mat el_ptr(1, var_count, CV_32F);
 14     int i;
 15     vector<int> responses;
 16 
 17     _data->release();
 18     _responses->release();
 19     FILE *f;
 20     fopen_s(&f, filename.c_str(), "rt");
 21     if (!f)
 22     {
 23         cout << "Could not read the database " << filename << endl;
 24         return false;
 25     }
 26 
 27     for (;;)
 28     {
 29         char* ptr;
 30         if (!fgets(buf, M, f) || !strchr(buf, ','))
 31             break;
 32         responses.push_back((int)buf[0]);
 33         ptr = buf + 2;
 34         for (i = 0; i < var_count; i++)
 35         {
 36             int n = 0;
 37             sscanf_s(ptr, "%f%n", &el_ptr.at<float>(i), &n);
 38             ptr += n + 1;
 39         }
 40         if (i < var_count)
 41             break;
 42         _data->push_back(el_ptr);
 43     }
 44     fclose(f);
 45     Mat(responses).copyTo(*_responses);
 46     return true;
 47 }
 48 
 49 
 50 //准备训练数据
 51 Ptr<TrainData> prepare_train_data(const Mat& data, const Mat& responses, int ntrain_samples)
 52 {
 53     Mat sample_idx = Mat::zeros(1, data.rows, CV_8U);
 54     Mat train_samples = sample_idx.colRange(0, ntrain_samples);
 55     train_samples.setTo(Scalar::all(1));
 56 
 57     int nvars = data.cols;
 58     Mat var_type(nvars + 1, 1, CV_8U);
 59     var_type.setTo(Scalar::all(VAR_ORDERED));
 60     var_type.at<uchar>(nvars) = VAR_CATEGORICAL;
 61 
 62     return TrainData::create(data, ROW_SAMPLE, responses,
 63         noArray(), sample_idx, noArray(), var_type);
 64 }
 65 
 66 //设置迭代条件
 67 inline TermCriteria TC(int iters, double eps)
 68 {
 69     return TermCriteria(TermCriteria::MAX_ITER + (eps > 0 ? TermCriteria::EPS : 0), iters, eps);
 70 }
 71 
 72 //分类预测
 73 void test_and_save_classifier(const Ptr<StatModel>& model, const Mat& data, const Mat& responses,
 74     int ntrain_samples, int rdelta)
 75 {
 76     int i, nsamples_all = data.rows;
 77     double train_hr = 0, test_hr = 0;
 78 
 79     // compute prediction error on train and test data
 80     for (i = 0; i < nsamples_all; i++)
 81     {
 82         Mat sample = data.row(i);
 83 
 84         float r = model->predict(sample);
 85         r = std::abs(r + rdelta - responses.at<int>(i)) <= FLT_EPSILON ? 1.f : 0.f;
 86 
 87         if (i < ntrain_samples)
 88             train_hr += r;
 89         else
 90             test_hr += r;
 91     }
 92 
 93     test_hr /= nsamples_all - ntrain_samples;
 94     train_hr = ntrain_samples > 0 ? train_hr / ntrain_samples : 1.;
 95 
 96     printf("Recognition rate: train = %.1f%%, test = %.1f%%
",
 97         train_hr*100., test_hr*100.);
 98 }
 99 
100 //随机树分类
101 bool build_rtrees_classifier(const string& data_filename)
102 {
103     Mat data;
104     Mat responses;
105     read_num_class_data(data_filename, 16, &data, &responses);
106 
107     int nsamples_all = data.rows;
108     int ntrain_samples = (int)(nsamples_all*0.8);
109 
110     Ptr<RTrees> model;
111     Ptr<TrainData> tdata = prepare_train_data(data, responses, ntrain_samples);
112     model = RTrees::create();
113     model->setMaxDepth(10);
114     model->setMinSampleCount(10);
115     model->setRegressionAccuracy(0);
116     model->setUseSurrogates(false);
117     model->setMaxCategories(15);
118     model->setPriors(Mat());
119     model->setCalculateVarImportance(true);
120     model->setActiveVarCount(4);
121     model->setTermCriteria(TC(100, 0.01f));
122     model->train(tdata);
123     test_and_save_classifier(model, data, responses, ntrain_samples, 0);
124     cout << "Number of trees: " << model->getRoots().size() << endl;
125 
126     // Print variable importance
127     Mat var_importance = model->getVarImportance();
128     if (!var_importance.empty())
129     {
130         double rt_imp_sum = sum(var_importance)[0];
131         printf("var#	importance (in %%):
");
132         int i, n = (int)var_importance.total();
133         for (i = 0; i < n; i++)
134             printf("%-2d	%-4.1f
", i, 100.f*var_importance.at<float>(i) / rt_imp_sum);
135     }
136 
137     return true;
138 }
139 
140 //adaboost分类
141 bool build_boost_classifier(const string& data_filename)
142 {
143     const int class_count = 26;
144     Mat data;
145     Mat responses;
146     Mat weak_responses;
147 
148     read_num_class_data(data_filename, 16, &data, &responses);
149     int i, j, k;
150     Ptr<Boost> model;
151 
152     int nsamples_all = data.rows;
153     int ntrain_samples = (int)(nsamples_all*0.5);
154     int var_count = data.cols;
155 
156     Mat new_data(ntrain_samples*class_count, var_count + 1, CV_32F);
157     Mat new_responses(ntrain_samples*class_count, 1, CV_32S);
158 
159     for (i = 0; i < ntrain_samples; i++)
160     {
161         const float* data_row = data.ptr<float>(i);
162         for (j = 0; j < class_count; j++)
163         {
164             float* new_data_row = (float*)new_data.ptr<float>(i*class_count + j);
165             memcpy(new_data_row, data_row, var_count * sizeof(data_row[0]));
166             new_data_row[var_count] = (float)j;
167             new_responses.at<int>(i*class_count + j) = responses.at<int>(i) == j + 'A';
168         }
169     }
170 
171     Mat var_type(1, var_count + 2, CV_8U);
172     var_type.setTo(Scalar::all(VAR_ORDERED));
173     var_type.at<uchar>(var_count) = var_type.at<uchar>(var_count + 1) = VAR_CATEGORICAL;
174 
175     Ptr<TrainData> tdata = TrainData::create(new_data, ROW_SAMPLE, new_responses,
176         noArray(), noArray(), noArray(), var_type);
177     vector<double> priors(2);
178     priors[0] = 1;
179     priors[1] = 26;
180 
181     model = Boost::create();
182     model->setBoostType(Boost::GENTLE);
183     model->setWeakCount(100);
184     model->setWeightTrimRate(0.95);
185     model->setMaxDepth(5);
186     model->setUseSurrogates(false);
187     model->setPriors(Mat(priors));
188     model->train(tdata);
189     Mat temp_sample(1, var_count + 1, CV_32F);
190     float* tptr = temp_sample.ptr<float>();
191 
192     // compute prediction error on train and test data
193     double train_hr = 0, test_hr = 0;
194     for (i = 0; i < nsamples_all; i++)
195     {
196         int best_class = 0;
197         double max_sum = -DBL_MAX;
198         const float* ptr = data.ptr<float>(i);
199         for (k = 0; k < var_count; k++)
200             tptr[k] = ptr[k];
201 
202         for (j = 0; j < class_count; j++)
203         {
204             tptr[var_count] = (float)j;
205             float s = model->predict(temp_sample, noArray(), StatModel::RAW_OUTPUT);
206             if (max_sum < s)
207             {
208                 max_sum = s;
209                 best_class = j + 'A';
210             }
211         }
212 
213         double r = std::abs(best_class - responses.at<int>(i)) < FLT_EPSILON ? 1 : 0;
214         if (i < ntrain_samples)
215             train_hr += r;
216         else
217             test_hr += r;
218     }
219 
220     test_hr /= nsamples_all - ntrain_samples;
221     train_hr = ntrain_samples > 0 ? train_hr / ntrain_samples : 1.;
222     printf("Recognition rate: train = %.1f%%, test = %.1f%%
",
223         train_hr*100., test_hr*100.);
224 
225     cout << "Number of trees: " << model->getRoots().size() << endl;
226     return true;
227 }
228 
229 //多层感知机分类（ANN）
230 bool build_mlp_classifier(const string& data_filename)
231 {
232     const int class_count = 26;
233     Mat data;
234     Mat responses;
235 
236     read_num_class_data(data_filename, 16, &data, &responses);
237     Ptr<ANN_MLP> model;
238 
239     int nsamples_all = data.rows;
240     int ntrain_samples = (int)(nsamples_all*0.8);
241     Mat train_data = data.rowRange(0, ntrain_samples);
242     Mat train_responses = Mat::zeros(ntrain_samples, class_count, CV_32F);
243 
244     // 1. unroll the responses
245     cout << "Unrolling the responses...
";
246     for (int i = 0; i < ntrain_samples; i++)
247     {
248         int cls_label = responses.at<int>(i) - 'A';
249         train_responses.at<float>(i, cls_label) = 1.f;
250     }
251 
252     // 2. train classifier
253     int layer_sz[] = { data.cols, 100, 100, class_count };
254     int nlayers = (int)(sizeof(layer_sz) / sizeof(layer_sz[0]));
255     Mat layer_sizes(1, nlayers, CV_32S, layer_sz);
256 
257 #if 1
258     int method = ANN_MLP::BACKPROP;
259     double method_param = 0.001;
260     int max_iter = 300;
261 #else
262     int method = ANN_MLP::RPROP;
263     double method_param = 0.1;
264     int max_iter = 1000;
265 #endif
266 
267     Ptr<TrainData> tdata = TrainData::create(train_data, ROW_SAMPLE, train_responses);
268     model = ANN_MLP::create();
269     model->setLayerSizes(layer_sizes);
270     model->setActivationFunction(ANN_MLP::SIGMOID_SYM, 0, 0);
271     model->setTermCriteria(TC(max_iter, 0));
272     model->setTrainMethod(method, method_param);
273     model->train(tdata);
274     return true;
275 }
276 
277 //K最近邻分类
278 bool build_knearest_classifier(const string& data_filename, int K)
279 {
280     Mat data;
281     Mat responses;
282     read_num_class_data(data_filename, 16, &data, &responses);
283     int nsamples_all = data.rows;
284     int ntrain_samples = (int)(nsamples_all*0.8);
285 
286     Ptr<TrainData> tdata = prepare_train_data(data, responses, ntrain_samples);
287     Ptr<KNearest> model = KNearest::create();
288     model->setDefaultK(K);
289     model->setIsClassifier(true);
290     model->train(tdata);
291 
292     test_and_save_classifier(model, data, responses, ntrain_samples, 0);
293     return true;
294 }
295 
296 //贝叶斯分类
297 bool build_nbayes_classifier(const string& data_filename)
298 {
299     Mat data;
300     Mat responses;
301     read_num_class_data(data_filename, 16, &data, &responses);
302 
303     int nsamples_all = data.rows;
304     int ntrain_samples = (int)(nsamples_all*0.8);
305 
306     Ptr<NormalBayesClassifier> model;
307     Ptr<TrainData> tdata = prepare_train_data(data, responses, ntrain_samples);
308     model = NormalBayesClassifier::create();
309     model->train(tdata);
310 
311     test_and_save_classifier(model, data, responses, ntrain_samples, 0);
312     return true;
313 }
314 
315 
316 //svm分类
317 bool build_svm_classifier(const string& data_filename)
318 {
319     Mat data;
320     Mat responses;
321     read_num_class_data(data_filename, 16, &data, &responses);
322 
323     int nsamples_all = data.rows;
324     int ntrain_samples = (int)(nsamples_all*0.8);
325 
326     Ptr<SVM> model;
327     Ptr<TrainData> tdata = prepare_train_data(data, responses, ntrain_samples);
328     model = SVM::create();
329     model->setType(SVM::C_SVC);
330     model->setKernel(SVM::LINEAR);
331     model->setC(1);
332     model->train(tdata);
333 
334     test_and_save_classifier(model, data, responses, ntrain_samples, 0);
335     return true;
336 }
337 
338 int main()
339 {
340     string data_filename = "D:\Program Files\opencv\sources\samples\data\letter-recognition.data";  //字母数据
341 
342     cout << "svm分类：" << endl;
343     build_svm_classifier(data_filename);
344 
345     cout << "贝叶斯分类：" << endl;
346     build_nbayes_classifier(data_filename);
347 
348     cout << "K最近邻分类：" << endl;
349     build_knearest_classifier(data_filename, 10);
350 
351     cout << "随机树分类：" << endl;
352     build_rtrees_classifier(data_filename);
353 
354     cout << "adaboost分类：" << endl;
355     build_boost_classifier(data_filename);
356 
357     cout << "ANN（多层感知机)分类：" << endl;
358     build_mlp_classifier(data_filename);
359 
360     system("pause");
361     return 0;
362 }