Galaxy/annCh_train.cpp

#include <numeric>
#include <ctime>

#include "easypr/train/annCh_train.h"
#include "easypr/config.h"
#include "easypr/core/chars_identify.h"
#include "easypr/core/feature.h"
#include "easypr/core/core_func.h"
#include "easypr/util/util.h"
#include "easypr/train/create_data.h"

namespace easypr {

  AnnChTrain::AnnChTrain(const char* chars_folder, const char* xml)
    : chars_folder_(chars_folder), ann_xml_(xml)
  {
  ann_ = cv::ml::ANN_MLP::create();
  type = 1;
  kv_ = std::shared_ptr<Kv>(new Kv);
  kv_->load("resources/text/province_mapping");
  extractFeature = getGrayPlusProject;
}

  void AnnChTrain::train()
  {
    int classNumber = 0;
    int input_number = 0;
    int hidden_number = 0;
    int output_number = 0;

    bool useLBP = false;
    if (useLBP)
      input_number = kCharLBPPatterns * kCharLBPGridX * kCharLBPGridY;
    else
      input_number = kGrayCharHeight * kGrayCharWidth;

    input_number += 64;

  classNumber = kChineseNumber;
  hidden_number = kCharHiddenNeurans;
  output_number = classNumber;
  cv::Mat layers;

    int first_hidden_neurons = 48;
    int second_hidden_neurons = 32;

    int N = input_number;
    int m = output_number;
    //int first_hidden_neurons = int(std::sqrt((m + 2) * N) + 2 * std::sqrt(N / (m + 2)));
    //int second_hidden_neurons = int(m * std::sqrt(N / (m + 2)));

    bool useTLFN = false;
    if (!useTLFN) {
      layers.create(1, 3, CV_32SC1);
      layers.at<int>(0) = input_number;
      layers.at<int>(1) = hidden_number;
      layers.at<int>(2) = output_number;
    }
    else {
      fprintf(stdout, ">> Use two-layers neural networks,\n");
      fprintf(stdout, ">> First_hidden_neurons: %d \n", first_hidden_neurons);
      fprintf(stdout, ">> Second_hidden_neurons: %d \n", second_hidden_neurons);

      layers.create(1, 4, CV_32SC1);
      layers.at<int>(0) = input_number;
      layers.at<int>(1) = first_hidden_neurons;
      layers.at<int>(2) = second_hidden_neurons;
      layers.at<int>(3) = output_number;
    }

  ann_->setLayerSizes(layers);
  ann_->setActivationFunction(cv::ml::ANN_MLP::SIGMOID_SYM, 1, 1);
  ann_->setTrainMethod(cv::ml::ANN_MLP::TrainingMethods::BACKPROP);
  ann_->setTermCriteria(cvTermCriteria(CV_TERMCRIT_ITER, 30000, 0.0001));
  ann_->setBackpropWeightScale(0.1);
  ann_->setBackpropMomentumScale(0.1);

  auto files = Utils::getFiles(chars_folder_);
  if (files.size() == 0) {
    fprintf(stdout, "No file found in the train folder!\n");
    fprintf(stdout, "You should create a folder named \"tmp\" in EasyPR main folder.\n");
    fprintf(stdout, "Copy train data folder(like \"annCh\") under \"tmp\". \n");
    return;
  }

  // using raw data or raw + synthic data.
  trainVal(m_number_for_count);
}

std::pair<std::string, std::string> AnnChTrain::identifyGrayChinese(cv::Mat input) {
  Mat feature;
  extractFeature(input, feature);
  float maxVal = -2;
  int result = 0;

  cv::Mat output(1, kChineseNumber, CV_32FC1);
  ann_->predict(feature, output);

  for (int j = 0; j < kChineseNumber; j++) {
    float val = output.at<float>(j);
    //std::cout << "j:" << j << "val:" << val << std::endl;
    if (val > maxVal) {
      maxVal = val;
      result = j;
    }
  }

  auto index = result + kCharsTotalNumber - kChineseNumber;
  const char* key = kChars[index];
  std::string s = key;
  std::string province = kv_->get(s);

  return std::make_pair(s, province);
}

  void AnnChTrain::test() {
    //TODO
}

void AnnChTrain::trainVal(size_t number_for_count) {
  assert(chars_folder_);
  cv::Mat train_samples;
  std::vector<cv::Mat>  train_images, val_images;
  std::vector<int> train_label, val_labels;
  float percentage = 0.7f;
  int classNumber = kChineseNumber;

  for (int i = 0; i < classNumber; ++i) { 
    auto char_key = kChars[i + kCharsTotalNumber - classNumber];
    char sub_folder[512] = { 0 };
    sprintf(sub_folder, "%s/%s", chars_folder_, char_key);
    std::string test_char(char_key);
    // if (test_char != "zh_yun") continue;

    fprintf(stdout, ">> Testing characters %s in %s \n", char_key, sub_folder);
    auto chars_files = utils::getFiles(sub_folder);
    size_t char_size = chars_files.size();
    fprintf(stdout, ">> Characters count: %d \n", (int)char_size);

    std::vector<cv::Mat> matVec;
    matVec.reserve(number_for_count);
    for (auto file : chars_files) {
      std::cout << file << std::endl;
      auto img = cv::imread(file, IMREAD_GRAYSCALE);  // a grayscale image 
      Mat img_resize;
      img_resize.create(kGrayCharHeight, kGrayCharWidth, CV_8UC1);
      resize(img, img_resize, img_resize.size(), 0, 0, INTER_LINEAR);
      matVec.push_back(img_resize);
    }
    // genrate the synthetic images
    for (int t = 0; t < (int)number_for_count - (int)char_size; t++) {
      int rand_range = char_size + t;
      int ran_num = rand() % rand_range;
      auto img = matVec.at(ran_num);
      SHOW_IMAGE(img, 0);
      auto simg = generateSyntheticImage(img);
      SHOW_IMAGE(simg, 0);
      matVec.push_back(simg);
    }
    fprintf(stdout, ">> Characters count: %d \n", (int)matVec.size());

    // random sort the mat;
    srand(unsigned(time(NULL)));
    random_shuffle(matVec.begin(), matVec.end());

    int mat_size = (int)matVec.size();
    int split_index = int((float)mat_size * percentage);
    for (int j = mat_size - 1; j >= 0; j--) {
      Mat img = matVec.at(j);
      if (1) {
        Mat feature;
        extractFeature(img, feature);
        if (j <= split_index) {
          train_samples.push_back(feature);
          train_images.push_back(img);
          train_label.push_back(i);
        }
        else {
          val_images.push_back(img);
          val_labels.push_back(i);
        }
      }
    }
  } 
  // generate train data
  train_samples.convertTo(train_samples, CV_32F);
  cv::Mat train_classes = cv::Mat::zeros((int)train_label.size(), classNumber, CV_32F);
  for (int i = 0; i < train_classes.rows; ++i)
    train_classes.at<float>(i, train_label[i]) = 1.f;
  auto train_data = cv::ml::TrainData::create(train_samples, cv::ml::SampleTypes::ROW_SAMPLE, train_classes);

  // train the data, calculate the cost time
  std::cout << "Training ANN chinese model, please wait..." << std::endl;
  long start = utils::getTimestamp();
  ann_->train(train_data);
  long end = utils::getTimestamp();
  ann_->save(ann_xml_);
  std::cout << "Your ANN Model was saved to " << ann_xml_ << std::endl;
  std::cout << "Training done. Time elapse: " << (end - start) / (1000 * 60) << "minute" << std::endl;

  // test the accuracy_rate in train
  if (1) {
    int corrects_all = 0, sum_all = train_images.size();
    std::cout << "train_images size: " << sum_all << std::endl;
    for (size_t i = 0; i < train_images.size(); ++i) {
      cv::Mat img = train_images.at(i);
      int label = train_label.at(i);
      auto char_key = kChars[label + kCharsTotalNumber - classNumber];
      std::pair<std::string, std::string> ch = identifyGrayChinese(img);
      if (ch.first == char_key)
        corrects_all++;
    }
    float accuracy_rate = (float)corrects_all / (float)sum_all;
    std::cout << "Train error_rate: " << (1.f - accuracy_rate) * 100.f << "% "<< std::endl;
  }

  // test the accuracy_rate in val
  if (1) {
    int corrects_all = 0, sum_all = val_images.size();
    std::cout << "val_images: " << sum_all << std::endl;
    for (size_t i = 0; i < val_images.size(); ++i) {
      cv::Mat img = val_images.at(i);
      int label = val_labels.at(i);
      auto char_key = kChars[label + kCharsTotalNumber - classNumber];
      std::pair<std::string, std::string> ch = identifyGrayChinese(img);
      if (ch.first == char_key)
        corrects_all++;
    }
    float accuracy_rate = (float)corrects_all / (float)sum_all;
    std::cout << "Test error_rate: " << (1.f - accuracy_rate) * 100.f << "% "<< std::endl;
  }

}

cv::Ptr<cv::ml::TrainData> AnnChTrain::tdata() {
  assert(chars_folder_);

  cv::Mat samples;
  std::vector<int> labels;

  std::cout << "Collecting chars in " << chars_folder_ << std::endl;

  int classNumber = 0;
  if (type == 0) classNumber = kCharsTotalNumber;
  if (type == 1) classNumber = kChineseNumber;

  for (int i = 0; i < classNumber; ++i) {
    auto char_key = kChars[i + kCharsTotalNumber - classNumber];
    char sub_folder[512] = {0};

    sprintf(sub_folder, "%s/%s", chars_folder_, char_key);
    std::cout << "  >> Featuring characters " << char_key << " in "
              << sub_folder << std::endl;

    auto chars_files = utils::getFiles(sub_folder);
    for (auto file : chars_files) {
      auto img = cv::imread(file, 0);  // a grayscale image
      auto fps = charFeatures2(img, kPredictSize);

      samples.push_back(fps);
      labels.push_back(i);
    }
  }

  cv::Mat samples_;
  samples.convertTo(samples_, CV_32F);
  cv::Mat train_classes =
    cv::Mat::zeros((int)labels.size(), classNumber, CV_32F);

  for (int i = 0; i < train_classes.rows; ++i) {
    train_classes.at<float>(i, labels[i]) = 1.f;
  }

  return cv::ml::TrainData::create(samples_, cv::ml::SampleTypes::ROW_SAMPLE,
                                   train_classes);
}
}