tesseract::TrainingSampleSet Class Reference

#include <trainingsampleset.h>

Public Member Functions
	TrainingSampleSet (const FontInfoTable &fontinfo_table)
	~TrainingSampleSet ()
bool	Serialize (FILE *fp) const
bool	DeSerialize (bool swap, FILE *fp)
int	num_samples () const
int	num_raw_samples () const
int	NumFonts () const
const UNICHARSET &	unicharset () const
int	charsetsize () const
const FontInfoTable &	fontinfo_table () const
void	LoadUnicharset (const char *filename)
int	AddSample (const char *unichar, TrainingSample *sample)
void	AddSample (int unichar_id, TrainingSample *sample)
int	NumClassSamples (int font_id, int class_id, bool randomize) const
const TrainingSample *	GetSample (int index) const
const TrainingSample *	GetSample (int font_id, int class_id, int index) const
TrainingSample *	MutableSample (int font_id, int class_id, int index)
STRING	SampleToString (const TrainingSample &sample) const
const BitVector &	GetCloudFeatures (int font_id, int class_id) const
const GenericVector< int > &	GetCanonicalFeatures (int font_id, int class_id) const
float	UnicharDistance (const UnicharAndFonts &uf1, const UnicharAndFonts &uf2, bool matched_fonts, const IntFeatureMap &feature_map)
float	ClusterDistance (int font_id1, int class_id1, int font_id2, int class_id2, const IntFeatureMap &feature_map)
float	ComputeClusterDistance (int font_id1, int class_id1, int font_id2, int class_id2, const IntFeatureMap &feature_map) const
int	ReliablySeparable (int font_id1, int class_id1, int font_id2, int class_id2, const IntFeatureMap &feature_map, bool thorough) const
int	GlobalSampleIndex (int font_id, int class_id, int index) const
const TrainingSample *	GetCanonicalSample (int font_id, int class_id) const
float	GetCanonicalDist (int font_id, int class_id) const
TrainingSample *	mutable_sample (int index)
TrainingSample *	extract_sample (int index)
void	IndexFeatures (const IntFeatureSpace &feature_space)
void	KillSample (TrainingSample *sample)
void	DeleteDeadSamples ()
bool	DeleteableSample (const TrainingSample *sample)
void	OrganizeByFontAndClass ()
void	SetupFontIdMap ()
void	ComputeCanonicalSamples (const IntFeatureMap &map, bool debug)
void	ReplicateAndRandomizeSamples ()
void	ComputeCanonicalFeatures ()
void	ComputeCloudFeatures (int feature_space_size)
void	AddAllFontsForClass (int class_id, Shape *shape) const
void	DisplaySamplesWithFeature (int f_index, const Shape &shape, const IntFeatureSpace &feature_space, ScrollView::Color color, ScrollView *window) const

Detailed Description

Definition at line 43 of file trainingsampleset.h.

Constructor & Destructor Documentation

TrainingSampleSet()

tesseract::TrainingSampleSet::TrainingSampleSet ( const FontInfoTable &  fontinfo_table )

explicit

Definition at line 70 of file trainingsampleset.cpp.

  : num_raw_samples_(0), unicharset_size_(0),
    font_class_array_(nullptr), fontinfo_table_(font_table) {
}

~TrainingSampleSet()

tesseract::TrainingSampleSet::~TrainingSampleSet

(

)

Definition at line 75 of file trainingsampleset.cpp.

                                      {
  delete font_class_array_;
}

Member Function Documentation

AddAllFontsForClass()

void tesseract::TrainingSampleSet::AddAllFontsForClass	(	int	class_id,
		Shape *	shape
	)		const

Definition at line 734 of file trainingsampleset.cpp.

                                                                            {
  for (int f = 0; f < font_id_map_.CompactSize(); ++f) {
    const int font_id = font_id_map_.CompactToSparse(f);
    shape->AddToShape(class_id, font_id);
  }
}

AddSample() [1/2]

int tesseract::TrainingSampleSet::AddSample	(	const char *	unichar,
		TrainingSample *	sample
	)

Definition at line 129 of file trainingsampleset.cpp.

                                                                            {
  if (!unicharset_.contains_unichar(unichar)) {
    unicharset_.unichar_insert(unichar);
    if (unicharset_.size() > MAX_NUM_CLASSES) {
      tprintf("Error: Size of unicharset in TrainingSampleSet::AddSample is "
              "greater than MAX_NUM_CLASSES\n");
      return -1;
    }
  }
  UNICHAR_ID char_id = unicharset_.unichar_to_id(unichar);
  AddSample(char_id, sample);
  return char_id;
}

AddSample() [2/2]

void tesseract::TrainingSampleSet::AddSample	(	int	unichar_id,
		TrainingSample *	sample
	)

Definition at line 145 of file trainingsampleset.cpp.

                                                                        {
  sample->set_class_id(unichar_id);
  samples_.push_back(sample);
  num_raw_samples_ = samples_.size();
  unicharset_size_ = unicharset_.size();
}

charsetsize()

int tesseract::TrainingSampleSet::charsetsize ( ) const

inline

Definition at line 67 of file trainingsampleset.h.

                          {
    return unicharset_size_;
  }

ClusterDistance()

float tesseract::TrainingSampleSet::ClusterDistance	(	int	font_id1,
		int	class_id1,
		int	font_id2,
		int	class_id2,
		const IntFeatureMap &	feature_map
	)

Definition at line 296 of file trainingsampleset.cpp.

                                                                           {
  ASSERT_HOST(font_class_array_ != nullptr);
  int font_index1 = font_id_map_.SparseToCompact(font_id1);
  int font_index2 = font_id_map_.SparseToCompact(font_id2);
  if (font_index1 < 0 || font_index2 < 0)
    return 0.0f;
  FontClassInfo& fc_info = (*font_class_array_)(font_index1, class_id1);
  if (font_id1 == font_id2) {
    // Special case cache for speed.
    if (fc_info.unichar_distance_cache.size() == 0)
      fc_info.unichar_distance_cache.init_to_size(unicharset_size_, -1.0f);
    if (fc_info.unichar_distance_cache[class_id2] < 0) {
      // Distance has to be calculated.
      float result = ComputeClusterDistance(font_id1, class_id1,
                                            font_id2, class_id2,
                                            feature_map);
      fc_info.unichar_distance_cache[class_id2] = result;
      // Copy to the symmetric cache entry.
      FontClassInfo& fc_info2 = (*font_class_array_)(font_index2, class_id2);
      if (fc_info2.unichar_distance_cache.size() == 0)
        fc_info2.unichar_distance_cache.init_to_size(unicharset_size_, -1.0f);
      fc_info2.unichar_distance_cache[class_id1] = result;
    }
    return fc_info.unichar_distance_cache[class_id2];
  } else if (class_id1 == class_id2) {
    // Another special-case cache for equal class-id.
    if (fc_info.font_distance_cache.size() == 0)
      fc_info.font_distance_cache.init_to_size(font_id_map_.CompactSize(),
                                               -1.0f);
    if (fc_info.font_distance_cache[font_index2] < 0) {
      // Distance has to be calculated.
      float result = ComputeClusterDistance(font_id1, class_id1,
                                            font_id2, class_id2,
                                            feature_map);
      fc_info.font_distance_cache[font_index2] = result;
      // Copy to the symmetric cache entry.
      FontClassInfo& fc_info2 = (*font_class_array_)(font_index2, class_id2);
      if (fc_info2.font_distance_cache.size() == 0)
        fc_info2.font_distance_cache.init_to_size(font_id_map_.CompactSize(),
                                                  -1.0f);
      fc_info2.font_distance_cache[font_index1] = result;
    }
    return fc_info.font_distance_cache[font_index2];
  }
  // Both font and class are different. Linear search for class_id2/font_id2
  // in what is a hopefully short list of distances.
  int cache_index = 0;
  while (cache_index < fc_info.distance_cache.size() &&
         (fc_info.distance_cache[cache_index].unichar_id != class_id2 ||
          fc_info.distance_cache[cache_index].font_id != font_id2))
    ++cache_index;
  if (cache_index == fc_info.distance_cache.size()) {
    // Distance has to be calculated.
    float result = ComputeClusterDistance(font_id1, class_id1,
                                          font_id2, class_id2,
                                          feature_map);
    FontClassDistance fc_dist = { class_id2, font_id2, result };
    fc_info.distance_cache.push_back(fc_dist);
    // Copy to the symmetric cache entry. We know it isn't there already, as
    // we always copy to the symmetric entry.
    FontClassInfo& fc_info2 = (*font_class_array_)(font_index2, class_id2);
    fc_dist.unichar_id = class_id1;
    fc_dist.font_id = font_id1;
    fc_info2.distance_cache.push_back(fc_dist);
  }
  return fc_info.distance_cache[cache_index].distance;
}

ComputeCanonicalFeatures()

void tesseract::TrainingSampleSet::ComputeCanonicalFeatures

(

)

Definition at line 694 of file trainingsampleset.cpp.

                                                 {
  ASSERT_HOST(font_class_array_ != nullptr);
  const int font_size = font_id_map_.CompactSize();
  for (int font_index = 0; font_index < font_size; ++font_index) {
    const int font_id = font_id_map_.CompactToSparse(font_index);
    for (int c = 0; c < unicharset_size_; ++c) {
      int num_samples = NumClassSamples(font_id, c, false);
      if (num_samples == 0)
        continue;
      const TrainingSample* sample = GetCanonicalSample(font_id, c);
      FontClassInfo& fcinfo = (*font_class_array_)(font_index, c);
      fcinfo.canonical_features = sample->indexed_features();
    }
  }
}

ComputeCanonicalSamples()

void tesseract::TrainingSampleSet::ComputeCanonicalSamples	(	const IntFeatureMap &	map,
		bool	debug
	)

Definition at line 568 of file trainingsampleset.cpp.

                                                            {
  ASSERT_HOST(font_class_array_ != nullptr);
  IntFeatureDist f_table;
  if (debug) tprintf("feature table size %d\n", map.sparse_size());
  f_table.Init(&map);
  int worst_s1 = 0;
  int worst_s2 = 0;
  double global_worst_dist = 0.0;
  // Compute distances independently for each font and char index.
  int font_size = font_id_map_.CompactSize();
  for (int font_index = 0; font_index < font_size; ++font_index) {
    int font_id = font_id_map_.CompactToSparse(font_index);
    for (int c = 0; c < unicharset_size_; ++c) {
      int samples_found = 0;
      FontClassInfo& fcinfo = (*font_class_array_)(font_index, c);
      if (fcinfo.samples.size() == 0 ||
          (kTestChar >= 0 && c != kTestChar)) {
        fcinfo.canonical_sample = -1;
        fcinfo.canonical_dist = 0.0f;
        if (debug) tprintf("Skipping class %d\n", c);
        continue;
      }
      // The canonical sample will be the one with the min_max_dist, which
      // is the sample with the lowest maximum distance to all other samples.
      double min_max_dist = 2.0;
      // We keep track of the farthest apart pair (max_s1, max_s2) which
      // are max_max_dist apart, so we can see how bad the variability is.
      double max_max_dist = 0.0;
      int max_s1 = 0;
      int max_s2 = 0;
      fcinfo.canonical_sample = fcinfo.samples[0];
      fcinfo.canonical_dist = 0.0f;
      for (int i = 0; i < fcinfo.samples.size(); ++i) {
        int s1 = fcinfo.samples[i];
        const GenericVector<int>& features1 = samples_[s1]->indexed_features();
        f_table.Set(features1, features1.size(), true);
        double max_dist = 0.0;
        // Run the full squared-order search for similar samples. It is still
        // reasonably fast because f_table.FeatureDistance is fast, but we
        // may have to reconsider if we start playing with too many samples
        // of a single char/font.
        for (int j = 0; j < fcinfo.samples.size(); ++j) {
          int s2 = fcinfo.samples[j];
          if (samples_[s2]->class_id() != c  ||
              samples_[s2]->font_id() != font_id ||
              s2 == s1)
            continue;
          GenericVector<int> features2 = samples_[s2]->indexed_features();
          double dist = f_table.FeatureDistance(features2);
          if (dist > max_dist) {
            max_dist = dist;
            if (dist > max_max_dist) {
              max_max_dist = dist;
              max_s1 = s1;
              max_s2 = s2;
            }
          }
        }
        // Using Set(..., false) is far faster than re initializing, due to
        // the sparseness of the feature space.
        f_table.Set(features1, features1.size(), false);
        samples_[s1]->set_max_dist(max_dist);
        ++samples_found;
        if (max_dist < min_max_dist) {
          fcinfo.canonical_sample = s1;
          fcinfo.canonical_dist = max_dist;
        }
        UpdateRange(max_dist, &min_max_dist, &max_max_dist);
      }
      if (max_max_dist > global_worst_dist) {
        // Keep a record of the worst pair over all characters/fonts too.
        global_worst_dist = max_max_dist;
        worst_s1 = max_s1;
        worst_s2 = max_s2;
      }
      if (debug) {
        tprintf("Found %d samples of class %d=%s, font %d, "
                "dist range [%g, %g], worst pair= %s, %s\n",
                samples_found, c, unicharset_.debug_str(c).c_str(),
                font_index, min_max_dist, max_max_dist,
                SampleToString(*samples_[max_s1]).c_str(),
                SampleToString(*samples_[max_s2]).c_str());
      }
    }
  }
  if (debug) {
    tprintf("Global worst dist = %g, between sample %d and %d\n",
            global_worst_dist, worst_s1, worst_s2);
  }
}

ComputeCloudFeatures()

void tesseract::TrainingSampleSet::ComputeCloudFeatures

(

int

feature_space_size

)

Definition at line 712 of file trainingsampleset.cpp.

                                                                   {
  ASSERT_HOST(font_class_array_ != nullptr);
  int font_size = font_id_map_.CompactSize();
  for (int font_index = 0; font_index < font_size; ++font_index) {
    int font_id = font_id_map_.CompactToSparse(font_index);
    for (int c = 0; c < unicharset_size_; ++c) {
      int num_samples = NumClassSamples(font_id, c, false);
      if (num_samples == 0)
        continue;
      FontClassInfo& fcinfo = (*font_class_array_)(font_index, c);
      fcinfo.cloud_features.Init(feature_space_size);
      for (int s = 0; s < num_samples; ++s) {
        const TrainingSample* sample = GetSample(font_id, c, s);
        const GenericVector<int>& sample_features = sample->indexed_features();
        for (int i = 0; i < sample_features.size(); ++i)
          fcinfo.cloud_features.SetBit(sample_features[i]);
      }
    }
  }
}

ComputeClusterDistance()

float tesseract::TrainingSampleSet::ComputeClusterDistance	(	int	font_id1,
		int	class_id1,
		int	font_id2,
		int	class_id2,
		const IntFeatureMap &	feature_map
	)		const

Definition at line 367 of file trainingsampleset.cpp.

                                            {
  int dist = ReliablySeparable(font_id1, class_id1, font_id2, class_id2,
                               feature_map, false);
  dist += ReliablySeparable(font_id2, class_id2, font_id1, class_id1,
                            feature_map, false);
  int denominator = GetCanonicalFeatures(font_id1, class_id1).size();
  denominator += GetCanonicalFeatures(font_id2, class_id2).size();
  return static_cast<float>(dist) / denominator;
}

DeleteableSample()

bool tesseract::TrainingSampleSet::DeleteableSample

(

const TrainingSample *

sample

)

Definition at line 506 of file trainingsampleset.cpp.

                                                                     {
  return sample == nullptr || sample->class_id() < 0;
}

DeleteDeadSamples()

void tesseract::TrainingSampleSet::DeleteDeadSamples

(

)

Definition at line 497 of file trainingsampleset.cpp.

                                          {
  using namespace std::placeholders;  // for _1
  samples_.compact(std::bind(&TrainingSampleSet::DeleteableSample, this, _1));
  num_raw_samples_ = samples_.size();
  // Samples must be re-organized now we have deleted a few.
}

DeSerialize()

bool tesseract::TrainingSampleSet::DeSerialize	(	bool	swap,
		FILE *	fp
	)

Definition at line 94 of file trainingsampleset.cpp.

                                                       {
  if (!samples_.DeSerialize(swap, fp)) return false;
  num_raw_samples_ = samples_.size();
  if (!unicharset_.load_from_file(fp)) return false;
  if (!font_id_map_.DeSerialize(swap, fp)) return false;
  delete font_class_array_;
  font_class_array_ = nullptr;
  int8_t not_null;
  if (fread(&not_null, sizeof(not_null), 1, fp) != 1) return false;
  if (not_null) {
    FontClassInfo empty;
    font_class_array_ = new GENERIC_2D_ARRAY<FontClassInfo >(1, 1 , empty);
    if (!font_class_array_->DeSerializeClasses(swap, fp)) return false;
  }
  unicharset_size_ = unicharset_.size();
  return true;
}

DisplaySamplesWithFeature()

void tesseract::TrainingSampleSet::DisplaySamplesWithFeature	(	int	f_index,
		const Shape &	shape,
		const IntFeatureSpace &	feature_space,
		ScrollView::Color	color,
		ScrollView *	window
	)		const

Definition at line 743 of file trainingsampleset.cpp.

                                                                            {
  for (int s = 0; s < num_raw_samples(); ++s) {
    const TrainingSample* sample = GetSample(s);
    if (shape.ContainsUnichar(sample->class_id())) {
      GenericVector<int> indexed_features;
      space.IndexAndSortFeatures(sample->features(), sample->num_features(),
                                 &indexed_features);
      for (int f = 0; f < indexed_features.size(); ++f) {
        if (indexed_features[f] == f_index) {
          sample->DisplayFeatures(color, window);
        }
      }
    }
  }
}

extract_sample()

TrainingSample* tesseract::TrainingSampleSet::extract_sample ( int  index )

inline

Definition at line 165 of file trainingsampleset.h.

                                            {
    TrainingSample* sample = samples_[index];
    samples_[index] = nullptr;
    return sample;
  }

fontinfo_table()

const FontInfoTable& tesseract::TrainingSampleSet::fontinfo_table ( ) const

inline

Definition at line 70 of file trainingsampleset.h.

                                              {
    return fontinfo_table_;
  }

GetCanonicalDist()

float tesseract::TrainingSampleSet::GetCanonicalDist	(	int	font_id,
		int	class_id
	)		const

Definition at line 474 of file trainingsampleset.cpp.

                                                                         {
  ASSERT_HOST(font_class_array_ != nullptr);
  int font_index = font_id_map_.SparseToCompact(font_id);
  if (font_index < 0) return 0.0f;
  if ((*font_class_array_)(font_index, class_id).canonical_sample >= 0)
    return (*font_class_array_)(font_index, class_id).canonical_dist;
  else
    return 0.0f;
}

GetCanonicalFeatures()

const GenericVector< int > & tesseract::TrainingSampleSet::GetCanonicalFeatures	(	int	font_id,
		int	class_id
	)		const

Definition at line 219 of file trainingsampleset.cpp.

                                     {
  int font_index = font_id_map_.SparseToCompact(font_id);
  ASSERT_HOST(font_index >= 0);
  return (*font_class_array_)(font_index, class_id).canonical_features;
}

GetCanonicalSample()

const TrainingSample * tesseract::TrainingSampleSet::GetCanonicalSample	(	int	font_id,
		int	class_id
	)		const

Definition at line 462 of file trainingsampleset.cpp.

                                     {
  ASSERT_HOST(font_class_array_ != nullptr);
  int font_index = font_id_map_.SparseToCompact(font_id);
  if (font_index < 0) return nullptr;
  const int sample_index = (*font_class_array_)(font_index,
                                          class_id).canonical_sample;
  return sample_index >= 0 ? samples_[sample_index] : nullptr;
}

GetCloudFeatures()

const BitVector & tesseract::TrainingSampleSet::GetCloudFeatures	(	int	font_id,
		int	class_id
	)		const

Definition at line 211 of file trainingsampleset.cpp.

                                     {
  int font_index = font_id_map_.SparseToCompact(font_id);
  ASSERT_HOST(font_index >= 0);
  return (*font_class_array_)(font_index, class_id).cloud_features;
}

GetSample() [1/2]

const TrainingSample * tesseract::TrainingSampleSet::GetSample	(	int	font_id,
		int	class_id,
		int	index
	)		const

Definition at line 180 of file trainingsampleset.cpp.

                                                                    {
  ASSERT_HOST(font_class_array_ != nullptr);
  int font_index = font_id_map_.SparseToCompact(font_id);
  if (font_index < 0) return nullptr;
  int sample_index = (*font_class_array_)(font_index, class_id).samples[index];
  return samples_[sample_index];
}

GetSample() [2/2]

const TrainingSample * tesseract::TrainingSampleSet::GetSample

(

int

index

)

const

Definition at line 174 of file trainingsampleset.cpp.

                                                                  {
  return samples_[index];
}

GlobalSampleIndex()

int tesseract::TrainingSampleSet::GlobalSampleIndex	(	int	font_id,
		int	class_id,
		int	index
	)		const

Definition at line 452 of file trainingsampleset.cpp.

                                                          {
  ASSERT_HOST(font_class_array_ != nullptr);
  int font_index = font_id_map_.SparseToCompact(font_id);
  if (font_index < 0) return -1;
  return (*font_class_array_)(font_index, class_id).samples[index];
}

IndexFeatures()

void tesseract::TrainingSampleSet::IndexFeatures

(

const IntFeatureSpace &

feature_space

)

Definition at line 485 of file trainingsampleset.cpp.

                                                                          {
  for (int s = 0; s < samples_.size(); ++s)
    samples_[s]->IndexFeatures(feature_space);
}

KillSample()

void tesseract::TrainingSampleSet::KillSample

(

TrainingSample *

sample

)

Definition at line 492 of file trainingsampleset.cpp.

                                                         {
  sample->set_sample_index(-1);
}

LoadUnicharset()

void tesseract::TrainingSampleSet::LoadUnicharset

(

const char *

filename

)

Definition at line 113 of file trainingsampleset.cpp.

                                                           {
  if (!unicharset_.load_from_file(filename)) {
    tprintf("Failed to load unicharset from file %s\n"
            "Building unicharset from scratch...\n",
            filename);
    unicharset_.clear();
    // Add special characters as they were removed by the clear.
    UNICHARSET empty;
    unicharset_.AppendOtherUnicharset(empty);
  }
  unicharset_size_ = unicharset_.size();
}

mutable_sample()

TrainingSample* tesseract::TrainingSampleSet::mutable_sample ( int  index )

inline

Definition at line 161 of file trainingsampleset.h.

                                            {
    return samples_[index];
  }

MutableSample()

TrainingSample * tesseract::TrainingSampleSet::MutableSample	(	int	font_id,
		int	class_id,
		int	index
	)

Definition at line 191 of file trainingsampleset.cpp.

                                                            {
  ASSERT_HOST(font_class_array_ != nullptr);
  int font_index = font_id_map_.SparseToCompact(font_id);
  if (font_index < 0) return nullptr;
  int sample_index = (*font_class_array_)(font_index, class_id).samples[index];
  return samples_[sample_index];
}

num_raw_samples()

int tesseract::TrainingSampleSet::num_raw_samples ( ) const

inline

Definition at line 58 of file trainingsampleset.h.

                              {
    return num_raw_samples_;
  }

num_samples()

int tesseract::TrainingSampleSet::num_samples ( ) const

inline

Definition at line 55 of file trainingsampleset.h.

                          {
    return samples_.size();
  }

NumClassSamples()

int tesseract::TrainingSampleSet::NumClassSamples	(	int	font_id,
		int	class_id,
		bool	randomize
	)		const

Definition at line 156 of file trainingsampleset.cpp.

                                                             {
  ASSERT_HOST(font_class_array_ != nullptr);
  if (font_id < 0 || class_id < 0 ||
      font_id >= font_id_map_.SparseSize() || class_id >= unicharset_size_) {
    // There are no samples because the font or class doesn't exist.
    return 0;
  }
  int font_index = font_id_map_.SparseToCompact(font_id);
  if (font_index < 0)
    return 0;  // The font has no samples.
  if (randomize)
    return (*font_class_array_)(font_index, class_id).samples.size();
  else
    return (*font_class_array_)(font_index, class_id).num_raw_samples;
}

NumFonts()

int tesseract::TrainingSampleSet::NumFonts ( ) const

inline

Definition at line 61 of file trainingsampleset.h.

                       {
    return font_id_map_.SparseSize();
  }

OrganizeByFontAndClass()

void tesseract::TrainingSampleSet::OrganizeByFontAndClass

(

)

Definition at line 511 of file trainingsampleset.cpp.

                                               {
  // Font indexes are sparse, so we used a map to compact them, so we can
  // have an efficient 2-d array of fonts and character classes.
  SetupFontIdMap();
  int compact_font_size = font_id_map_.CompactSize();
  // Get a 2-d array of generic vectors.
  delete font_class_array_;
  FontClassInfo empty;
  font_class_array_ = new GENERIC_2D_ARRAY<FontClassInfo>(
      compact_font_size, unicharset_size_, empty);
  for (int s = 0; s < samples_.size(); ++s) {
    int font_id = samples_[s]->font_id();
    int class_id = samples_[s]->class_id();
    if (font_id < 0 || font_id >= font_id_map_.SparseSize()) {
      tprintf("Font id = %d/%d, class id = %d/%d on sample %d\n",
              font_id, font_id_map_.SparseSize(), class_id, unicharset_size_,
              s);
    }
    ASSERT_HOST(font_id >= 0 && font_id < font_id_map_.SparseSize());
    ASSERT_HOST(class_id >= 0 && class_id < unicharset_size_);
    int font_index = font_id_map_.SparseToCompact(font_id);
    (*font_class_array_)(font_index, class_id).samples.push_back(s);
  }
  // Set the num_raw_samples member of the FontClassInfo, to set the boundary
  // between the raw samples and the replicated ones.
  for (int f = 0; f < compact_font_size; ++f) {
    for (int c = 0; c < unicharset_size_; ++c)
      (*font_class_array_)(f, c).num_raw_samples =
          (*font_class_array_)(f, c).samples.size();
  }
  // This is the global number of samples and also marks the boundary between
  // real and replicated samples.
  num_raw_samples_ = samples_.size();
}

ReliablySeparable()

int tesseract::TrainingSampleSet::ReliablySeparable	(	int	font_id1,
		int	class_id1,
		int	font_id2,
		int	class_id2,
		const IntFeatureMap &	feature_map,
		bool	thorough
	)		const

Definition at line 413 of file trainingsampleset.cpp.

                                                              {
  int result = 0;
  const TrainingSample* sample2 = GetCanonicalSample(font_id2, class_id2);
  if (sample2 == nullptr)
    return 0;  // There are no canonical features.
  const GenericVector<int>& canonical2 = GetCanonicalFeatures(font_id2,
                                                              class_id2);
  const BitVector& cloud1 = GetCloudFeatures(font_id1, class_id1);
  if (cloud1.size() == 0)
    return canonical2.size();  // There are no cloud features.
 
  // Find a canonical2 feature that is not in cloud1.
  for (int f = 0; f < canonical2.size(); ++f) {
    const int feature = canonical2[f];
    if (cloud1[feature])
      continue;
    // Gather the near neighbours of f.
    GenericVector<int> good_features;
    AddNearFeatures(feature_map, feature, 1, &good_features);
    // Check that none of the good_features are in the cloud.
    int i;
    for (i = 0; i < good_features.size(); ++i) {
      int good_f = good_features[i];
      if (cloud1[good_f]) {
        break;
      }
    }
    if (i < good_features.size())
       continue;  // Found one in the cloud.
    ++result;
  }
  return result;
}

ReplicateAndRandomizeSamples()

void tesseract::TrainingSampleSet::ReplicateAndRandomizeSamples

(

)

Definition at line 665 of file trainingsampleset.cpp.

                                                     {
  ASSERT_HOST(font_class_array_ != nullptr);
  int font_size = font_id_map_.CompactSize();
  for (int font_index = 0; font_index < font_size; ++font_index) {
    for (int c = 0; c < unicharset_size_; ++c) {
      FontClassInfo& fcinfo = (*font_class_array_)(font_index, c);
      int sample_count = fcinfo.samples.size();
      int min_samples = 2 * std::max(kSampleRandomSize, sample_count);
      if (sample_count > 0 && sample_count < min_samples) {
        int base_count = sample_count;
        for (int base_index = 0; sample_count < min_samples; ++sample_count) {
          int src_index = fcinfo.samples[base_index++];
          if (base_index >= base_count) base_index = 0;
          TrainingSample* sample = samples_[src_index]->RandomizedCopy(
              sample_count % kSampleRandomSize);
          int sample_index = samples_.size();
          sample->set_sample_index(sample_index);
          samples_.push_back(sample);
          fcinfo.samples.push_back(sample_index);
        }
      }
    }
  }
}

SampleToString()

STRING tesseract::TrainingSampleSet::SampleToString

(

const TrainingSample &

sample

)

const

Definition at line 202 of file trainingsampleset.cpp.

                                                                           {
  STRING boxfile_str;
  MakeBoxFileStr(unicharset_.id_to_unichar(sample.class_id()),
                 sample.bounding_box(), sample.page_num(), &boxfile_str);
  return STRING(fontinfo_table_.get(sample.font_id()).name) + " " + boxfile_str;
}

Serialize()

bool tesseract::TrainingSampleSet::Serialize

(

FILE *

fp

)

const

Definition at line 80 of file trainingsampleset.cpp.

                                                {
  if (!samples_.Serialize(fp)) return false;
  if (!unicharset_.save_to_file(fp)) return false;
  if (!font_id_map_.Serialize(fp)) return false;
  int8_t not_null = font_class_array_ != nullptr;
  if (fwrite(&not_null, sizeof(not_null), 1, fp) != 1) return false;
  if (not_null) {
    if (!font_class_array_->SerializeClasses(fp)) return false;
  }
  return true;
}

SetupFontIdMap()

void tesseract::TrainingSampleSet::SetupFontIdMap

(

)

Definition at line 548 of file trainingsampleset.cpp.

                                       {
  // Number of samples for each font_id.
  GenericVector<int> font_counts;
  for (int s = 0; s < samples_.size(); ++s) {
    const int font_id = samples_[s]->font_id();
    while (font_id >= font_counts.size())
      font_counts.push_back(0);
    ++font_counts[font_id];
  }
  font_id_map_.Init(font_counts.size(), false);
  for (int f = 0; f < font_counts.size(); ++f) {
    font_id_map_.SetMap(f, font_counts[f] > 0);
  }
  font_id_map_.Setup();
}

UnicharDistance()

float tesseract::TrainingSampleSet::UnicharDistance	(	const UnicharAndFonts &	uf1,
		const UnicharAndFonts &	uf2,
		bool	matched_fonts,
		const IntFeatureMap &	feature_map
	)

Definition at line 230 of file trainingsampleset.cpp.

                                                                           {
  int num_fonts1 = uf1.font_ids.size();
  int c1 = uf1.unichar_id;
  int num_fonts2 = uf2.font_ids.size();
  int c2 = uf2.unichar_id;
  double dist_sum = 0.0;
  int dist_count = 0;
  const bool debug = false;
  if (matched_fonts) {
    // Compute distances only where fonts match.
    for (int i = 0; i < num_fonts1; ++i) {
      int f1 = uf1.font_ids[i];
      for (int j = 0; j < num_fonts2; ++j) {
        int f2 = uf2.font_ids[j];
        if (f1 == f2) {
          dist_sum += ClusterDistance(f1, c1, f2, c2, feature_map);
          ++dist_count;
        }
      }
    }
  } else if (num_fonts1 * num_fonts2 <= kSquareLimit) {
    // Small enough sets to compute all the distances.
    for (int i = 0; i < num_fonts1; ++i) {
      int f1 = uf1.font_ids[i];
      for (int j = 0; j < num_fonts2; ++j) {
        int f2 = uf2.font_ids[j];
        dist_sum += ClusterDistance(f1, c1, f2, c2, feature_map);
        if (debug) {
            tprintf("Cluster dist %d %d %d %d = %g\n",
                    f1, c1, f2, c2,
                    ClusterDistance(f1, c1, f2, c2, feature_map));
        }
        ++dist_count;
      }
    }
  } else {
    // Subsample distances, using the largest set once, and stepping through
    // the smaller set so as to ensure that all the pairs are different.
    int increment = kPrime1 != num_fonts2 ? kPrime1 : kPrime2;
    int index = 0;
    int num_samples = std::max(num_fonts1, num_fonts2);
    for (int i = 0; i < num_samples; ++i, index += increment) {
      int f1 = uf1.font_ids[i % num_fonts1];
      int f2 = uf2.font_ids[index % num_fonts2];
      if (debug) {
          tprintf("Cluster dist %d %d %d %d = %g\n",
                  f1, c1, f2, c2, ClusterDistance(f1, c1, f2, c2, feature_map));
      }
      dist_sum += ClusterDistance(f1, c1, f2, c2, feature_map);
      ++dist_count;
    }
  }
  if (dist_count == 0) {
    if (matched_fonts)
      return UnicharDistance(uf1, uf2, false, feature_map);
    return 0.0f;
  }
  return dist_sum / dist_count;
}

unicharset()

const UNICHARSET& tesseract::TrainingSampleSet::unicharset ( ) const

inline

Definition at line 64 of file trainingsampleset.h.

                                       {
    return unicharset_;
  }

---

The documentation for this class was generated from the following files:

• /tesseract/src/training/trainingsampleset.h
• /tesseract/src/training/trainingsampleset.cpp