adaptmatch.cpp

Go to the documentation of this file.

/******************************************************************************
 ** Filename:    adaptmatch.cpp
 ** Purpose:     High level adaptive matcher.
 ** Author:      Dan Johnson
 **
 ** (c) Copyright Hewlett-Packard Company, 1988.
 ** Licensed under the Apache License, Version 2.0 (the "License");
 ** you may not use this file except in compliance with the License.
 ** You may obtain a copy of the License at
 ** http://www.apache.org/licenses/LICENSE-2.0
 ** Unless required by applicable law or agreed to in writing, software
 ** distributed under the License is distributed on an "AS IS" BASIS,
 ** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 ** See the License for the specific language governing permissions and
 ** limitations under the License.
 ******************************************************************************/
 
/*-----------------------------------------------------------------------------
          Include Files and Type Defines
-----------------------------------------------------------------------------*/
#ifdef HAVE_CONFIG_H
#  include "config_auto.h"
#endif
 
#include "adaptive.h"        // for ADAPT_CLASS
#include "ambigs.h"          // for UnicharIdVector, UnicharAmbigs
#include "bitvec.h"          // for FreeBitVector, NewBitVector, BIT_VECTOR
#include "blobs.h"           // for TBLOB, TWERD
#include "classify.h"        // for Classify, CST_FRAGMENT, CST_WHOLE
#include "dict.h"            // for Dict
#include "errcode.h"         // for ASSERT_HOST
#include "featdefs.h"        // for CharNormDesc
#include "float2int.h"       // for BASELINE_Y_SHIFT
#include "fontinfo.h"        // for ScoredFont, FontSet
#include "intfx.h"           // for BlobToTrainingSample, INT_FX_RESULT_S...
#include "intmatcher.h"      // for CP_RESULT_STRUCT, IntegerMatcher
#include "intproto.h"        // for INT_FEATURE_STRUCT, (anonymous), Clas...
#include "matchdefs.h"       // for CLASS_ID, FEATURE_ID, PROTO_ID, NO_PROTO
#include "mfoutline.h"       // for baseline, character, MF_SCALE_FACTOR
#include "normalis.h"        // for DENORM, kBlnBaselineOffset, kBlnXHeight
#include "normfeat.h"        // for ActualOutlineLength, CharNormLength
#include "ocrfeatures.h"     // for FEATURE_STRUCT, FEATURE
#include "oldlist.h"         // for push, delete_d
#include "outfeat.h"         // for OutlineFeatDir, OutlineFeatLength
#include "pageres.h"         // for WERD_RES
#include "params.h"          // for IntParam, BoolParam, DoubleParam, Str...
#include "picofeat.h"        // for PicoFeatDir, PicoFeatX, PicoFeatY
#include "protos.h"          // for PROTO_STRUCT, FillABC
#include "ratngs.h"          // for BLOB_CHOICE_IT, BLOB_CHOICE_LIST, BLO...
#include "rect.h"            // for TBOX
#include "scrollview.h"      // for ScrollView, ScrollView::BROWN, Scroll...
#include "seam.h"            // for SEAM
#include "shapeclassifier.h" // for ShapeClassifier
#include "shapetable.h"      // for UnicharRating, ShapeTable, Shape, Uni...
#include "tessclassifier.h"  // for TessClassifier
#include "tessdatamanager.h" // for TessdataManager, TESSDATA_INTTEMP
#include "tprintf.h"         // for tprintf
#include "trainingsample.h"  // for TrainingSample
#include "unicharset.h"      // for UNICHARSET, CHAR_FRAGMENT, UNICHAR_SPACE
#include "unicity_table.h"   // for UnicityTable
 
#include <tesseract/unichar.h> // for UNICHAR_ID, INVALID_UNICHAR_ID
#include "helpers.h"           // for IntCastRounded, ClipToRange
#include "serialis.h"          // for TFile
 
#include <algorithm> // for max, min
#include <cassert>   // for assert
#include <cmath>     // for fabs
#include <cstdint>   // for INT32_MAX, UINT8_MAX
#include <cstdio>    // for fflush, fclose, fopen, stdout, FILE
#include <cstring>   // for strstr, memset, strcmp
 
namespace tesseract {
 
// TODO: The parameter classify_enable_adaptive_matcher can cause
// a segmentation fault if it is set to false (issue #256),
// so override it here.
#define classify_enable_adaptive_matcher true
 
#define ADAPT_TEMPLATE_SUFFIX ".a"
 
#define MAX_MATCHES 10
#define UNLIKELY_NUM_FEAT 200
#define NO_DEBUG 0
#define MAX_ADAPTABLE_WERD_SIZE 40
 
#define ADAPTABLE_WERD_ADJUSTMENT (0.05)
 
#define Y_DIM_OFFSET (Y_SHIFT - BASELINE_Y_SHIFT)
 
#define WORST_POSSIBLE_RATING (0.0f)
 
struct ADAPT_RESULTS {
  int32_t BlobLength;
  bool HasNonfragment;
  UNICHAR_ID best_unichar_id;
  int best_match_index;
  float best_rating;
  std::vector<UnicharRating> match;
  std::vector<CP_RESULT_STRUCT> CPResults;
 
  inline void Initialize() {
    BlobLength = INT32_MAX;
    HasNonfragment = false;
    ComputeBest();
  }
  // Computes best_unichar_id, best_match_index and best_rating.
  void ComputeBest() {
    best_unichar_id = INVALID_UNICHAR_ID;
    best_match_index = -1;
    best_rating = WORST_POSSIBLE_RATING;
    for (unsigned i = 0; i < match.size(); ++i) {
      if (match[i].rating > best_rating) {
        best_rating = match[i].rating;
        best_unichar_id = match[i].unichar_id;
        best_match_index = i;
      }
    }
  }
};
 
struct PROTO_KEY {
  ADAPT_TEMPLATES_STRUCT *Templates;
  CLASS_ID ClassId;
  int ConfigId;
};
 
// Sort function to sort ratings appropriately by descending rating.
static bool SortDescendingRating(const UnicharRating &a, const UnicharRating &b) {
  if (a.rating != b.rating) {
    return a.rating > b.rating;
  } else {
    return a.unichar_id < b.unichar_id;
  }
}
 
/*-----------------------------------------------------------------------------
          Private Macros
-----------------------------------------------------------------------------*/
inline bool MarginalMatch(float confidence, float matcher_great_threshold) {
  return (1.0f - confidence) > matcher_great_threshold;
}
 
/*-----------------------------------------------------------------------------
          Private Function Prototypes
-----------------------------------------------------------------------------*/
// Returns the index of the given id in results, if present, or the size of the
// vector (index it will go at) if not present.
static unsigned FindScoredUnichar(UNICHAR_ID id, const ADAPT_RESULTS &results) {
  for (unsigned i = 0; i < results.match.size(); i++) {
    if (results.match[i].unichar_id == id) {
      return i;
    }
  }
  return results.match.size();
}
 
// Returns the current rating for a unichar id if we have rated it, defaulting
// to WORST_POSSIBLE_RATING.
static float ScoredUnichar(UNICHAR_ID id, const ADAPT_RESULTS &results) {
  unsigned index = FindScoredUnichar(id, results);
  if (index >= results.match.size()) {
    return WORST_POSSIBLE_RATING;
  }
  return results.match[index].rating;
}
 
void InitMatcherRatings(float *Rating);
 
int MakeTempProtoPerm(void *item1, void *item2);
 
void SetAdaptiveThreshold(float Threshold);
 
/*-----------------------------------------------------------------------------
              Public Code
-----------------------------------------------------------------------------*/
void Classify::AdaptiveClassifier(TBLOB *Blob, BLOB_CHOICE_LIST *Choices) {
  assert(Choices != nullptr);
  auto *Results = new ADAPT_RESULTS;
  Results->Initialize();
 
  ASSERT_HOST(AdaptedTemplates != nullptr);
 
  DoAdaptiveMatch(Blob, Results);
 
  RemoveBadMatches(Results);
  std::sort(Results->match.begin(), Results->match.end(), SortDescendingRating);
  RemoveExtraPuncs(Results);
  Results->ComputeBest();
  ConvertMatchesToChoices(Blob->denorm(), Blob->bounding_box(), Results, Choices);
 
  // TODO(rays) Move to before ConvertMatchesToChoices!
  if (LargeSpeckle(*Blob) || Choices->empty()) {
    AddLargeSpeckleTo(Results->BlobLength, Choices);
  }
 
  if (matcher_debug_level >= 1) {
    tprintf("AD Matches =  ");
    PrintAdaptiveMatchResults(*Results);
  }
 
#ifndef GRAPHICS_DISABLED
  if (classify_enable_adaptive_debugger) {
    DebugAdaptiveClassifier(Blob, Results);
  }
#endif
 
  delete Results;
} /* AdaptiveClassifier */
 
#ifndef GRAPHICS_DISABLED
 
// If *win is nullptr, sets it to a new ScrollView() object with title msg.
// Clears the window and draws baselines.
void Classify::RefreshDebugWindow(ScrollView **win, const char *msg, int y_offset,
                                  const TBOX &wbox) {
  const int kSampleSpaceWidth = 500;
  if (*win == nullptr) {
    *win = new ScrollView(msg, 100, y_offset, kSampleSpaceWidth * 2, 200, kSampleSpaceWidth * 2,
                          200, true);
  }
  (*win)->Clear();
  (*win)->Pen(64, 64, 64);
  (*win)->Line(-kSampleSpaceWidth, kBlnBaselineOffset, kSampleSpaceWidth, kBlnBaselineOffset);
  (*win)->Line(-kSampleSpaceWidth, kBlnXHeight + kBlnBaselineOffset, kSampleSpaceWidth,
               kBlnXHeight + kBlnBaselineOffset);
  (*win)->ZoomToRectangle(wbox.left(), wbox.top(), wbox.right(), wbox.bottom());
}
 
#endif // !GRAPHICS_DISABLED
 
// Learns the given word using its chopped_word, seam_array, denorm,
// box_word, best_state, and correct_text to learn both correctly and
// incorrectly segmented blobs. If fontname is not nullptr, then LearnBlob
// is called and the data will be saved in an internal buffer.
// Otherwise AdaptToBlob is called for adaption within a document.
void Classify::LearnWord(const char *fontname, WERD_RES *word) {
  int word_len = word->correct_text.size();
  if (word_len == 0) {
    return;
  }
 
  float *thresholds = nullptr;
  if (fontname == nullptr) {
    // Adaption mode.
    if (!EnableLearning || word->best_choice == nullptr) {
      return; // Can't or won't adapt.
    }
 
    if (classify_learning_debug_level >= 1) {
      tprintf("\n\nAdapting to word = %s\n", word->best_choice->debug_string().c_str());
    }
    thresholds = new float[word_len];
    word->ComputeAdaptionThresholds(getDict().certainty_scale, matcher_perfect_threshold,
                                    matcher_good_threshold, matcher_rating_margin, thresholds);
  }
  int start_blob = 0;
 
#ifndef GRAPHICS_DISABLED
  if (classify_debug_character_fragments) {
    if (learn_fragmented_word_debug_win_ != nullptr) {
      learn_fragmented_word_debug_win_->Wait();
    }
    RefreshDebugWindow(&learn_fragments_debug_win_, "LearnPieces", 400,
                       word->chopped_word->bounding_box());
    RefreshDebugWindow(&learn_fragmented_word_debug_win_, "LearnWord", 200,
                       word->chopped_word->bounding_box());
    word->chopped_word->plot(learn_fragmented_word_debug_win_);
    ScrollView::Update();
  }
#endif // !GRAPHICS_DISABLED
 
  for (int ch = 0; ch < word_len; ++ch) {
    if (classify_debug_character_fragments) {
      tprintf("\nLearning %s\n", word->correct_text[ch].c_str());
    }
    if (word->correct_text[ch].length() > 0) {
      float threshold = thresholds != nullptr ? thresholds[ch] : 0.0f;
 
      LearnPieces(fontname, start_blob, word->best_state[ch], threshold, CST_WHOLE,
                  word->correct_text[ch].c_str(), word);
 
      if (word->best_state[ch] > 1 && !disable_character_fragments) {
        // Check that the character breaks into meaningful fragments
        // that each match a whole character with at least
        // classify_character_fragments_garbage_certainty_threshold
        bool garbage = false;
        int frag;
        for (frag = 0; frag < word->best_state[ch]; ++frag) {
          TBLOB *frag_blob = word->chopped_word->blobs[start_blob + frag];
          if (classify_character_fragments_garbage_certainty_threshold < 0) {
            garbage |= LooksLikeGarbage(frag_blob);
          }
        }
        // Learn the fragments.
        if (!garbage) {
          bool pieces_all_natural = word->PiecesAllNatural(start_blob, word->best_state[ch]);
          if (pieces_all_natural || !prioritize_division) {
            for (frag = 0; frag < word->best_state[ch]; ++frag) {
              std::vector<std::string> tokens = split(word->correct_text[ch], ' ');
 
              tokens[0] = CHAR_FRAGMENT::to_string(tokens[0].c_str(), frag, word->best_state[ch],
                                                   pieces_all_natural);
 
              std::string full_string;
              for (unsigned i = 0; i < tokens.size(); i++) {
                full_string += tokens[i];
                if (i != tokens.size() - 1) {
                  full_string += ' ';
                }
              }
              LearnPieces(fontname, start_blob + frag, 1, threshold, CST_FRAGMENT,
                          full_string.c_str(), word);
            }
          }
        }
      }
 
      // TODO(rays): re-enable this part of the code when we switch to the
      // new classifier that needs to see examples of garbage.
      /*
if (word->best_state[ch] > 1) {
  // If the next blob is good, make junk with the rightmost fragment.
  if (ch + 1 < word_len && word->correct_text[ch + 1].length() > 0) {
    LearnPieces(fontname, start_blob + word->best_state[ch] - 1,
                word->best_state[ch + 1] + 1,
                threshold, CST_IMPROPER, INVALID_UNICHAR, word);
  }
  // If the previous blob is good, make junk with the leftmost fragment.
  if (ch > 0 && word->correct_text[ch - 1].length() > 0) {
    LearnPieces(fontname, start_blob - word->best_state[ch - 1],
                word->best_state[ch - 1] + 1,
                threshold, CST_IMPROPER, INVALID_UNICHAR, word);
  }
}
// If the next blob is good, make a join with it.
if (ch + 1 < word_len && word->correct_text[ch + 1].length() > 0) {
  std::string joined_text = word->correct_text[ch];
  joined_text += word->correct_text[ch + 1];
  LearnPieces(fontname, start_blob,
              word->best_state[ch] + word->best_state[ch + 1],
              threshold, CST_NGRAM, joined_text.c_str(), word);
}
*/
    }
    start_blob += word->best_state[ch];
  }
  delete[] thresholds;
} // LearnWord.
 
// Builds a blob of length fragments, from the word, starting at start,
// and then learns it, as having the given correct_text.
// If fontname is not nullptr, then LearnBlob is called and the data will be
// saved in an internal buffer for static training.
// Otherwise AdaptToBlob is called for adaption within a document.
// threshold is a magic number required by AdaptToChar and generated by
// ComputeAdaptionThresholds.
// Although it can be partly inferred from the string, segmentation is
// provided to explicitly clarify the character segmentation.
void Classify::LearnPieces(const char *fontname, int start, int length, float threshold,
                           CharSegmentationType segmentation, const char *correct_text,
                           WERD_RES *word) {
  // TODO(daria) Remove/modify this if/when we want
  // to train and/or adapt to n-grams.
  if (segmentation != CST_WHOLE && (segmentation != CST_FRAGMENT || disable_character_fragments)) {
    return;
  }
 
  if (length > 1) {
    SEAM::JoinPieces(word->seam_array, word->chopped_word->blobs, start, start + length - 1);
  }
  TBLOB *blob = word->chopped_word->blobs[start];
  // Rotate the blob if needed for classification.
  TBLOB *rotated_blob = blob->ClassifyNormalizeIfNeeded();
  if (rotated_blob == nullptr) {
    rotated_blob = blob;
  }
 
#ifndef GRAPHICS_DISABLED
  // Draw debug windows showing the blob that is being learned if needed.
  if (strcmp(classify_learn_debug_str.c_str(), correct_text) == 0) {
    RefreshDebugWindow(&learn_debug_win_, "LearnPieces", 600, word->chopped_word->bounding_box());
    rotated_blob->plot(learn_debug_win_, ScrollView::GREEN, ScrollView::BROWN);
    learn_debug_win_->Update();
    learn_debug_win_->Wait();
  }
  if (classify_debug_character_fragments && segmentation == CST_FRAGMENT) {
    ASSERT_HOST(learn_fragments_debug_win_ != nullptr); // set up in LearnWord
    blob->plot(learn_fragments_debug_win_, ScrollView::BLUE, ScrollView::BROWN);
    learn_fragments_debug_win_->Update();
  }
#endif // !GRAPHICS_DISABLED
 
  if (fontname != nullptr) {
    classify_norm_method.set_value(character); // force char norm spc 30/11/93
    tess_bn_matching.set_value(false);         // turn it off
    tess_cn_matching.set_value(false);
    DENORM bl_denorm, cn_denorm;
    INT_FX_RESULT_STRUCT fx_info;
    SetupBLCNDenorms(*rotated_blob, classify_nonlinear_norm, &bl_denorm, &cn_denorm, &fx_info);
    LearnBlob(fontname, rotated_blob, cn_denorm, fx_info, correct_text);
  } else if (unicharset.contains_unichar(correct_text)) {
    UNICHAR_ID class_id = unicharset.unichar_to_id(correct_text);
    int font_id = word->fontinfo != nullptr ? fontinfo_table_.get_index(*word->fontinfo) : 0;
    if (classify_learning_debug_level >= 1) {
      tprintf("Adapting to char = %s, thr= %g font_id= %d\n", unicharset.id_to_unichar(class_id),
              threshold, font_id);
    }
    // If filename is not nullptr we are doing recognition
    // (as opposed to training), so we must have already set word fonts.
    AdaptToChar(rotated_blob, class_id, font_id, threshold, AdaptedTemplates);
    if (BackupAdaptedTemplates != nullptr) {
      // Adapt the backup templates too. They will be used if the primary gets
      // too full.
      AdaptToChar(rotated_blob, class_id, font_id, threshold, BackupAdaptedTemplates);
    }
  } else if (classify_debug_level >= 1) {
    tprintf("Can't adapt to %s not in unicharset\n", correct_text);
  }
  if (rotated_blob != blob) {
    delete rotated_blob;
  }
 
  SEAM::BreakPieces(word->seam_array, word->chopped_word->blobs, start, start + length - 1);
} // LearnPieces.
 
/*---------------------------------------------------------------------------*/
void Classify::EndAdaptiveClassifier() {
  std::string Filename;
  FILE *File;
 
  if (AdaptedTemplates != nullptr && classify_enable_adaptive_matcher &&
      classify_save_adapted_templates) {
    Filename = imagefile + ADAPT_TEMPLATE_SUFFIX;
    File = fopen(Filename.c_str(), "wb");
    if (File == nullptr) {
      tprintf("Unable to save adapted templates to %s!\n", Filename.c_str());
    } else {
      tprintf("\nSaving adapted templates to %s ...", Filename.c_str());
      fflush(stdout);
      WriteAdaptedTemplates(File, AdaptedTemplates);
      tprintf("\n");
      fclose(File);
    }
  }
 
  delete AdaptedTemplates;
  AdaptedTemplates = nullptr;
  delete BackupAdaptedTemplates;
  BackupAdaptedTemplates = nullptr;
 
  if (PreTrainedTemplates != nullptr) {
    delete PreTrainedTemplates;
    PreTrainedTemplates = nullptr;
  }
  getDict().EndDangerousAmbigs();
  FreeNormProtos();
  if (AllProtosOn != nullptr) {
    FreeBitVector(AllProtosOn);
    FreeBitVector(AllConfigsOn);
    FreeBitVector(AllConfigsOff);
    FreeBitVector(TempProtoMask);
    AllProtosOn = nullptr;
    AllConfigsOn = nullptr;
    AllConfigsOff = nullptr;
    TempProtoMask = nullptr;
  }
  delete shape_table_;
  shape_table_ = nullptr;
  delete static_classifier_;
  static_classifier_ = nullptr;
} /* EndAdaptiveClassifier */
 
/*---------------------------------------------------------------------------*/
void Classify::InitAdaptiveClassifier(TessdataManager *mgr) {
  if (!classify_enable_adaptive_matcher) {
    return;
  }
  if (AllProtosOn != nullptr) {
    EndAdaptiveClassifier(); // Don't leak with multiple inits.
  }
 
  // If there is no language_data_path_prefix, the classifier will be
  // adaptive only.
  if (language_data_path_prefix.length() > 0 && mgr != nullptr) {
    TFile fp;
    ASSERT_HOST(mgr->GetComponent(TESSDATA_INTTEMP, &fp));
    PreTrainedTemplates = ReadIntTemplates(&fp);
 
    if (mgr->GetComponent(TESSDATA_SHAPE_TABLE, &fp)) {
      shape_table_ = new ShapeTable(unicharset);
      if (!shape_table_->DeSerialize(&fp)) {
        tprintf("Error loading shape table!\n");
        delete shape_table_;
        shape_table_ = nullptr;
      }
    }
 
    ASSERT_HOST(mgr->GetComponent(TESSDATA_PFFMTABLE, &fp));
    ReadNewCutoffs(&fp, CharNormCutoffs);
 
    ASSERT_HOST(mgr->GetComponent(TESSDATA_NORMPROTO, &fp));
    NormProtos = ReadNormProtos(&fp);
    static_classifier_ = new TessClassifier(false, this);
  }
 
  InitIntegerFX();
 
  AllProtosOn = NewBitVector(MAX_NUM_PROTOS);
  AllConfigsOn = NewBitVector(MAX_NUM_CONFIGS);
  AllConfigsOff = NewBitVector(MAX_NUM_CONFIGS);
  TempProtoMask = NewBitVector(MAX_NUM_PROTOS);
  set_all_bits(AllProtosOn, WordsInVectorOfSize(MAX_NUM_PROTOS));
  set_all_bits(AllConfigsOn, WordsInVectorOfSize(MAX_NUM_CONFIGS));
  zero_all_bits(AllConfigsOff, WordsInVectorOfSize(MAX_NUM_CONFIGS));
 
  for (uint16_t &BaselineCutoff : BaselineCutoffs) {
    BaselineCutoff = 0;
  }
 
  if (classify_use_pre_adapted_templates) {
    TFile fp;
    std::string Filename = imagefile;
    Filename += ADAPT_TEMPLATE_SUFFIX;
    if (!fp.Open(Filename.c_str(), nullptr)) {
      AdaptedTemplates = new ADAPT_TEMPLATES_STRUCT(unicharset);
    } else {
      tprintf("\nReading pre-adapted templates from %s ...\n", Filename.c_str());
      fflush(stdout);
      AdaptedTemplates = ReadAdaptedTemplates(&fp);
      tprintf("\n");
      PrintAdaptedTemplates(stdout, AdaptedTemplates);
 
      for (unsigned i = 0; i < AdaptedTemplates->Templates->NumClasses; i++) {
        BaselineCutoffs[i] = CharNormCutoffs[i];
      }
    }
  } else {
    delete AdaptedTemplates;
    AdaptedTemplates = new ADAPT_TEMPLATES_STRUCT(unicharset);
  }
} /* InitAdaptiveClassifier */
 
void Classify::ResetAdaptiveClassifierInternal() {
  if (classify_learning_debug_level > 0) {
    tprintf("Resetting adaptive classifier (NumAdaptationsFailed=%d)\n", NumAdaptationsFailed);
  }
  delete AdaptedTemplates;
  AdaptedTemplates = new ADAPT_TEMPLATES_STRUCT(unicharset);
  delete BackupAdaptedTemplates;
  BackupAdaptedTemplates = nullptr;
  NumAdaptationsFailed = 0;
}
 
// If there are backup adapted templates, switches to those, otherwise resets
// the main adaptive classifier (because it is full.)
void Classify::SwitchAdaptiveClassifier() {
  if (BackupAdaptedTemplates == nullptr) {
    ResetAdaptiveClassifierInternal();
    return;
  }
  if (classify_learning_debug_level > 0) {
    tprintf("Switch to backup adaptive classifier (NumAdaptationsFailed=%d)\n",
            NumAdaptationsFailed);
  }
  delete AdaptedTemplates;
  AdaptedTemplates = BackupAdaptedTemplates;
  BackupAdaptedTemplates = nullptr;
  NumAdaptationsFailed = 0;
}
 
// Resets the backup adaptive classifier to empty.
void Classify::StartBackupAdaptiveClassifier() {
  delete BackupAdaptedTemplates;
  BackupAdaptedTemplates = new ADAPT_TEMPLATES_STRUCT(unicharset);
}
 
/*---------------------------------------------------------------------------*/
void Classify::SettupPass1() {
  EnableLearning = classify_enable_learning;
 
  getDict().SettupStopperPass1();
 
} /* SettupPass1 */
 
/*---------------------------------------------------------------------------*/
void Classify::SettupPass2() {
  EnableLearning = false;
  getDict().SettupStopperPass2();
 
} /* SettupPass2 */
 
/*---------------------------------------------------------------------------*/
void Classify::InitAdaptedClass(TBLOB *Blob, CLASS_ID ClassId, int FontinfoId, ADAPT_CLASS_STRUCT *Class,
                                ADAPT_TEMPLATES_STRUCT *Templates) {
  FEATURE_SET Features;
  int Fid, Pid;
  FEATURE Feature;
  int NumFeatures;
  PROTO_STRUCT *Proto;
  INT_CLASS_STRUCT *IClass;
  TEMP_CONFIG_STRUCT *Config;
 
  classify_norm_method.set_value(baseline);
  Features = ExtractOutlineFeatures(Blob);
  NumFeatures = Features->NumFeatures;
  if (NumFeatures > UNLIKELY_NUM_FEAT || NumFeatures <= 0) {
    delete Features;
    return;
  }
 
  Config = new TEMP_CONFIG_STRUCT(NumFeatures - 1, FontinfoId);
  TempConfigFor(Class, 0) = Config;
 
  /* this is a kludge to construct cutoffs for adapted templates */
  if (Templates == AdaptedTemplates) {
    BaselineCutoffs[ClassId] = CharNormCutoffs[ClassId];
  }
 
  IClass = ClassForClassId(Templates->Templates, ClassId);
 
  for (Fid = 0; Fid < Features->NumFeatures; Fid++) {
    Pid = AddIntProto(IClass);
    assert(Pid != NO_PROTO);
 
    Feature = Features->Features[Fid];
    auto TempProto = new TEMP_PROTO_STRUCT;
    Proto = &(TempProto->Proto);
 
    /* compute proto params - NOTE that Y_DIM_OFFSET must be used because
   ConvertProto assumes that the Y dimension varies from -0.5 to 0.5
   instead of the -0.25 to 0.75 used in baseline normalization */
    Proto->Angle = Feature->Params[OutlineFeatDir];
    Proto->X = Feature->Params[OutlineFeatX];
    Proto->Y = Feature->Params[OutlineFeatY] - Y_DIM_OFFSET;
    Proto->Length = Feature->Params[OutlineFeatLength];
    FillABC(Proto);
 
    TempProto->ProtoId = Pid;
    SET_BIT(Config->Protos, Pid);
 
    ConvertProto(Proto, Pid, IClass);
    AddProtoToProtoPruner(Proto, Pid, IClass, classify_learning_debug_level >= 2);
 
    Class->TempProtos = push(Class->TempProtos, TempProto);
  }
  delete Features;
 
  AddIntConfig(IClass);
  ConvertConfig(AllProtosOn, 0, IClass);
 
  if (classify_learning_debug_level >= 1) {
    tprintf("Added new class '%s' with class id %d and %d protos.\n",
            unicharset.id_to_unichar(ClassId), ClassId, NumFeatures);
#ifndef GRAPHICS_DISABLED
    if (classify_learning_debug_level > 1) {
      DisplayAdaptedChar(Blob, IClass);
    }
#endif
  }
 
  if (IsEmptyAdaptedClass(Class)) {
    (Templates->NumNonEmptyClasses)++;
  }
} /* InitAdaptedClass */
 
/*---------------------------------------------------------------------------*/
int Classify::GetAdaptiveFeatures(TBLOB *Blob, INT_FEATURE_ARRAY IntFeatures,
                                  FEATURE_SET *FloatFeatures) {
  FEATURE_SET Features;
  int NumFeatures;
 
  classify_norm_method.set_value(baseline);
  Features = ExtractPicoFeatures(Blob);
 
  NumFeatures = Features->NumFeatures;
  if (NumFeatures == 0 || NumFeatures > UNLIKELY_NUM_FEAT) {
    delete Features;
    return 0;
  }
 
  ComputeIntFeatures(Features, IntFeatures);
  *FloatFeatures = Features;
 
  return NumFeatures;
} /* GetAdaptiveFeatures */
 
/*-----------------------------------------------------------------------------
              Private Code
-----------------------------------------------------------------------------*/
/*---------------------------------------------------------------------------*/
bool Classify::AdaptableWord(WERD_RES *word) {
  if (word->best_choice == nullptr) {
    return false;
  }
  auto BestChoiceLength = word->best_choice->length();
  float adaptable_score = getDict().segment_penalty_dict_case_ok + ADAPTABLE_WERD_ADJUSTMENT;
  return // rules that apply in general - simplest to compute first
      BestChoiceLength > 0 && BestChoiceLength == word->rebuild_word->NumBlobs() &&
      BestChoiceLength <= MAX_ADAPTABLE_WERD_SIZE &&
      // This basically ensures that the word is at least a dictionary match
      // (freq word, user word, system dawg word, etc).
      // Since all the other adjustments will make adjust factor higher
      // than higher than adaptable_score=1.1+0.05=1.15
      // Since these are other flags that ensure that the word is dict word,
      // this check could be at times redundant.
      word->best_choice->adjust_factor() <= adaptable_score &&
      // Make sure that alternative choices are not dictionary words.
      word->AlternativeChoiceAdjustmentsWorseThan(adaptable_score);
}
 
/*---------------------------------------------------------------------------*/
void Classify::AdaptToChar(TBLOB *Blob, CLASS_ID ClassId, int FontinfoId, float Threshold,
                           ADAPT_TEMPLATES_STRUCT *adaptive_templates) {
  int NumFeatures;
  INT_FEATURE_ARRAY IntFeatures;
  UnicharRating int_result;
  INT_CLASS_STRUCT *IClass;
  ADAPT_CLASS_STRUCT *Class;
  TEMP_CONFIG_STRUCT *TempConfig;
  FEATURE_SET FloatFeatures;
  int NewTempConfigId;
 
  if (!LegalClassId(ClassId)) {
    return;
  }
 
  int_result.unichar_id = ClassId;
  Class = adaptive_templates->Class[ClassId];
  assert(Class != nullptr);
  if (IsEmptyAdaptedClass(Class)) {
    InitAdaptedClass(Blob, ClassId, FontinfoId, Class, adaptive_templates);
  } else {
    IClass = ClassForClassId(adaptive_templates->Templates, ClassId);
 
    NumFeatures = GetAdaptiveFeatures(Blob, IntFeatures, &FloatFeatures);
    if (NumFeatures <= 0) {
      return; // Features already freed by GetAdaptiveFeatures.
    }
 
    // Only match configs with the matching font.
    BIT_VECTOR MatchingFontConfigs = NewBitVector(MAX_NUM_PROTOS);
    for (int cfg = 0; cfg < IClass->NumConfigs; ++cfg) {
      if (GetFontinfoId(Class, cfg) == FontinfoId) {
        SET_BIT(MatchingFontConfigs, cfg);
      } else {
        reset_bit(MatchingFontConfigs, cfg);
      }
    }
    im_.Match(IClass, AllProtosOn, MatchingFontConfigs, NumFeatures, IntFeatures, &int_result,
              classify_adapt_feature_threshold, NO_DEBUG, matcher_debug_separate_windows);
    FreeBitVector(MatchingFontConfigs);
 
    SetAdaptiveThreshold(Threshold);
 
    if (1.0f - int_result.rating <= Threshold) {
      if (ConfigIsPermanent(Class, int_result.config)) {
        if (classify_learning_debug_level >= 1) {
          tprintf("Found good match to perm config %d = %4.1f%%.\n", int_result.config,
                  int_result.rating * 100.0);
        }
        delete FloatFeatures;
        return;
      }
 
      TempConfig = TempConfigFor(Class, int_result.config);
      IncreaseConfidence(TempConfig);
      if (TempConfig->NumTimesSeen > Class->MaxNumTimesSeen) {
        Class->MaxNumTimesSeen = TempConfig->NumTimesSeen;
      }
      if (classify_learning_debug_level >= 1) {
        tprintf("Increasing reliability of temp config %d to %d.\n", int_result.config,
                TempConfig->NumTimesSeen);
      }
 
      if (TempConfigReliable(ClassId, TempConfig)) {
        MakePermanent(adaptive_templates, ClassId, int_result.config, Blob);
        UpdateAmbigsGroup(ClassId, Blob);
      }
    } else {
      if (classify_learning_debug_level >= 1) {
        tprintf("Found poor match to temp config %d = %4.1f%%.\n", int_result.config,
                int_result.rating * 100.0);
#ifndef GRAPHICS_DISABLED
        if (classify_learning_debug_level > 2) {
          DisplayAdaptedChar(Blob, IClass);
        }
#endif
      }
      NewTempConfigId = MakeNewTemporaryConfig(adaptive_templates, ClassId, FontinfoId, NumFeatures,
                                               IntFeatures, FloatFeatures);
      if (NewTempConfigId >= 0 &&
          TempConfigReliable(ClassId, TempConfigFor(Class, NewTempConfigId))) {
        MakePermanent(adaptive_templates, ClassId, NewTempConfigId, Blob);
        UpdateAmbigsGroup(ClassId, Blob);
      }
 
#ifndef GRAPHICS_DISABLED
      if (classify_learning_debug_level > 1) {
        DisplayAdaptedChar(Blob, IClass);
      }
#endif
    }
    delete FloatFeatures;
  }
} /* AdaptToChar */
 
#ifndef GRAPHICS_DISABLED
 
void Classify::DisplayAdaptedChar(TBLOB *blob, INT_CLASS_STRUCT *int_class) {
  INT_FX_RESULT_STRUCT fx_info;
  std::vector<INT_FEATURE_STRUCT> bl_features;
  TrainingSample *sample =
      BlobToTrainingSample(*blob, classify_nonlinear_norm, &fx_info, &bl_features);
  if (sample == nullptr) {
    return;
  }
 
  UnicharRating int_result;
  im_.Match(int_class, AllProtosOn, AllConfigsOn, bl_features.size(), &bl_features[0], &int_result,
            classify_adapt_feature_threshold, NO_DEBUG, matcher_debug_separate_windows);
  tprintf("Best match to temp config %d = %4.1f%%.\n", int_result.config,
          int_result.rating * 100.0);
  if (classify_learning_debug_level >= 2) {
    uint32_t ConfigMask;
    ConfigMask = 1 << int_result.config;
    ShowMatchDisplay();
    im_.Match(int_class, AllProtosOn, static_cast<BIT_VECTOR>(&ConfigMask), bl_features.size(),
              &bl_features[0], &int_result, classify_adapt_feature_threshold, 6 | 0x19,
              matcher_debug_separate_windows);
    UpdateMatchDisplay();
  }
 
  delete sample;
}
 
#endif
 
void Classify::AddNewResult(const UnicharRating &new_result, ADAPT_RESULTS *results) {
  auto old_match = FindScoredUnichar(new_result.unichar_id, *results);
 
  if (new_result.rating + matcher_bad_match_pad < results->best_rating ||
      (old_match < results->match.size() &&
       new_result.rating <= results->match[old_match].rating)) {
    return; // New one not good enough.
  }
 
  if (!unicharset.get_fragment(new_result.unichar_id)) {
    results->HasNonfragment = true;
  }
 
  if (old_match < results->match.size()) {
    results->match[old_match].rating = new_result.rating;
  } else {
    results->match.push_back(new_result);
  }
 
  if (new_result.rating > results->best_rating &&
      // Ensure that fragments do not affect best rating, class and config.
      // This is needed so that at least one non-fragmented character is
      // always present in the results.
      // TODO(daria): verify that this helps accuracy and does not
      // hurt performance.
      !unicharset.get_fragment(new_result.unichar_id)) {
    results->best_match_index = old_match;
    results->best_rating = new_result.rating;
    results->best_unichar_id = new_result.unichar_id;
  }
} /* AddNewResult */
 
/*---------------------------------------------------------------------------*/
void Classify::AmbigClassifier(const std::vector<INT_FEATURE_STRUCT> &int_features,
                               const INT_FX_RESULT_STRUCT &fx_info, const TBLOB *blob,
                               INT_TEMPLATES_STRUCT *templates, ADAPT_CLASS_STRUCT **classes,
                               UNICHAR_ID *ambiguities, ADAPT_RESULTS *results) {
  if (int_features.empty()) {
    return;
  }
  auto *CharNormArray = new uint8_t[unicharset.size()];
  UnicharRating int_result;
 
  results->BlobLength = GetCharNormFeature(fx_info, templates, nullptr, CharNormArray);
  bool debug = matcher_debug_level >= 2 || classify_debug_level > 1;
  if (debug) {
    tprintf("AM Matches =  ");
  }
 
  int top = blob->bounding_box().top();
  int bottom = blob->bounding_box().bottom();
  while (*ambiguities >= 0) {
    CLASS_ID class_id = *ambiguities;
 
    int_result.unichar_id = class_id;
    im_.Match(ClassForClassId(templates, class_id), AllProtosOn, AllConfigsOn, int_features.size(),
              &int_features[0], &int_result, classify_adapt_feature_threshold, NO_DEBUG,
              matcher_debug_separate_windows);
 
    ExpandShapesAndApplyCorrections(nullptr, debug, class_id, bottom, top, 0, results->BlobLength,
                                    classify_integer_matcher_multiplier, CharNormArray, &int_result,
                                    results);
    ambiguities++;
  }
  delete[] CharNormArray;
} /* AmbigClassifier */
 
/*---------------------------------------------------------------------------*/
void Classify::MasterMatcher(INT_TEMPLATES_STRUCT *templates, int16_t num_features,
                             const INT_FEATURE_STRUCT *features, const uint8_t *norm_factors,
                             ADAPT_CLASS_STRUCT **classes, int debug, int matcher_multiplier,
                             const TBOX &blob_box, const std::vector<CP_RESULT_STRUCT> &results,
                             ADAPT_RESULTS *final_results) {
  int top = blob_box.top();
  int bottom = blob_box.bottom();
  UnicharRating int_result;
  for (auto &&result : results) {
    CLASS_ID class_id = result.Class;
    BIT_VECTOR protos = classes != nullptr ? classes[class_id]->PermProtos : AllProtosOn;
    BIT_VECTOR configs = classes != nullptr ? classes[class_id]->PermConfigs : AllConfigsOn;
 
    int_result.unichar_id = class_id;
    im_.Match(ClassForClassId(templates, class_id), protos, configs, num_features, features,
              &int_result, classify_adapt_feature_threshold, debug, matcher_debug_separate_windows);
    bool is_debug = matcher_debug_level >= 2 || classify_debug_level > 1;
    ExpandShapesAndApplyCorrections(classes, is_debug, class_id, bottom, top, result.Rating,
                                    final_results->BlobLength, matcher_multiplier, norm_factors,
                                    &int_result, final_results);
  }
}
 
// Converts configs to fonts, and if the result is not adapted, and a
// shape_table_ is present, the shape is expanded to include all
// unichar_ids represented, before applying a set of corrections to the
// distance rating in int_result, (see ComputeCorrectedRating.)
// The results are added to the final_results output.
void Classify::ExpandShapesAndApplyCorrections(ADAPT_CLASS_STRUCT **classes, bool debug, int class_id,
                                               int bottom, int top, float cp_rating,
                                               int blob_length, int matcher_multiplier,
                                               const uint8_t *cn_factors, UnicharRating *int_result,
                                               ADAPT_RESULTS *final_results) {
  if (classes != nullptr) {
    // Adapted result. Convert configs to fontinfo_ids.
    int_result->adapted = true;
    for (auto &font : int_result->fonts) {
      font.fontinfo_id = GetFontinfoId(classes[class_id], font.fontinfo_id);
    }
  } else {
    // Pre-trained result. Map fonts using font_sets_.
    int_result->adapted = false;
    for (auto &font : int_result->fonts) {
      font.fontinfo_id = ClassAndConfigIDToFontOrShapeID(class_id, font.fontinfo_id);
    }
    if (shape_table_ != nullptr) {
      // Two possible cases:
      // 1. Flat shapetable. All unichar-ids of the shapes referenced by
      // int_result->fonts are the same. In this case build a new vector of
      // mapped fonts and replace the fonts in int_result.
      // 2. Multi-unichar shapetable. Variable unichars in the shapes referenced
      // by int_result. In this case, build a vector of UnicharRating to
      // gather together different font-ids for each unichar. Also covers case1.
      std::vector<UnicharRating> mapped_results;
      for (auto &f : int_result->fonts) {
        int shape_id = f.fontinfo_id;
        const Shape &shape = shape_table_->GetShape(shape_id);
        for (int c = 0; c < shape.size(); ++c) {
          int unichar_id = shape[c].unichar_id;
          if (!unicharset.get_enabled(unichar_id)) {
            continue;
          }
          // Find the mapped_result for unichar_id.
          unsigned r = 0;
          for (r = 0; r < mapped_results.size() && mapped_results[r].unichar_id != unichar_id;
               ++r) {
          }
          if (r == mapped_results.size()) {
            mapped_results.push_back(*int_result);
            mapped_results[r].unichar_id = unichar_id;
            mapped_results[r].fonts.clear();
          }
          for (int font_id : shape[c].font_ids) {
            mapped_results[r].fonts.emplace_back(font_id, f.score);
          }
        }
      }
      for (auto &m : mapped_results) {
        m.rating = ComputeCorrectedRating(debug, m.unichar_id, cp_rating, int_result->rating,
                                          int_result->feature_misses, bottom, top, blob_length,
                                          matcher_multiplier, cn_factors);
        AddNewResult(m, final_results);
      }
      return;
    }
  }
  if (unicharset.get_enabled(class_id)) {
    int_result->rating = ComputeCorrectedRating(debug, class_id, cp_rating, int_result->rating,
                                                int_result->feature_misses, bottom, top,
                                                blob_length, matcher_multiplier, cn_factors);
    AddNewResult(*int_result, final_results);
  }
}
 
// Applies a set of corrections to the confidence im_rating,
// including the cn_correction, miss penalty and additional penalty
// for non-alnums being vertical misfits. Returns the corrected confidence.
double Classify::ComputeCorrectedRating(bool debug, int unichar_id, double cp_rating,
                                        double im_rating, int feature_misses, int bottom, int top,
                                        int blob_length, int matcher_multiplier,
                                        const uint8_t *cn_factors) {
  // Compute class feature corrections.
  double cn_corrected = im_.ApplyCNCorrection(1.0 - im_rating, blob_length, cn_factors[unichar_id],
                                              matcher_multiplier);
  double miss_penalty = tessedit_class_miss_scale * feature_misses;
  double vertical_penalty = 0.0;
  // Penalize non-alnums for being vertical misfits.
  if (!unicharset.get_isalpha(unichar_id) && !unicharset.get_isdigit(unichar_id) &&
      cn_factors[unichar_id] != 0 && classify_misfit_junk_penalty > 0.0) {
    int min_bottom, max_bottom, min_top, max_top;
    unicharset.get_top_bottom(unichar_id, &min_bottom, &max_bottom, &min_top, &max_top);
    if (debug) {
      tprintf("top=%d, vs [%d, %d], bottom=%d, vs [%d, %d]\n", top, min_top, max_top, bottom,
              min_bottom, max_bottom);
    }
    if (top < min_top || top > max_top || bottom < min_bottom || bottom > max_bottom) {
      vertical_penalty = classify_misfit_junk_penalty;
    }
  }
  double result = 1.0 - (cn_corrected + miss_penalty + vertical_penalty);
  if (result < WORST_POSSIBLE_RATING) {
    result = WORST_POSSIBLE_RATING;
  }
  if (debug) {
    tprintf("%s: %2.1f%%(CP%2.1f, IM%2.1f + CN%.2f(%d) + MP%2.1f + VP%2.1f)\n",
            unicharset.id_to_unichar(unichar_id), result * 100.0, cp_rating * 100.0,
            (1.0 - im_rating) * 100.0, (cn_corrected - (1.0 - im_rating)) * 100.0,
            cn_factors[unichar_id], miss_penalty * 100.0, vertical_penalty * 100.0);
  }
  return result;
}
 
/*---------------------------------------------------------------------------*/
UNICHAR_ID *Classify::BaselineClassifier(TBLOB *Blob,
                                         const std::vector<INT_FEATURE_STRUCT> &int_features,
                                         const INT_FX_RESULT_STRUCT &fx_info,
                                         ADAPT_TEMPLATES_STRUCT *Templates, ADAPT_RESULTS *Results) {
  if (int_features.empty()) {
    return nullptr;
  }
  auto *CharNormArray = new uint8_t[unicharset.size()];
  ClearCharNormArray(CharNormArray);
 
  Results->BlobLength = IntCastRounded(fx_info.Length / kStandardFeatureLength);
  PruneClasses(Templates->Templates, int_features.size(), -1, &int_features[0], CharNormArray,
               BaselineCutoffs, &Results->CPResults);
 
  if (matcher_debug_level >= 2 || classify_debug_level > 1) {
    tprintf("BL Matches =  ");
  }
 
  MasterMatcher(Templates->Templates, int_features.size(), &int_features[0], CharNormArray,
                Templates->Class, matcher_debug_flags, 0, Blob->bounding_box(), Results->CPResults,
                Results);
 
  delete[] CharNormArray;
  CLASS_ID ClassId = Results->best_unichar_id;
  if (ClassId == INVALID_UNICHAR_ID || Results->best_match_index < 0) {
    return nullptr;
  }
 
  return Templates->Class[ClassId]
      ->Config[Results->match[Results->best_match_index].config]
      .Perm->Ambigs;
} /* BaselineClassifier */
 
/*---------------------------------------------------------------------------*/
int Classify::CharNormClassifier(TBLOB *blob, const TrainingSample &sample,
                                 ADAPT_RESULTS *adapt_results) {
  // This is the length that is used for scaling ratings vs certainty.
  adapt_results->BlobLength = IntCastRounded(sample.outline_length() / kStandardFeatureLength);
  std::vector<UnicharRating> unichar_results;
  static_classifier_->UnicharClassifySample(sample, blob->denorm().pix(), 0, -1, &unichar_results);
  // Convert results to the format used internally by AdaptiveClassifier.
  for (auto &r : unichar_results) {
    AddNewResult(r, adapt_results);
  }
  return sample.num_features();
} /* CharNormClassifier */
 
// As CharNormClassifier, but operates on a TrainingSample and outputs to
// a vector of ShapeRating without conversion to classes.
int Classify::CharNormTrainingSample(bool pruner_only, int keep_this, const TrainingSample &sample,
                                     std::vector<UnicharRating> *results) {
  results->clear();
  std::unique_ptr<ADAPT_RESULTS> adapt_results(new ADAPT_RESULTS());
  adapt_results->Initialize();
  // Compute the bounding box of the features.
  uint32_t num_features = sample.num_features();
  // Only the top and bottom of the blob_box are used by MasterMatcher, so
  // fabricate right and left using top and bottom.
  TBOX blob_box(sample.geo_feature(GeoBottom), sample.geo_feature(GeoBottom),
                sample.geo_feature(GeoTop), sample.geo_feature(GeoTop));
  // Compute the char_norm_array from the saved cn_feature.
  FEATURE norm_feature = sample.GetCNFeature();
  std::vector<uint8_t> char_norm_array(unicharset.size());
  auto num_pruner_classes = std::max(static_cast<unsigned>(unicharset.size()), PreTrainedTemplates->NumClasses);
  std::vector<uint8_t> pruner_norm_array(num_pruner_classes);
  adapt_results->BlobLength = static_cast<int>(ActualOutlineLength(norm_feature) * 20 + 0.5f);
  ComputeCharNormArrays(norm_feature, PreTrainedTemplates, &char_norm_array[0], &pruner_norm_array[0]);
 
  PruneClasses(PreTrainedTemplates, num_features, keep_this, sample.features(), &pruner_norm_array[0],
               shape_table_ != nullptr ? &shapetable_cutoffs_[0] : CharNormCutoffs,
               &adapt_results->CPResults);
  if (keep_this >= 0) {
    adapt_results->CPResults[0].Class = keep_this;
    adapt_results->CPResults.resize(1);
  }
  if (pruner_only) {
    // Convert pruner results to output format.
    for (auto &it : adapt_results->CPResults) {
      int class_id = it.Class;
      results->push_back(UnicharRating(class_id, 1.0f - it.Rating));
    }
  } else {
    MasterMatcher(PreTrainedTemplates, num_features, sample.features(), &char_norm_array[0], nullptr,
                  matcher_debug_flags, classify_integer_matcher_multiplier, blob_box,
                  adapt_results->CPResults, adapt_results.get());
    // Convert master matcher results to output format.
    for (auto &i : adapt_results->match) {
      results->push_back(i);
    }
    if (results->size() > 1) {
      std::sort(results->begin(), results->end(), SortDescendingRating);
    }
  }
  return num_features;
} /* CharNormTrainingSample */
 
/*---------------------------------------------------------------------------*/
void Classify::ClassifyAsNoise(ADAPT_RESULTS *results) {
  float rating = results->BlobLength / matcher_avg_noise_size;
  rating *= rating;
  rating /= 1 + rating;
 
  AddNewResult(UnicharRating(UNICHAR_SPACE, 1.0f - rating), results);
} /* ClassifyAsNoise */
 
void Classify::ConvertMatchesToChoices(const DENORM &denorm, const TBOX &box,
                                       ADAPT_RESULTS *Results, BLOB_CHOICE_LIST *Choices) {
  assert(Choices != nullptr);
  float Rating;
  float Certainty;
  BLOB_CHOICE_IT temp_it;
  bool contains_nonfrag = false;
  temp_it.set_to_list(Choices);
  int choices_length = 0;
  // With no shape_table_ maintain the previous MAX_MATCHES as the maximum
  // number of returned results, but with a shape_table_ we want to have room
  // for at least the biggest shape (which might contain hundreds of Indic
  // grapheme fragments) and more, so use double the size of the biggest shape
  // if that is more than the default.
  int max_matches = MAX_MATCHES;
  if (shape_table_ != nullptr) {
    max_matches = shape_table_->MaxNumUnichars() * 2;
    if (max_matches < MAX_MATCHES) {
      max_matches = MAX_MATCHES;
    }
  }
 
  float best_certainty = -FLT_MAX;
  for (auto &it : Results->match) {
    const UnicharRating &result = it;
    bool adapted = result.adapted;
    bool current_is_frag = (unicharset.get_fragment(result.unichar_id) != nullptr);
    if (temp_it.length() + 1 == max_matches && !contains_nonfrag && current_is_frag) {
      continue; // look for a non-fragmented character to fill the
                // last spot in Choices if only fragments are present
    }
    // BlobLength can never be legally 0, this means recognition failed.
    // But we must return a classification result because some invoking
    // functions (chopper/permuter) do not anticipate a null blob choice.
    // So we need to assign a poor, but not infinitely bad score.
    if (Results->BlobLength == 0) {
      Certainty = -20;
      Rating = 100; // should be -certainty * real_blob_length
    } else {
      Rating = Certainty = (1.0f - result.rating);
      Rating *= rating_scale * Results->BlobLength;
      Certainty *= -(getDict().certainty_scale);
    }
    // Adapted results, by their very nature, should have good certainty.
    // Those that don't are at best misleading, and often lead to errors,
    // so don't accept adapted results that are too far behind the best result,
    // whether adapted or static.
    // TODO(rays) find some way of automatically tuning these constants.
    if (Certainty > best_certainty) {
      best_certainty = std::min(Certainty, static_cast<float>(classify_adapted_pruning_threshold));
    } else if (adapted && Certainty / classify_adapted_pruning_factor < best_certainty) {
      continue; // Don't accept bad adapted results.
    }
 
    float min_xheight, max_xheight, yshift;
    denorm.XHeightRange(result.unichar_id, unicharset, box, &min_xheight, &max_xheight, &yshift);
    auto *choice = new BLOB_CHOICE(
        result.unichar_id, Rating, Certainty, unicharset.get_script(result.unichar_id), min_xheight,
        max_xheight, yshift, adapted ? BCC_ADAPTED_CLASSIFIER : BCC_STATIC_CLASSIFIER);
    choice->set_fonts(result.fonts);
    temp_it.add_to_end(choice);
    contains_nonfrag |= !current_is_frag; // update contains_nonfrag
    choices_length++;
    if (choices_length >= max_matches) {
      break;
    }
  }
  Results->match.resize(choices_length);
} // ConvertMatchesToChoices
 
/*---------------------------------------------------------------------------*/
#ifndef GRAPHICS_DISABLED
void Classify::DebugAdaptiveClassifier(TBLOB *blob, ADAPT_RESULTS *Results) {
  if (static_classifier_ == nullptr) {
    return;
  }
  INT_FX_RESULT_STRUCT fx_info;
  std::vector<INT_FEATURE_STRUCT> bl_features;
  TrainingSample *sample = BlobToTrainingSample(*blob, false, &fx_info, &bl_features);
  if (sample == nullptr) {
    return;
  }
  static_classifier_->DebugDisplay(*sample, blob->denorm().pix(), Results->best_unichar_id);
} /* DebugAdaptiveClassifier */
#endif
 
/*---------------------------------------------------------------------------*/
void Classify::DoAdaptiveMatch(TBLOB *Blob, ADAPT_RESULTS *Results) {
  UNICHAR_ID *Ambiguities;
 
  INT_FX_RESULT_STRUCT fx_info;
  std::vector<INT_FEATURE_STRUCT> bl_features;
  TrainingSample *sample =
      BlobToTrainingSample(*Blob, classify_nonlinear_norm, &fx_info, &bl_features);
  if (sample == nullptr) {
    return;
  }
 
  // TODO: With LSTM, static_classifier_ is nullptr.
  // Return to avoid crash in CharNormClassifier.
  if (static_classifier_ == nullptr) {
    delete sample;
    return;
  }
 
  if (AdaptedTemplates->NumPermClasses < matcher_permanent_classes_min || tess_cn_matching) {
    CharNormClassifier(Blob, *sample, Results);
  } else {
    Ambiguities = BaselineClassifier(Blob, bl_features, fx_info, AdaptedTemplates, Results);
    if ((!Results->match.empty() &&
         MarginalMatch(Results->best_rating, matcher_reliable_adaptive_result) &&
         !tess_bn_matching) ||
        Results->match.empty()) {
      CharNormClassifier(Blob, *sample, Results);
    } else if (Ambiguities && *Ambiguities >= 0 && !tess_bn_matching) {
      AmbigClassifier(bl_features, fx_info, Blob, PreTrainedTemplates, AdaptedTemplates->Class,
                      Ambiguities, Results);
    }
  }
 
  // Force the blob to be classified as noise
  // if the results contain only fragments.
  // TODO(daria): verify that this is better than
  // just adding a nullptr classification.
  if (!Results->HasNonfragment || Results->match.empty()) {
    ClassifyAsNoise(Results);
  }
  delete sample;
} /* DoAdaptiveMatch */
 
/*---------------------------------------------------------------------------*/
UNICHAR_ID *Classify::GetAmbiguities(TBLOB *Blob, CLASS_ID CorrectClass) {
  auto *Results = new ADAPT_RESULTS();
  UNICHAR_ID *Ambiguities;
 
  Results->Initialize();
  INT_FX_RESULT_STRUCT fx_info;
  std::vector<INT_FEATURE_STRUCT> bl_features;
  TrainingSample *sample =
      BlobToTrainingSample(*Blob, classify_nonlinear_norm, &fx_info, &bl_features);
  if (sample == nullptr) {
    delete Results;
    return nullptr;
  }
 
  CharNormClassifier(Blob, *sample, Results);
  delete sample;
  RemoveBadMatches(Results);
  std::sort(Results->match.begin(), Results->match.end(), SortDescendingRating);
 
  /* copy the class id's into an string of ambiguities - don't copy if
   the correct class is the only class id matched */
  Ambiguities = new UNICHAR_ID[Results->match.size() + 1];
  if (Results->match.size() > 1 ||
      (Results->match.size() == 1 && Results->match[0].unichar_id != CorrectClass)) {
    unsigned i;
    for (i = 0; i < Results->match.size(); i++) {
      Ambiguities[i] = Results->match[i].unichar_id;
    }
    Ambiguities[i] = -1;
  } else {
    Ambiguities[0] = -1;
  }
 
  delete Results;
  return Ambiguities;
} /* GetAmbiguities */
 
// Returns true if the given blob looks too dissimilar to any character
// present in the classifier templates.
bool Classify::LooksLikeGarbage(TBLOB *blob) {
  auto *ratings = new BLOB_CHOICE_LIST();
  AdaptiveClassifier(blob, ratings);
  BLOB_CHOICE_IT ratings_it(ratings);
  const UNICHARSET &unicharset = getDict().getUnicharset();
  if (classify_debug_character_fragments) {
    print_ratings_list("======================\nLooksLikeGarbage() got ", ratings, unicharset);
  }
  for (ratings_it.mark_cycle_pt(); !ratings_it.cycled_list(); ratings_it.forward()) {
    if (unicharset.get_fragment(ratings_it.data()->unichar_id()) != nullptr) {
      continue;
    }
    float certainty = ratings_it.data()->certainty();
    delete ratings;
    return certainty < classify_character_fragments_garbage_certainty_threshold;
  }
  delete ratings;
  return true; // no whole characters in ratings
}
 
/*---------------------------------------------------------------------------*/
int Classify::GetCharNormFeature(const INT_FX_RESULT_STRUCT &fx_info, INT_TEMPLATES_STRUCT *templates,
                                 uint8_t *pruner_norm_array, uint8_t *char_norm_array) {
  auto norm_feature = new FEATURE_STRUCT(&CharNormDesc);
  float baseline = kBlnBaselineOffset;
  float scale = MF_SCALE_FACTOR;
  norm_feature->Params[CharNormY] = (fx_info.Ymean - baseline) * scale;
  norm_feature->Params[CharNormLength] = fx_info.Length * scale / LENGTH_COMPRESSION;
  norm_feature->Params[CharNormRx] = fx_info.Rx * scale;
  norm_feature->Params[CharNormRy] = fx_info.Ry * scale;
  // Deletes norm_feature.
  ComputeCharNormArrays(norm_feature, templates, char_norm_array, pruner_norm_array);
  return IntCastRounded(fx_info.Length / kStandardFeatureLength);
} /* GetCharNormFeature */
 
// Computes the char_norm_array for the unicharset and, if not nullptr, the
// pruner_array as appropriate according to the existence of the shape_table.
void Classify::ComputeCharNormArrays(FEATURE_STRUCT *norm_feature, INT_TEMPLATES_STRUCT *templates,
                                     uint8_t *char_norm_array, uint8_t *pruner_array) {
  ComputeIntCharNormArray(*norm_feature, char_norm_array);
  //if (pruner_array != nullptr) {
    if (shape_table_ == nullptr) {
      ComputeIntCharNormArray(*norm_feature, pruner_array);
    } else {
      memset(&pruner_array[0], UINT8_MAX, templates->NumClasses * sizeof(pruner_array[0]));
      // Each entry in the pruner norm array is the MIN of all the entries of
      // the corresponding unichars in the CharNormArray.
      for (unsigned id = 0; id < templates->NumClasses; ++id) {
        int font_set_id = templates->Class[id]->font_set_id;
        const FontSet &fs = fontset_table_.at(font_set_id);
        for (auto f : fs) {
          const Shape &shape = shape_table_->GetShape(f);
          for (int c = 0; c < shape.size(); ++c) {
            if (char_norm_array[shape[c].unichar_id] < pruner_array[id]) {
              pruner_array[id] = char_norm_array[shape[c].unichar_id];
            }
          }
        }
      }
    }
  //}
  delete norm_feature;
}
 
/*---------------------------------------------------------------------------*/
int Classify::MakeNewTemporaryConfig(ADAPT_TEMPLATES_STRUCT *Templates, CLASS_ID ClassId, int FontinfoId,
                                     int NumFeatures, INT_FEATURE_ARRAY Features,
                                     FEATURE_SET FloatFeatures) {
  INT_CLASS_STRUCT *IClass;
  ADAPT_CLASS_STRUCT *Class;
  PROTO_ID OldProtos[MAX_NUM_PROTOS];
  FEATURE_ID BadFeatures[MAX_NUM_INT_FEATURES];
  int NumOldProtos;
  int NumBadFeatures;
  int MaxProtoId, OldMaxProtoId;
  int MaskSize;
  int ConfigId;
  int i;
  int debug_level = NO_DEBUG;
 
  if (classify_learning_debug_level >= 3) {
    debug_level = PRINT_MATCH_SUMMARY | PRINT_FEATURE_MATCHES | PRINT_PROTO_MATCHES;
  }
 
  IClass = ClassForClassId(Templates->Templates, ClassId);
  Class = Templates->Class[ClassId];
 
  if (IClass->NumConfigs >= MAX_NUM_CONFIGS) {
    ++NumAdaptationsFailed;
    if (classify_learning_debug_level >= 1) {
      tprintf("Cannot make new temporary config: maximum number exceeded.\n");
    }
    return -1;
  }
 
  OldMaxProtoId = IClass->NumProtos - 1;
 
  NumOldProtos = im_.FindGoodProtos(IClass, AllProtosOn, AllConfigsOff, NumFeatures, Features,
                                    OldProtos, classify_adapt_proto_threshold, debug_level);
 
  MaskSize = WordsInVectorOfSize(MAX_NUM_PROTOS);
  zero_all_bits(TempProtoMask, MaskSize);
  for (i = 0; i < NumOldProtos; i++) {
    SET_BIT(TempProtoMask, OldProtos[i]);
  }
 
  NumBadFeatures = im_.FindBadFeatures(IClass, TempProtoMask, AllConfigsOn, NumFeatures, Features,
                                       BadFeatures, classify_adapt_feature_threshold, debug_level);
 
  MaxProtoId =
      MakeNewTempProtos(FloatFeatures, NumBadFeatures, BadFeatures, IClass, Class, TempProtoMask);
  if (MaxProtoId == NO_PROTO) {
    ++NumAdaptationsFailed;
    if (classify_learning_debug_level >= 1) {
      tprintf("Cannot make new temp protos: maximum number exceeded.\n");
    }
    return -1;
  }
 
  ConfigId = AddIntConfig(IClass);
  ConvertConfig(TempProtoMask, ConfigId, IClass);
  auto Config = new TEMP_CONFIG_STRUCT(MaxProtoId, FontinfoId);
  TempConfigFor(Class, ConfigId) = Config;
  copy_all_bits(TempProtoMask, Config->Protos, Config->ProtoVectorSize);
 
  if (classify_learning_debug_level >= 1) {
    tprintf(
        "Making new temp config %d fontinfo id %d"
        " using %d old and %d new protos.\n",
        ConfigId, Config->FontinfoId, NumOldProtos, MaxProtoId - OldMaxProtoId);
  }
 
  return ConfigId;
} /* MakeNewTemporaryConfig */
 
/*---------------------------------------------------------------------------*/
PROTO_ID Classify::MakeNewTempProtos(FEATURE_SET Features, int NumBadFeat, FEATURE_ID BadFeat[],
                                     INT_CLASS_STRUCT *IClass, ADAPT_CLASS_STRUCT *Class,
                                     BIT_VECTOR TempProtoMask) {
  FEATURE_ID *ProtoStart;
  FEATURE_ID *ProtoEnd;
  FEATURE_ID *LastBad;
  PROTO_STRUCT *Proto;
  FEATURE F1, F2;
  float X1, X2, Y1, Y2;
  float A1, A2, AngleDelta;
  float SegmentLength;
  PROTO_ID Pid;
 
  for (ProtoStart = BadFeat, LastBad = ProtoStart + NumBadFeat; ProtoStart < LastBad;
       ProtoStart = ProtoEnd) {
    F1 = Features->Features[*ProtoStart];
    X1 = F1->Params[PicoFeatX];
    Y1 = F1->Params[PicoFeatY];
    A1 = F1->Params[PicoFeatDir];
 
    for (ProtoEnd = ProtoStart + 1, SegmentLength = GetPicoFeatureLength(); ProtoEnd < LastBad;
         ProtoEnd++, SegmentLength += GetPicoFeatureLength()) {
      F2 = Features->Features[*ProtoEnd];
      X2 = F2->Params[PicoFeatX];
      Y2 = F2->Params[PicoFeatY];
      A2 = F2->Params[PicoFeatDir];
 
      AngleDelta = std::fabs(A1 - A2);
      if (AngleDelta > 0.5f) {
        AngleDelta = 1 - AngleDelta;
      }
 
      if (AngleDelta > matcher_clustering_max_angle_delta || std::fabs(X1 - X2) > SegmentLength ||
          std::fabs(Y1 - Y2) > SegmentLength) {
        break;
      }
    }
 
    F2 = Features->Features[*(ProtoEnd - 1)];
    X2 = F2->Params[PicoFeatX];
    Y2 = F2->Params[PicoFeatY];
    A2 = F2->Params[PicoFeatDir];
 
    Pid = AddIntProto(IClass);
    if (Pid == NO_PROTO) {
      return (NO_PROTO);
    }
 
    auto TempProto = new TEMP_PROTO_STRUCT;
    Proto = &(TempProto->Proto);
 
    /* compute proto params - NOTE that Y_DIM_OFFSET must be used because
   ConvertProto assumes that the Y dimension varies from -0.5 to 0.5
   instead of the -0.25 to 0.75 used in baseline normalization */
    Proto->Length = SegmentLength;
    Proto->Angle = A1;
    Proto->X = (X1 + X2) / 2;
    Proto->Y = (Y1 + Y2) / 2 - Y_DIM_OFFSET;
    FillABC(Proto);
 
    TempProto->ProtoId = Pid;
    SET_BIT(TempProtoMask, Pid);
 
    ConvertProto(Proto, Pid, IClass);
    AddProtoToProtoPruner(Proto, Pid, IClass, classify_learning_debug_level >= 2);
 
    Class->TempProtos = push(Class->TempProtos, TempProto);
  }
  return IClass->NumProtos - 1;
} /* MakeNewTempProtos */
 
/*---------------------------------------------------------------------------*/
void Classify::MakePermanent(ADAPT_TEMPLATES_STRUCT *Templates, CLASS_ID ClassId, int ConfigId,
                             TBLOB *Blob) {
  UNICHAR_ID *Ambigs;
  PROTO_KEY ProtoKey;
 
  auto Class = Templates->Class[ClassId];
  auto Config = TempConfigFor(Class, ConfigId);
 
  MakeConfigPermanent(Class, ConfigId);
  if (Class->NumPermConfigs == 0) {
    Templates->NumPermClasses++;
  }
  Class->NumPermConfigs++;
 
  // Initialize permanent config.
  Ambigs = GetAmbiguities(Blob, ClassId);
  auto Perm = new PERM_CONFIG_STRUCT;
  Perm->Ambigs = Ambigs;
  Perm->FontinfoId = Config->FontinfoId;
 
  // Free memory associated with temporary config (since ADAPTED_CONFIG
  // is a union we need to clean up before we record permanent config).
  ProtoKey.Templates = Templates;
  ProtoKey.ClassId = ClassId;
  ProtoKey.ConfigId = ConfigId;
  Class->TempProtos = delete_d(Class->TempProtos, &ProtoKey, MakeTempProtoPerm);
  delete Config;
 
  // Record permanent config.
  PermConfigFor(Class, ConfigId) = Perm;
 
  if (classify_learning_debug_level >= 1) {
    tprintf(
        "Making config %d for %s (ClassId %d) permanent:"
        " fontinfo id %d, ambiguities '",
        ConfigId, getDict().getUnicharset().debug_str(ClassId).c_str(), ClassId,
        PermConfigFor(Class, ConfigId)->FontinfoId);
    for (UNICHAR_ID *AmbigsPointer = Ambigs; *AmbigsPointer >= 0; ++AmbigsPointer) {
      tprintf("%s", unicharset.id_to_unichar(*AmbigsPointer));
    }
    tprintf("'.\n");
  }
} /* MakePermanent */
 
/*---------------------------------------------------------------------------*/
int MakeTempProtoPerm(void *item1, void *item2) {
  auto TempProto = static_cast<TEMP_PROTO_STRUCT *>(item1);
  auto ProtoKey = static_cast<PROTO_KEY *>(item2);
 
  auto Class = ProtoKey->Templates->Class[ProtoKey->ClassId];
  auto Config = TempConfigFor(Class, ProtoKey->ConfigId);
 
  if (TempProto->ProtoId > Config->MaxProtoId || !test_bit(Config->Protos, TempProto->ProtoId)) {
    return false;
  }
 
  MakeProtoPermanent(Class, TempProto->ProtoId);
  AddProtoToClassPruner(&(TempProto->Proto), ProtoKey->ClassId, ProtoKey->Templates->Templates);
  delete TempProto;
 
  return true;
} /* MakeTempProtoPerm */
 
/*---------------------------------------------------------------------------*/
void Classify::PrintAdaptiveMatchResults(const ADAPT_RESULTS &results) {
  for (auto &it : results.match) {
    tprintf("%s  ", unicharset.debug_str(it.unichar_id).c_str());
    it.Print();
  }
} /* PrintAdaptiveMatchResults */
 
/*---------------------------------------------------------------------------*/
void Classify::RemoveBadMatches(ADAPT_RESULTS *Results) {
  unsigned Next, NextGood;
  float BadMatchThreshold;
  static const char *romans = "i v x I V X";
  BadMatchThreshold = Results->best_rating - matcher_bad_match_pad;
 
  if (classify_bln_numeric_mode) {
    UNICHAR_ID unichar_id_one =
        unicharset.contains_unichar("1") ? unicharset.unichar_to_id("1") : -1;
    UNICHAR_ID unichar_id_zero =
        unicharset.contains_unichar("0") ? unicharset.unichar_to_id("0") : -1;
    float scored_one = ScoredUnichar(unichar_id_one, *Results);
    float scored_zero = ScoredUnichar(unichar_id_zero, *Results);
 
    for (Next = NextGood = 0; Next < Results->match.size(); Next++) {
      const UnicharRating &match = Results->match[Next];
      if (match.rating >= BadMatchThreshold) {
        if (!unicharset.get_isalpha(match.unichar_id) ||
            strstr(romans, unicharset.id_to_unichar(match.unichar_id)) != nullptr) {
        } else if (unicharset.eq(match.unichar_id, "l") && scored_one < BadMatchThreshold) {
          Results->match[Next].unichar_id = unichar_id_one;
        } else if (unicharset.eq(match.unichar_id, "O") && scored_zero < BadMatchThreshold) {
          Results->match[Next].unichar_id = unichar_id_zero;
        } else {
          Results->match[Next].unichar_id = INVALID_UNICHAR_ID; // Don't copy.
        }
        if (Results->match[Next].unichar_id != INVALID_UNICHAR_ID) {
          if (NextGood == Next) {
            ++NextGood;
          } else {
            Results->match[NextGood++] = Results->match[Next];
          }
        }
      }
    }
  } else {
    for (Next = NextGood = 0; Next < Results->match.size(); Next++) {
      if (Results->match[Next].rating >= BadMatchThreshold) {
        if (NextGood == Next) {
          ++NextGood;
        } else {
          Results->match[NextGood++] = Results->match[Next];
        }
      }
    }
  }
  Results->match.resize(NextGood);
} /* RemoveBadMatches */
 
/*----------------------------------------------------------------------------*/
void Classify::RemoveExtraPuncs(ADAPT_RESULTS *Results) {
  unsigned Next, NextGood;
  int punc_count; /*no of garbage characters */
  int digit_count;
  /*garbage characters */
  static char punc_chars[] = ". , ; : / ` ~ ' - = \\ | \" ! _ ^";
  static char digit_chars[] = "0 1 2 3 4 5 6 7 8 9";
 
  punc_count = 0;
  digit_count = 0;
  for (Next = NextGood = 0; Next < Results->match.size(); Next++) {
    const UnicharRating &match = Results->match[Next];
    bool keep = true;
    if (strstr(punc_chars, unicharset.id_to_unichar(match.unichar_id)) != nullptr) {
      if (punc_count >= 2) {
        keep = false;
      }
      punc_count++;
    } else {
      if (strstr(digit_chars, unicharset.id_to_unichar(match.unichar_id)) != nullptr) {
        if (digit_count >= 1) {
          keep = false;
        }
        digit_count++;
      }
    }
    if (keep) {
      if (NextGood == Next) {
        ++NextGood;
      } else {
        Results->match[NextGood++] = match;
      }
    }
  }
  Results->match.resize(NextGood);
} /* RemoveExtraPuncs */
 
/*---------------------------------------------------------------------------*/
void Classify::SetAdaptiveThreshold(float Threshold) {
  Threshold = (Threshold == matcher_good_threshold) ? 0.9f : (1 - Threshold);
  classify_adapt_proto_threshold.set_value(ClipToRange<int>(255 * Threshold, 0, 255));
  classify_adapt_feature_threshold.set_value(ClipToRange<int>(255 * Threshold, 0, 255));
} /* SetAdaptiveThreshold */
 
#ifndef GRAPHICS_DISABLED
 
/*---------------------------------------------------------------------------*/
void Classify::ShowBestMatchFor(int shape_id, const INT_FEATURE_STRUCT *features,
                                int num_features) {
  uint32_t config_mask;
  if (UnusedClassIdIn(PreTrainedTemplates, shape_id)) {
    tprintf("No built-in templates for class/shape %d\n", shape_id);
    return;
  }
  if (num_features <= 0) {
    tprintf("Illegal blob (char norm features)!\n");
    return;
  }
  UnicharRating cn_result;
  classify_norm_method.set_value(character);
  im_.Match(ClassForClassId(PreTrainedTemplates, shape_id), AllProtosOn, AllConfigsOn, num_features,
            features, &cn_result, classify_adapt_feature_threshold, NO_DEBUG,
            matcher_debug_separate_windows);
  tprintf("\n");
  config_mask = 1 << cn_result.config;
 
  tprintf("Static Shape ID: %d\n", shape_id);
  ShowMatchDisplay();
  im_.Match(ClassForClassId(PreTrainedTemplates, shape_id), AllProtosOn, &config_mask, num_features,
            features, &cn_result, classify_adapt_feature_threshold, matcher_debug_flags,
            matcher_debug_separate_windows);
  UpdateMatchDisplay();
} /* ShowBestMatchFor */
 
#endif // !GRAPHICS_DISABLED
 
// Returns a string for the classifier class_id: either the corresponding
// unicharset debug_str or the shape_table_ debug str.
std::string Classify::ClassIDToDebugStr(const INT_TEMPLATES_STRUCT *templates, int class_id,
                                        int config_id) const {
  std::string class_string;
  if (templates == PreTrainedTemplates && shape_table_ != nullptr) {
    int shape_id = ClassAndConfigIDToFontOrShapeID(class_id, config_id);
    class_string = shape_table_->DebugStr(shape_id);
  } else {
    class_string = unicharset.debug_str(class_id);
  }
  return class_string;
}
 
// Converts a classifier class_id index to a shape_table_ index
int Classify::ClassAndConfigIDToFontOrShapeID(int class_id, int int_result_config) const {
  int font_set_id = PreTrainedTemplates->Class[class_id]->font_set_id;
  // Older inttemps have no font_ids.
  if (font_set_id < 0) {
    return kBlankFontinfoId;
  }
  const FontSet &fs = fontset_table_.at(font_set_id);
  return fs.at(int_result_config);
}
 
// Converts a shape_table_ index to a classifier class_id index (not a
// unichar-id!). Uses a search, so not fast.
int Classify::ShapeIDToClassID(int shape_id) const {
  for (unsigned id = 0; id < PreTrainedTemplates->NumClasses; ++id) {
    int font_set_id = PreTrainedTemplates->Class[id]->font_set_id;
    ASSERT_HOST(font_set_id >= 0);
    const FontSet &fs = fontset_table_.at(font_set_id);
    for (auto f : fs) {
      if (f == shape_id) {
        return id;
      }
    }
  }
  tprintf("Shape %d not found\n", shape_id);
  return -1;
}
 
// Returns true if the given TEMP_CONFIG_STRUCT is good enough to make it
// a permanent config.
bool Classify::TempConfigReliable(CLASS_ID class_id, const TEMP_CONFIG_STRUCT *config) {
  if (classify_learning_debug_level >= 1) {
    tprintf("NumTimesSeen for config of %s is %d\n",
            getDict().getUnicharset().debug_str(class_id).c_str(), config->NumTimesSeen);
  }
  if (config->NumTimesSeen >= matcher_sufficient_examples_for_prototyping) {
    return true;
  } else if (config->NumTimesSeen < matcher_min_examples_for_prototyping) {
    return false;
  } else if (use_ambigs_for_adaption) {
    // Go through the ambigs vector and see whether we have already seen
    // enough times all the characters represented by the ambigs vector.
    const UnicharIdVector *ambigs = getDict().getUnicharAmbigs().AmbigsForAdaption(class_id);
    int ambigs_size = (ambigs == nullptr) ? 0 : ambigs->size();
    for (int ambig = 0; ambig < ambigs_size; ++ambig) {
      ADAPT_CLASS_STRUCT *ambig_class = AdaptedTemplates->Class[(*ambigs)[ambig]];
      assert(ambig_class != nullptr);
      if (ambig_class->NumPermConfigs == 0 &&
          ambig_class->MaxNumTimesSeen < matcher_min_examples_for_prototyping) {
        if (classify_learning_debug_level >= 1) {
          tprintf(
              "Ambig %s has not been seen enough times,"
              " not making config for %s permanent\n",
              getDict().getUnicharset().debug_str((*ambigs)[ambig]).c_str(),
              getDict().getUnicharset().debug_str(class_id).c_str());
        }
        return false;
      }
    }
  }
  return true;
}
 
void Classify::UpdateAmbigsGroup(CLASS_ID class_id, TBLOB *Blob) {
  const UnicharIdVector *ambigs = getDict().getUnicharAmbigs().ReverseAmbigsForAdaption(class_id);
  int ambigs_size = (ambigs == nullptr) ? 0 : ambigs->size();
  if (classify_learning_debug_level >= 1) {
    tprintf("Running UpdateAmbigsGroup for %s class_id=%d\n",
            getDict().getUnicharset().debug_str(class_id).c_str(), class_id);
  }
  for (int ambig = 0; ambig < ambigs_size; ++ambig) {
    CLASS_ID ambig_class_id = (*ambigs)[ambig];
    const ADAPT_CLASS_STRUCT *ambigs_class = AdaptedTemplates->Class[ambig_class_id];
    for (int cfg = 0; cfg < MAX_NUM_CONFIGS; ++cfg) {
      if (ConfigIsPermanent(ambigs_class, cfg)) {
        continue;
      }
      const TEMP_CONFIG_STRUCT *config = TempConfigFor(AdaptedTemplates->Class[ambig_class_id], cfg);
      if (config != nullptr && TempConfigReliable(ambig_class_id, config)) {
        if (classify_learning_debug_level >= 1) {
          tprintf("Making config %d of %s permanent\n", cfg,
                  getDict().getUnicharset().debug_str(ambig_class_id).c_str());
        }
        MakePermanent(AdaptedTemplates, ambig_class_id, cfg, Blob);
      }
    }
  }
}
 
} // namespace tesseract