// Returns a random number in [-range, range].
double Network::Random(double range) {
  ASSERT_HOST(randomizer_ != NULL);
  return randomizer_->SignedRand(range);
}

void Dict::ReplaceAmbig(int wrong_ngram_begin_index, int wrong_ngram_size,
                        UNICHAR_ID correct_ngram_id, WERD_CHOICE *werd_choice,
                        MATRIX *ratings) {
  int num_blobs_to_replace = 0;
  int begin_blob_index = 0;
  int i;
  // Rating and certainty for the new BLOB_CHOICE are derived from the
  // replaced choices.
  float new_rating = 0.0f;
  float new_certainty = 0.0f;
  BLOB_CHOICE* old_choice = nullptr;
  for (i = 0; i < wrong_ngram_begin_index + wrong_ngram_size; ++i) {
    if (i >= wrong_ngram_begin_index) {
      int num_blobs = werd_choice->state(i);
      int col = begin_blob_index + num_blobs_to_replace;
      int row = col + num_blobs - 1;
      BLOB_CHOICE_LIST* choices = ratings->get(col, row);
      ASSERT_HOST(choices != nullptr);
      old_choice = FindMatchingChoice(werd_choice->unichar_id(i), choices);
      ASSERT_HOST(old_choice != nullptr);
      new_rating += old_choice->rating();
      new_certainty += old_choice->certainty();
      num_blobs_to_replace += num_blobs;
    } else {
      begin_blob_index += werd_choice->state(i);
    }
  }
  new_certainty /= wrong_ngram_size;
  // If there is no entry in the ratings matrix, add it.
  MATRIX_COORD coord(begin_blob_index,
                     begin_blob_index + num_blobs_to_replace - 1);
  if (!coord.Valid(*ratings)) {
    ratings->IncreaseBandSize(coord.row - coord.col + 1);
  }
  if (ratings->get(coord.col, coord.row) == nullptr)
    ratings->put(coord.col, coord.row, new BLOB_CHOICE_LIST);
  BLOB_CHOICE_LIST* new_choices = ratings->get(coord.col, coord.row);
  BLOB_CHOICE* choice = FindMatchingChoice(correct_ngram_id, new_choices);
  if (choice != nullptr) {
    // Already there. Upgrade if new rating better.
    if (new_rating < choice->rating())
      choice->set_rating(new_rating);
    if (new_certainty < choice->certainty())
      choice->set_certainty(new_certainty);
    // DO NOT SORT!! It will mess up the iterator in LanguageModel::UpdateState.
  } else {
    // Need a new choice with the correct_ngram_id.
    choice = new BLOB_CHOICE(*old_choice);
    choice->set_unichar_id(correct_ngram_id);
    choice->set_rating(new_rating);
    choice->set_certainty(new_certainty);
    choice->set_classifier(BCC_AMBIG);
    choice->set_matrix_cell(coord.col, coord.row);
    BLOB_CHOICE_IT it (new_choices);
    it.add_to_end(choice);
  }
  // Remove current unichar from werd_choice. On the last iteration
  // set the correct replacement unichar instead of removing a unichar.
  for (int replaced_count = 0; replaced_count < wrong_ngram_size;
       ++replaced_count) {
    if (replaced_count + 1 == wrong_ngram_size) {
      werd_choice->set_blob_choice(wrong_ngram_begin_index,
                                   num_blobs_to_replace, choice);
    } else {
      werd_choice->remove_unichar_id(wrong_ngram_begin_index + 1);
    }
  }
  if (stopper_debug_level >= 1) {
      werd_choice->print("ReplaceAmbig() ");
      tprintf("Modified blob_choices: ");
      print_ratings_list("\n", new_choices, getUnicharset());
  }
}

inT16 worst_noise_blob(WERD_RES *word_res, float *worst_noise_score) {
  PBLOB_IT blob_it;
  inT16 blob_count;
  float noise_score[512];
  int i;
  int min_noise_blob;            //1st contender
  int max_noise_blob;            //last contender
  int non_noise_count;
  int worst_noise_blob;          //Worst blob
  float small_limit = bln_x_height * fixsp_small_outlines_size;
  float non_noise_limit = bln_x_height * 0.8;

  blob_it.set_to_list (word_res->outword->blob_list ());
  //normalised
  blob_count = blob_it.length ();
  ASSERT_HOST (blob_count <= 512);
  if (blob_count < 5)
    return -1;                   //too short to split
  /* Get the noise scores for all blobs */

  #ifndef SECURE_NAMES
  if (debug_fix_space_level > 5)
    tprintf ("FP fixspace Noise metrics for \"%s\": ",
      word_res->best_choice->string ().string ());
  #endif

  for (i = 0; i < blob_count; i++, blob_it.forward ()) {
    if (word_res->reject_map[i].accepted ())
      noise_score[i] = non_noise_limit;
    else
      noise_score[i] = blob_noise_score (blob_it.data ());

    if (debug_fix_space_level > 5)
      tprintf ("%1.1f ", noise_score[i]);
  }
  if (debug_fix_space_level > 5)
    tprintf ("\n");

  /* Now find the worst one which is far enough away from the end of the word */

  non_noise_count = 0;
  for (i = 0;
  (i < blob_count) && (non_noise_count < fixsp_non_noise_limit); i++) {
    if (noise_score[i] >= non_noise_limit)
      non_noise_count++;
  }
  if (non_noise_count < fixsp_non_noise_limit)
    return -1;
  min_noise_blob = i;

  non_noise_count = 0;
  for (i = blob_count - 1;
  (i >= 0) && (non_noise_count < fixsp_non_noise_limit); i--) {
    if (noise_score[i] >= non_noise_limit)
      non_noise_count++;
  }
  if (non_noise_count < fixsp_non_noise_limit)
    return -1;
  max_noise_blob = i;

  if (min_noise_blob > max_noise_blob)
    return -1;

  *worst_noise_score = small_limit;
  worst_noise_blob = -1;
  for (i = min_noise_blob; i <= max_noise_blob; i++) {
    if (noise_score[i] < *worst_noise_score) {
      worst_noise_blob = i;
      *worst_noise_score = noise_score[i];
    }
  }
  return worst_noise_blob;
}

// Top-level method to perform splitting based on current settings.
// Returns true if a split was actually performed.
// split_for_pageseg should be true if the splitting is being done prior to
// page segmentation. This mode uses the flag
// pageseg_devanagari_split_strategy to determine the splitting strategy.
bool ShiroRekhaSplitter::Split(bool split_for_pageseg) {
  SplitStrategy split_strategy = split_for_pageseg ? pageseg_split_strategy_ :
      ocr_split_strategy_;
  if (split_strategy == NO_SPLIT) {
    return false;  // Nothing to do.
  }
  ASSERT_HOST(split_strategy == MINIMAL_SPLIT ||
              split_strategy == MAXIMAL_SPLIT);
  ASSERT_HOST(orig_pix_);
  if (devanagari_split_debuglevel > 0) {
    tprintf("Splitting shiro-rekha ...\n");
    tprintf("Split strategy = %s\n",
            split_strategy == MINIMAL_SPLIT ? "Minimal" : "Maximal");
    tprintf("Initial pageseg available = %s\n",
            segmentation_block_list_ ? "yes" : "no");
  }
  // Create a copy of original image to store the splitting output.
  pixDestroy(&splitted_image_);
  splitted_image_ = pixCopy(NULL, orig_pix_);

  // Initialize debug image if required.
  if (devanagari_split_debugimage) {
    pixDestroy(&debug_image_);
    debug_image_ = pixConvertTo32(orig_pix_);
  }

  // Determine all connected components in the input image. A close operation
  // may be required prior to this, depending on the current settings.
  Pix* pix_for_ccs = pixClone(orig_pix_);
  if (perform_close_ && global_xheight_ != kUnspecifiedXheight &&
      !segmentation_block_list_) {
    if (devanagari_split_debuglevel > 0) {
      tprintf("Performing a global close operation..\n");
    }
    // A global measure is available for xheight, but no local information
    // exists.
    pixDestroy(&pix_for_ccs);
    pix_for_ccs = pixCopy(NULL, orig_pix_);
    PerformClose(pix_for_ccs, global_xheight_);
  }
  Pixa* ccs;
  Boxa* tmp_boxa = pixConnComp(pix_for_ccs, &ccs, 8);
  boxaDestroy(&tmp_boxa);
  pixDestroy(&pix_for_ccs);

  // Iterate over all connected components. Get their bounding boxes and clip
  // out the image regions corresponding to these boxes from the original image.
  // Conditionally run splitting on each of them.
  Boxa* regions_to_clear = boxaCreate(0);
  for (int i = 0; i < pixaGetCount(ccs); ++i) {
    Box* box = ccs->boxa->box[i];
    Pix* word_pix = pixClipRectangle(orig_pix_, box, NULL);
    ASSERT_HOST(word_pix);
    int xheight = GetXheightForCC(box);
    if (xheight == kUnspecifiedXheight && segmentation_block_list_ &&
        devanagari_split_debugimage) {
      pixRenderBoxArb(debug_image_, box, 1, 255, 0, 0);
    }
    // If some xheight measure is available, attempt to pre-eliminate small
    // blobs from the shiro-rekha process. This is primarily to save the CCs
    // corresponding to punctuation marks/small dots etc which are part of
    // larger graphemes.
    if (xheight == kUnspecifiedXheight ||
        (box->w > xheight / 3 && box->h > xheight / 2)) {
      SplitWordShiroRekha(split_strategy, word_pix, xheight,
                          box->x, box->y, regions_to_clear);
    } else if (devanagari_split_debuglevel > 0) {
      tprintf("CC dropped from splitting: %d,%d (%d, %d)\n",
              box->x, box->y, box->w, box->h);
    }
    pixDestroy(&word_pix);
  }
  // Actually clear the boxes now.
  for (int i = 0; i < boxaGetCount(regions_to_clear); ++i) {
    Box* box = boxaGetBox(regions_to_clear, i, L_CLONE);
    pixClearInRect(splitted_image_, box);
    boxDestroy(&box);
  }
  boxaDestroy(&regions_to_clear);
  pixaDestroy(&ccs);
  if (devanagari_split_debugimage) {
    DumpDebugImage(split_for_pageseg ? "pageseg_split_debug.png" :
                   "ocr_split_debug.png");
  }
  return true;
}

// Attempt to improve this by adding partitions or expanding partitions.
void ColPartitionSet::ImproveColumnCandidate(WidthCallback* cb,
                                             PartSetVector* src_sets) {
  int set_size = src_sets->size();
  // Iterate over the provided column sets, as each one may have something
  // to improve this.
  for (int i = 0; i < set_size; ++i) {
    ColPartitionSet* column_set = src_sets->get(i);
    if (column_set == NULL)
      continue;
    // Iterate over the parts in this and column_set, adding bigger or
    // new parts in column_set to this.
    ColPartition_IT part_it(&parts_);
    ASSERT_HOST(!part_it.empty());
    int prev_right = MIN_INT32;
    part_it.mark_cycle_pt();
    ColPartition_IT col_it(&column_set->parts_);
    for (col_it.mark_cycle_pt(); !col_it.cycled_list(); col_it.forward()) {
      ColPartition* col_part = col_it.data();
      if (col_part->blob_type() < BRT_UNKNOWN)
        continue;  // Ignore image partitions.
      int col_left = col_part->left_key();
      int col_right = col_part->right_key();
      // Sync-up part_it (in this) so it matches the col_part in column_set.
      ColPartition* part = part_it.data();
      while (!part_it.at_last() && part->right_key() < col_left) {
        prev_right = part->right_key();
        part_it.forward();
        part = part_it.data();
      }
      int part_left = part->left_key();
      int part_right = part->right_key();
      if (part_right < col_left || col_right < part_left) {
        // There is no overlap so this is a new partition.
        AddPartition(col_part->ShallowCopy(), &part_it);
        continue;
      }
      // Check the edges of col_part to see if they can improve part.
      bool part_width_ok = cb->Run(part->KeyWidth(part_left, part_right));
      if (col_left < part_left && col_left > prev_right) {
        // The left edge of the column is better and it doesn't overlap,
        // so we can potentially expand it.
        int col_box_left = col_part->BoxLeftKey();
        bool tab_width_ok = cb->Run(part->KeyWidth(col_left, part_right));
        bool box_width_ok = cb->Run(part->KeyWidth(col_box_left, part_right));
        if (tab_width_ok || (!part_width_ok )) {
          // The tab is leaving the good column metric at least as good as
          // it was before, so use the tab.
          part->CopyLeftTab(*col_part, false);
          part->SetColumnGoodness(cb);
        } else if (col_box_left < part_left &&
                   (box_width_ok || !part_width_ok)) {
          // The box is leaving the good column metric at least as good as
          // it was before, so use the box.
          part->CopyLeftTab(*col_part, true);
          part->SetColumnGoodness(cb);
        }
        part_left = part->left_key();
      }
      if (col_right > part_right &&
          (part_it.at_last() ||
           part_it.data_relative(1)->left_key() > col_right)) {
        // The right edge is better, so we can possibly expand it.
        int col_box_right = col_part->BoxRightKey();
        bool tab_width_ok = cb->Run(part->KeyWidth(part_left, col_right));
        bool box_width_ok = cb->Run(part->KeyWidth(part_left, col_box_right));
        if (tab_width_ok || (!part_width_ok )) {
          // The tab is leaving the good column metric at least as good as
          // it was before, so use the tab.
          part->CopyRightTab(*col_part, false);
          part->SetColumnGoodness(cb);
        } else if (col_box_right > part_right &&
                   (box_width_ok || !part_width_ok)) {
          // The box is leaving the good column metric at least as good as
          // it was before, so use the box.
          part->CopyRightTab(*col_part, true);
          part->SetColumnGoodness(cb);
        }
      }
    }
  }
  ComputeCoverage();
}

/**
 * Sets up auto page segmentation, determines the orientation, and corrects it.
 * Somewhat arbitrary chunk of functionality, factored out of AutoPageSeg to
 * facilitate testing.
 * photo_mask_pix is a pointer to a NULL pointer that will be filled on return
 * with the leptonica photo mask, which must be pixDestroyed by the caller.
 * to_blocks is an empty list that will be filled with (usually a single)
 * block that is used during layout analysis. This ugly API is required
 * because of the possibility of a unlv zone file.
 * TODO(rays) clean this up.
 * See AutoPageSeg for other arguments.
 * The returned ColumnFinder must be deleted after use.
 */
ColumnFinder* Tesseract::SetupPageSegAndDetectOrientation(
    PageSegMode pageseg_mode, BLOCK_LIST* blocks, Tesseract* osd_tess,
    OSResults* osr, TO_BLOCK_LIST* to_blocks, Pix** photo_mask_pix,
    Pix** music_mask_pix) {
  int vertical_x = 0;
  int vertical_y = 1;
  TabVector_LIST v_lines;
  TabVector_LIST h_lines;
  ICOORD bleft(0, 0);

  ASSERT_HOST(pix_binary_ != NULL);
  if (tessedit_dump_pageseg_images) {
    pixa_debug_.AddPix(pix_binary_, "PageSegInput");
  }
  // Leptonica is used to find the rule/separator lines in the input.
  LineFinder::FindAndRemoveLines(source_resolution_,
                                 textord_tabfind_show_vlines, pix_binary_,
                                 &vertical_x, &vertical_y, music_mask_pix,
                                 &v_lines, &h_lines);
  if (tessedit_dump_pageseg_images) {
    pixa_debug_.AddPix(pix_binary_, "NoLines");
  }
  // Leptonica is used to find a mask of the photo regions in the input.
  *photo_mask_pix = ImageFind::FindImages(pix_binary_, &pixa_debug_);
  if (tessedit_dump_pageseg_images) {
    pixa_debug_.AddPix(pix_binary_, "NoImages");
  }
  if (!PSM_COL_FIND_ENABLED(pageseg_mode)) v_lines.clear();

  // The rest of the algorithm uses the usual connected components.
  textord_.find_components(pix_binary_, blocks, to_blocks);

  TO_BLOCK_IT to_block_it(to_blocks);
  // There must be exactly one input block.
  // TODO(rays) handle new textline finding with a UNLV zone file.
  ASSERT_HOST(to_blocks->singleton());
  TO_BLOCK* to_block = to_block_it.data();
  TBOX blkbox = to_block->block->bounding_box();
  ColumnFinder* finder = NULL;
  int estimated_resolution = source_resolution_;
  if (source_resolution_ == kMinCredibleResolution) {
    // Try to estimate resolution from typical body text size.
    int res = IntCastRounded(to_block->line_size * kResolutionEstimationFactor);
    if (res > estimated_resolution && res < kMaxCredibleResolution) {
      estimated_resolution = res;
      tprintf("Estimating resolution as %d\n", estimated_resolution);
    }
  }

  if (to_block->line_size >= 2) {
    finder = new ColumnFinder(static_cast<int>(to_block->line_size),
                              blkbox.botleft(), blkbox.topright(),
                              estimated_resolution, textord_use_cjk_fp_model,
                              textord_tabfind_aligned_gap_fraction, &v_lines,
                              &h_lines, vertical_x, vertical_y);

    finder->SetupAndFilterNoise(pageseg_mode, *photo_mask_pix, to_block);

    if (equ_detect_) {
      equ_detect_->LabelSpecialText(to_block);
    }

    BLOBNBOX_CLIST osd_blobs;
    // osd_orientation is the number of 90 degree rotations to make the
    // characters upright. (See osdetect.h for precise definition.)
    // We want the text lines horizontal, (vertical text indicates vertical
    // textlines) which may conflict (eg vertically written CJK).
    int osd_orientation = 0;
    bool vertical_text = textord_tabfind_force_vertical_text ||
                         pageseg_mode == PSM_SINGLE_BLOCK_VERT_TEXT;
    if (!vertical_text && textord_tabfind_vertical_text &&
        PSM_ORIENTATION_ENABLED(pageseg_mode)) {
      vertical_text =
          finder->IsVerticallyAlignedText(textord_tabfind_vertical_text_ratio,
                                          to_block, &osd_blobs);
    }
    if (PSM_OSD_ENABLED(pageseg_mode) && osd_tess != NULL && osr != NULL) {
      GenericVector<int> osd_scripts;
      if (osd_tess != this) {
        // We are running osd as part of layout analysis, so constrain the
        // scripts to those allowed by *this.
        AddAllScriptsConverted(unicharset, osd_tess->unicharset, &osd_scripts);
        for (int s = 0; s < sub_langs_.size(); ++s) {
          AddAllScriptsConverted(sub_langs_[s]->unicharset,
                                 osd_tess->unicharset, &osd_scripts);
        }
      }
//.........这里部分代码省略.........

//.........这里部分代码省略.........
      DoError (ILLEGALSAMPLECOUNT, "Illegal sample count");
    Proto->NumSamples = SampleCount;

    Proto->Mean = ReadNFloats (File, N, NULL);
    if (Proto->Mean == NULL)
      DoError (ILLEGALMEANSPEC, "Illegal prototype mean");

    switch (Proto->Style) {
      case spherical:
        if (ReadNFloats (File, 1, &(Proto->Variance.Spherical)) == NULL)
          DoError (ILLEGALVARIANCESPEC, "Illegal prototype variance");
        Proto->Magnitude.Spherical =
          1.0 / sqrt ((double) (2.0 * PI * Proto->Variance.Spherical));
        Proto->TotalMagnitude =
          pow (Proto->Magnitude.Spherical, (float) N);
        Proto->LogMagnitude = log ((double) Proto->TotalMagnitude);
        Proto->Weight.Spherical = 1.0 / Proto->Variance.Spherical;
        Proto->Distrib = NULL;
        break;
      case elliptical:
        Proto->Variance.Elliptical = ReadNFloats (File, N, NULL);
        if (Proto->Variance.Elliptical == NULL)
          DoError (ILLEGALVARIANCESPEC, "Illegal prototype variance");
        Proto->Magnitude.Elliptical =
          (FLOAT32 *) Emalloc (N * sizeof (FLOAT32));
        Proto->Weight.Elliptical =
          (FLOAT32 *) Emalloc (N * sizeof (FLOAT32));
        Proto->TotalMagnitude = 1.0;
        for (i = 0; i < N; i++) {
          Proto->Magnitude.Elliptical[i] =
            1.0 /
            sqrt ((double) (2.0 * PI * Proto->Variance.Elliptical[i]));
          Proto->Weight.Elliptical[i] =
            1.0 / Proto->Variance.Elliptical[i];
          Proto->TotalMagnitude *= Proto->Magnitude.Elliptical[i];
        }
        Proto->LogMagnitude = log ((double) Proto->TotalMagnitude);
        Proto->Distrib = NULL;
        break;
      case mixed:
        Proto->Distrib =
          (DISTRIBUTION *) Emalloc (N * sizeof (DISTRIBUTION));
        for (i = 0; i < N; i++) {
          if (tfscanf(File, "%s", Token) != 1)
            DoError (ILLEGALDISTRIBUTION,
              "Illegal prototype distribution");
          switch (Token[0]) {
            case 'n':
              Proto->Distrib[i] = normal;
              break;
            case 'u':
              Proto->Distrib[i] = uniform;
              break;
            case 'r':
              Proto->Distrib[i] = D_random;
              break;
            default:
              DoError (ILLEGALDISTRIBUTION,
                "Illegal prototype distribution");
          }
        }
        Proto->Variance.Elliptical = ReadNFloats (File, N, NULL);
        if (Proto->Variance.Elliptical == NULL)
          DoError (ILLEGALVARIANCESPEC, "Illegal prototype variance");
        Proto->Magnitude.Elliptical =
          (FLOAT32 *) Emalloc (N * sizeof (FLOAT32));
        Proto->Weight.Elliptical =
          (FLOAT32 *) Emalloc (N * sizeof (FLOAT32));
        Proto->TotalMagnitude = 1.0;
        for (i = 0; i < N; i++) {
          switch (Proto->Distrib[i]) {
            case normal:
              Proto->Magnitude.Elliptical[i] = 1.0 /
                sqrt ((double)
                (2.0 * PI * Proto->Variance.Elliptical[i]));
              Proto->Weight.Elliptical[i] =
                1.0 / Proto->Variance.Elliptical[i];
              break;
            case uniform:
            case D_random:
              Proto->Magnitude.Elliptical[i] = 1.0 /
                (2.0 * Proto->Variance.Elliptical[i]);
              break;
            case DISTRIBUTION_COUNT:
              ASSERT_HOST(!"Distribution count not allowed!");
          }
          Proto->TotalMagnitude *= Proto->Magnitude.Elliptical[i];
        }
        Proto->LogMagnitude = log ((double) Proto->TotalMagnitude);
        break;
    }
    return (Proto);
  }
  else if (Status == EOF)
    return (NULL);
  else {
    DoError (ILLEGALSIGNIFICANCESPEC, "Illegal significance specification");
    return (NULL);
  }
}

/**
 * Sets up auto page segmentation, determines the orientation, and corrects it.
 * Somewhat arbitrary chunk of functionality, factored out of AutoPageSeg to
 * facilitate testing.
 * photo_mask_pix is a pointer to a NULL pointer that will be filled on return
 * with the leptonica photo mask, which must be pixDestroyed by the caller.
 * to_blocks is an empty list that will be filled with (usually a single)
 * block that is used during layout analysis. This ugly API is required
 * because of the possibility of a unlv zone file.
 * TODO(rays) clean this up.
 * See AutoPageSeg for other arguments.
 * The returned ColumnFinder must be deleted after use.
 */
ColumnFinder* Tesseract::SetupPageSegAndDetectOrientation(
    bool single_column, bool osd, bool only_osd,
    BLOCK_LIST* blocks, Tesseract* osd_tess, OSResults* osr,
    TO_BLOCK_LIST* to_blocks, Pix** photo_mask_pix, Pix** music_mask_pix) {
  int vertical_x = 0;
  int vertical_y = 1;
  TabVector_LIST v_lines;
  TabVector_LIST h_lines;
  ICOORD bleft(0, 0);

  ASSERT_HOST(pix_binary_ != NULL);
  if (tessedit_dump_pageseg_images) {
    pixWrite("tessinput.png", pix_binary_, IFF_PNG);
  }
  // Leptonica is used to find the rule/separator lines in the input.
  LineFinder::FindAndRemoveLines(source_resolution_,
                                 textord_tabfind_show_vlines, pix_binary_,
                                 &vertical_x, &vertical_y, music_mask_pix,
                                 &v_lines, &h_lines);
  if (tessedit_dump_pageseg_images)
    pixWrite("tessnolines.png", pix_binary_, IFF_PNG);
  // Leptonica is used to find a mask of the photo regions in the input.
  *photo_mask_pix = ImageFind::FindImages(pix_binary_);
  if (tessedit_dump_pageseg_images)
    pixWrite("tessnoimages.png", pix_binary_, IFF_PNG);
  if (single_column)
    v_lines.clear();

  // The rest of the algorithm uses the usual connected components.
  textord_.find_components(pix_binary_, blocks, to_blocks);

  TO_BLOCK_IT to_block_it(to_blocks);
  // There must be exactly one input block.
  // TODO(rays) handle new textline finding with a UNLV zone file.
  ASSERT_HOST(to_blocks->singleton());
  TO_BLOCK* to_block = to_block_it.data();
  TBOX blkbox = to_block->block->bounding_box();
  ColumnFinder* finder = NULL;

  if (to_block->line_size >= 2) {
    finder = new ColumnFinder(static_cast<int>(to_block->line_size),
                              blkbox.botleft(), blkbox.topright(),
                              source_resolution_,
                              &v_lines, &h_lines, vertical_x, vertical_y);

    finder->SetupAndFilterNoise(*photo_mask_pix, to_block);

    if (equ_detect_) {
      equ_detect_->LabelSpecialText(to_block);
    }

    BLOBNBOX_CLIST osd_blobs;
    // osd_orientation is the number of 90 degree rotations to make the
    // characters upright. (See osdetect.h for precise definition.)
    // We want the text lines horizontal, (vertical text indicates vertical
    // textlines) which may conflict (eg vertically written CJK).
    int osd_orientation = 0;
    bool vertical_text = finder->IsVerticallyAlignedText(to_block, &osd_blobs);
    if (osd && osd_tess != NULL && osr != NULL) {
      os_detect_blobs(&osd_blobs, osr, osd_tess);
      if (only_osd) {
        delete finder;
        return NULL;
      }
      osd_orientation = osr->best_result.orientation_id;
      double osd_score = osr->orientations[osd_orientation];
      double osd_margin = min_orientation_margin * 2;
      for (int i = 0; i < 4; ++i) {
        if (i != osd_orientation &&
            osd_score - osr->orientations[i] < osd_margin) {
          osd_margin = osd_score - osr->orientations[i];
        }
      }
      if (osd_margin < min_orientation_margin) {
        // The margin is weak.
        int best_script_id = osr->best_result.script_id;
        bool cjk = (best_script_id == osd_tess->unicharset.han_sid()) ||
            (best_script_id == osd_tess->unicharset.hiragana_sid()) ||
            (best_script_id == osd_tess->unicharset.katakana_sid());

        if (!cjk && !vertical_text && osd_orientation == 2) {
          // upside down latin text is improbable with such a weak margin.
          tprintf("OSD: Weak margin (%.2f), horiz textlines, not CJK: "
                  "Don't rotate.\n", osd_margin);
          osd_orientation = 0;
        } else {
          tprintf("OSD: Weak margin (%.2f) for %d blob text block, "
//.........这里部分代码省略.........

// Returns the mean confidence of the current object at the given level.
// The number should be interpreted as a percent probability. (0.0f-100.0f)
float LTRResultIterator::Confidence(PageIteratorLevel level) const {
  if (it_->word() == NULL) return 0.0f;  // Already at the end!
  float mean_certainty = 0.0f;
  int certainty_count = 0;
  PAGE_RES_IT res_it(*it_);
  WERD_CHOICE* best_choice = res_it.word()->best_choice;
  ASSERT_HOST(best_choice != NULL);
  switch (level) {
    case RIL_BLOCK:
      do {
        best_choice = res_it.word()->best_choice;
        ASSERT_HOST(best_choice != NULL);
        mean_certainty += best_choice->certainty();
        ++certainty_count;
        res_it.forward();
      } while (res_it.block() == res_it.prev_block());
      break;
    case RIL_PARA:
      do {
        best_choice = res_it.word()->best_choice;
        ASSERT_HOST(best_choice != NULL);
        mean_certainty += best_choice->certainty();
        ++certainty_count;
        res_it.forward();
      } while (res_it.block() == res_it.prev_block() &&
               res_it.row()->row->para() == res_it.prev_row()->row->para());
      break;
    case RIL_TEXTLINE:
      do {
        best_choice = res_it.word()->best_choice;
        ASSERT_HOST(best_choice != NULL);
        mean_certainty += best_choice->certainty();
        ++certainty_count;
        res_it.forward();
      } while (res_it.row() == res_it.prev_row());
      break;
    case RIL_WORD:
      mean_certainty += best_choice->certainty();
     ++certainty_count;
      break;
    case RIL_SYMBOL:
      BLOB_CHOICE_LIST_CLIST* choices = best_choice->blob_choices();
      if (choices != NULL) {
        BLOB_CHOICE_LIST_C_IT blob_choices_it(choices);
        for (int blob = 0; blob < blob_index_; ++blob)
          blob_choices_it.forward();
        BLOB_CHOICE_IT choice_it(blob_choices_it.data());
        for (choice_it.mark_cycle_pt();
             !choice_it.cycled_list();
             choice_it.forward()) {
          if (choice_it.data()->unichar_id() ==
              best_choice->unichar_id(blob_index_))
            break;
        }
        mean_certainty += choice_it.data()->certainty();
      } else {
        mean_certainty += best_choice->certainty();
      }
      ++certainty_count;
  }
  if (certainty_count > 0) {
    mean_certainty /= certainty_count;
    float confidence = 100 + 5 * mean_certainty;
    if (confidence < 0.0f) confidence = 0.0f;
    if (confidence > 100.0f) confidence = 100.0f;
    return confidence;
  }
  return 0.0f;
}

/**
 * Split input into space-separated tokens, strip trailing punctuation
 * from each, determine case properties, call UTF-8 flavor of cost
 * function on each word, and aggregate all into single mean word
 * cost.
 */
int WordUnigrams::Cost(const char_32 *key_str32,
                       LangModel *lang_mod,
                       CharSet *char_set) const {
  if (!key_str32)
    return 0;
  // convert string to UTF8 to split into space-separated words
  string key_str;
  CubeUtils::UTF32ToUTF8(key_str32, &key_str);
  vector<string> words;
  CubeUtils::SplitStringUsing(key_str, " \t", &words);

  // no words => no cost
  if (words.size() <= 0) {
    return 0;
  }

  // aggregate the costs of all the words
  int cost = 0;
  for (int word_idx = 0; word_idx < words.size(); word_idx++) {
    // convert each word back to UTF32 for analyzing case and punctuation
    string_32 str32;
    CubeUtils::UTF8ToUTF32(words[word_idx].c_str(), &str32);
    int len = CubeUtils::StrLen(str32.c_str());

    // strip all trailing punctuation
    string clean_str;
    int clean_len = len;
    bool trunc = false;
    while (clean_len > 0 &&
           lang_mod->IsTrailingPunc(str32.c_str()[clean_len - 1])) {
      --clean_len;
      trunc = true;
    }

    // If either the original string was not truncated (no trailing
    // punctuation) or the entire string was removed (all characters
    // are trailing punctuation), evaluate original word as is;
    // otherwise, copy all but the trailing punctuation characters
    char_32 *clean_str32 = NULL;
    if (clean_len == 0 || !trunc) {
      clean_str32 = CubeUtils::StrDup(str32.c_str());
    } else {
      clean_str32 = new char_32[clean_len + 1];
      for (int i = 0; i < clean_len; ++i) {
        clean_str32[i] = str32[i];
      }
      clean_str32[clean_len] = '\0';
    }
    ASSERT_HOST(clean_str32 != NULL);

    string str8;
    CubeUtils::UTF32ToUTF8(clean_str32, &str8);
    int word_cost = CostInternal(str8.c_str());

    // if case invariant, get costs of all-upper-case and all-lower-case
    // versions and return the min cost
    if (clean_len >= kMinLengthNumOrCaseInvariant &&
        CubeUtils::IsCaseInvariant(clean_str32, char_set)) {
      char_32 *lower_32 = CubeUtils::ToLower(clean_str32, char_set);
      if (lower_32) {
        string lower_8;
        CubeUtils::UTF32ToUTF8(lower_32, &lower_8);
        word_cost = MIN(word_cost, CostInternal(lower_8.c_str()));
        delete [] lower_32;
      }
      char_32 *upper_32 = CubeUtils::ToUpper(clean_str32, char_set);
      if (upper_32) {
        string upper_8;
        CubeUtils::UTF32ToUTF8(upper_32, &upper_8);
        word_cost = MIN(word_cost, CostInternal(upper_8.c_str()));
        delete [] upper_32;
      }
    }

    if (clean_len >= kMinLengthNumOrCaseInvariant) {
      // if characters are all numeric, incur 0 word cost
      bool is_numeric = true;
      for (int i = 0; i < clean_len; ++i) {
        if (!lang_mod->IsDigit(clean_str32[i]))
          is_numeric = false;
      }
      if (is_numeric)
        word_cost = 0;
    }
    delete [] clean_str32;
    cost += word_cost;
  }  // word_idx

  // return the mean cost
  return static_cast<int>(cost / static_cast<double>(words.size()));
}

/**
 * Segment the page according to the current value of tessedit_pageseg_mode.
 * pix_binary_ is used as the source image and should not be NULL.
 * On return the blocks list owns all the constructed page layout.
 */
int Tesseract::SegmentPage(const STRING* input_file, BLOCK_LIST* blocks,
                           Tesseract* osd_tess, OSResults* osr) {
  ASSERT_HOST(pix_binary_ != NULL);
  int width = pixGetWidth(pix_binary_);
  int height = pixGetHeight(pix_binary_);
  // Get page segmentation mode.
  PageSegMode pageseg_mode = static_cast<PageSegMode>(
      static_cast<int>(tessedit_pageseg_mode));
  // If a UNLV zone file can be found, use that instead of segmentation.
  if (!PSM_COL_FIND_ENABLED(pageseg_mode) &&
      input_file != NULL && input_file->length() > 0) {
    STRING name = *input_file;
    const char* lastdot = strrchr(name.string(), '.');
    if (lastdot != NULL)
      name[lastdot - name.string()] = '\0';
    read_unlv_file(name, width, height, blocks);
  }
  if (blocks->empty()) {
    // No UNLV file present. Work according to the PageSegMode.
    // First make a single block covering the whole image.
    BLOCK_IT block_it(blocks);
    BLOCK* block = new BLOCK("", TRUE, 0, 0, 0, 0, width, height);
    block->set_right_to_left(right_to_left());
    block_it.add_to_end(block);
  } else {
    // UNLV file present. Use PSM_SINGLE_BLOCK.
    pageseg_mode = PSM_SINGLE_BLOCK;
  }
  int auto_page_seg_ret_val = 0;
  TO_BLOCK_LIST to_blocks;
  if (PSM_OSD_ENABLED(pageseg_mode) || PSM_BLOCK_FIND_ENABLED(pageseg_mode) ||
      PSM_SPARSE(pageseg_mode)) {
    auto_page_seg_ret_val =
        AutoPageSeg(pageseg_mode, blocks, &to_blocks, osd_tess, osr);
    if (pageseg_mode == PSM_OSD_ONLY)
      return auto_page_seg_ret_val;
    // To create blobs from the image region bounds uncomment this line:
    //  to_blocks.clear();  // Uncomment to go back to the old mode.
  } else {
    deskew_ = FCOORD(1.0f, 0.0f);
    reskew_ = FCOORD(1.0f, 0.0f);
    if (pageseg_mode == PSM_CIRCLE_WORD) {
      Pix* pixcleaned = RemoveEnclosingCircle(pix_binary_);
      if (pixcleaned != NULL) {
        pixDestroy(&pix_binary_);
        pix_binary_ = pixcleaned;
      }
    }
  }

  if (auto_page_seg_ret_val < 0) {
    return -1;
  }

  if (blocks->empty()) {
    if (textord_debug_tabfind)
      tprintf("Empty page\n");
    return 0;  // AutoPageSeg found an empty page.
  }

  textord_.TextordPage(pageseg_mode, width, height, pix_binary_,
                       blocks, &to_blocks);
  return auto_page_seg_ret_val;
}

//.........这里部分代码省略.........
      }
      stats_.tilde_crunch_written = true;
      stats_.last_char_was_newline = false;
      stats_.write_results_empty_block = false;
    }

    if ((word->word->flag (W_EOL) && !stats_.last_char_was_newline) || force_eol) {
      /* Add a new line output */
      txt_chs[txt_index] = '\n';
      map_chs[txt_index++] = '\n';
                                 //end line
      ep_chars[ep_chars_index++] = newline_type;

                                 //Cos of the real newline
      stats_.tilde_crunch_written = false;
      stats_.last_char_was_newline = true;
      stats_.last_char_was_tilde = false;
    }
    txt_chs[txt_index] = '\0';
    map_chs[txt_index] = '\0';
    ep_chars[ep_chars_index] = '\0';  // terminate string
    word->ep_choice = new WERD_CHOICE(ep_chars, uchset);

    if (force_eol)
      stats_.write_results_empty_block = true;
    return;
  }

  /* NORMAL PROCESSING of non tilde crunched words */

  stats_.tilde_crunch_written = false;
  if (newline_type)
    stats_.last_char_was_newline = true;
  else
    stats_.last_char_was_newline = false;
  stats_.write_results_empty_block = force_eol;  // about to write a real word

  if (unlv_tilde_crunching &&
      stats_.last_char_was_tilde &&
      (word->word->space() == 0) &&
      !(word->word->flag(W_REP_CHAR) && tessedit_write_rep_codes) &&
      (word->best_choice->unichar_id(0) == space)) {
    /* Prevent adjacent tilde across words - we know that adjacent tildes within
       words have been removed */
    word->best_choice->remove_unichar_id(0);
    if (word->best_choice->blob_choices() != NULL) {
      BLOB_CHOICE_LIST_C_IT blob_choices_it(word->best_choice->blob_choices());
      if (!blob_choices_it.empty()) delete blob_choices_it.extract();
    }
    word->reject_map.remove_pos (0);
    word->box_word->DeleteBox(0);
  }
  if (newline_type ||
    (word->word->flag (W_REP_CHAR) && tessedit_write_rep_codes))
    stats_.last_char_was_tilde = false;
  else {
    if (word->reject_map.length () > 0) {
      if (word->best_choice->unichar_id(word->reject_map.length() - 1) == space)
        stats_.last_char_was_tilde = true;
      else
        stats_.last_char_was_tilde = false;
    }
    else if (word->word->space () > 0)
      stats_.last_char_was_tilde = false;
    /* else it is unchanged as there are no output chars */
  }

  ASSERT_HOST (word->best_choice->length() == word->reject_map.length());

  set_unlv_suspects(word);
  check_debug_pt (word, 120);
  if (tessedit_rejection_debug) {
    tprintf ("Dict word: \"%s\": %d\n",
             word->best_choice->debug_string().string(),
             dict_word(*(word->best_choice)));
  }
  if (word->word->flag (W_REP_CHAR) && tessedit_write_rep_codes) {
    repetition_code = "|^~R";
    wordstr_lengths = "\001\001\001\001";
    repetition_code += uchset.id_to_unichar(get_rep_char(word));
    wordstr_lengths += strlen(uchset.id_to_unichar(get_rep_char(word)));
    wordstr = &repetition_code;
  } else {
    if (tessedit_zero_rejection) {
      /* OVERRIDE ALL REJECTION MECHANISMS - ONLY REJECT TESS FAILURES */
      for (i = 0; i < word->best_choice->length(); ++i) {
        if (word->reject_map[i].rejected())
          word->reject_map[i].setrej_minimal_rej_accept();
      }
    }
    if (tessedit_minimal_rejection) {
      /* OVERRIDE ALL REJECTION MECHANISMS - ONLY REJECT TESS FAILURES */
      for (i = 0; i < word->best_choice->length(); ++i) {
        if ((word->best_choice->unichar_id(i) != space) &&
            word->reject_map[i].rejected())
          word->reject_map[i].setrej_minimal_rej_accept();
      }
    }
  }
}

// Segment the page according to the current value of tessedit_pageseg_mode.
// If the pix_binary_ member is not NULL, it is used as the source image,
// and copied to image, otherwise it just uses image as the input.
// On return the blocks list owns all the constructed page layout.
int Tesseract::SegmentPage(const STRING* input_file,
                           IMAGE* image, BLOCK_LIST* blocks) {
  int width = image->get_xsize();
  int height = image->get_ysize();
  int resolution = image->get_res();
#ifdef HAVE_LIBLEPT
  if (pix_binary_ != NULL) {
    width = pixGetWidth(pix_binary_);
    height = pixGetHeight(pix_binary_);
    resolution = pixGetXRes(pix_binary_);
  }
#endif
  // Zero resolution messes up the algorithms, so make sure it is credible.
  if (resolution < kMinCredibleResolution)
    resolution = kDefaultResolution;
  // Get page segmentation mode.
  PageSegMode pageseg_mode = static_cast<PageSegMode>(
      static_cast<int>(tessedit_pageseg_mode));
  // If a UNLV zone file can be found, use that instead of segmentation.
  if (pageseg_mode != tesseract::PSM_AUTO &&
      input_file != NULL && input_file->length() > 0) {
    STRING name = *input_file;
    const char* lastdot = strrchr(name.string(), '.');
    if (lastdot != NULL)
      name[lastdot - name.string()] = '\0';
    read_unlv_file(name, width, height, blocks);
  }
  bool single_column = pageseg_mode > PSM_AUTO;
  if (blocks->empty()) {
    // No UNLV file present. Work according to the PageSegMode.
    // First make a single block covering the whole image.
    BLOCK_IT block_it(blocks);
    BLOCK* block = new BLOCK("", TRUE, 0, 0, 0, 0, width, height);
    block_it.add_to_end(block);
  } else {
    // UNLV file present. Use PSM_SINGLE_COLUMN.
    pageseg_mode = PSM_SINGLE_COLUMN;
  }

  TO_BLOCK_LIST land_blocks, port_blocks;
  TBOX page_box;
  if (pageseg_mode <= PSM_SINGLE_COLUMN) {
    if (AutoPageSeg(width, height, resolution, single_column,
                    image, blocks, &port_blocks) < 0) {
      return -1;
    }
    // To create blobs from the image region bounds uncomment this line:
    //  port_blocks.clear();  // Uncomment to go back to the old mode.
  } else {
#if HAVE_LIBLEPT
    image->FromPix(pix_binary_);
#endif
    deskew_ = FCOORD(1.0f, 0.0f);
    reskew_ = FCOORD(1.0f, 0.0f);
  }
  if (blocks->empty()) {
    tprintf("Empty page\n");
    return 0;  // AutoPageSeg found an empty page.
  }

  if (port_blocks.empty()) {
    // AutoPageSeg was not used, so we need to find_components first.
    find_components(blocks, &land_blocks, &port_blocks, &page_box);
  } else {
    // AutoPageSeg does not need to find_components as it did that already.
    page_box.set_left(0);
    page_box.set_bottom(0);
    page_box.set_right(width);
    page_box.set_top(height);
    // Filter_blobs sets up the TO_BLOCKs the same as find_components does.
    filter_blobs(page_box.topright(), &port_blocks, true);
  }

  TO_BLOCK_IT to_block_it(&port_blocks);
  ASSERT_HOST(!port_blocks.empty());
  TO_BLOCK* to_block = to_block_it.data();
  if (pageseg_mode <= PSM_SINGLE_BLOCK ||
      to_block->line_size < 2) {
    // For now, AUTO, SINGLE_COLUMN and SINGLE_BLOCK all map to the old
    // textord. The difference is the number of blocks and how the are made.
    textord_page(page_box.topright(), blocks, &land_blocks, &port_blocks,
                 this);
  } else {
    // SINGLE_LINE, SINGLE_WORD and SINGLE_CHAR all need a single row.
    float gradient = make_single_row(page_box.topright(),
                                     to_block, &port_blocks, this);
    if (pageseg_mode == PSM_SINGLE_LINE) {
      // SINGLE_LINE uses the old word maker on the single line.
      make_words(page_box.topright(), gradient, blocks,
                 &land_blocks, &port_blocks, this);
    } else {
      // SINGLE_WORD and SINGLE_CHAR cram all the blobs into a
      // single word, and in SINGLE_CHAR mode, all the outlines
      // go in a single blob.
      make_single_word(pageseg_mode == PSM_SINGLE_CHAR,
                       to_block->get_rows(), to_block->block->row_list());
//.........这里部分代码省略.........

// Auto page segmentation. Divide the page image into blocks of uniform
// text linespacing and images.
// Width, height and resolution are derived from the input image.
// If the pix is non-NULL, then it is assumed to be the input, and it is
// copied to the image, otherwise the image is used directly.
// The output goes in the blocks list with corresponding TO_BLOCKs in the
// to_blocks list.
// If single_column is true, then no attempt is made to divide the image
// into columns, but multiple blocks are still made if the text is of
// non-uniform linespacing.
int Tesseract::AutoPageSeg(int width, int height, int resolution,
                           bool single_column, IMAGE* image,
                           BLOCK_LIST* blocks, TO_BLOCK_LIST* to_blocks) {
  int vertical_x = 0;
  int vertical_y = 1;
  TabVector_LIST v_lines;