applybox.cpp

来自「一个google的OCR源码」· C++ 代码 · 共 892 行 · 第 1/3 页

      }
      else
        ok_row_count++;
    }
  }

  min_samples = 9999;
  for (i = 0; i < unicharset_boxes.size(); i++) {
    if (tgt_char_counts[i] > labelled_char_counts[i]) {
      if (labelled_char_counts[i] <= 1) {
        tprintf
          ("APPLY_BOXES: FATALITY - %d labelled samples of \"%s\" - target is %d:\n",
          labelled_char_counts[i], unicharset_boxes.id_to_unichar(i), tgt_char_counts[i]);
        PrintString(unicharset_boxes.id_to_unichar(i));
      }
      else {
        rebalance_needed = TRUE;
        if (applybox_debug > 0)
          tprintf
            ("APPLY_BOXES: REBALANCE REQD \"%s\" - target of %d from %d labelled samples\n",
            unicharset_boxes.id_to_unichar(i), tgt_char_counts[i], labelled_char_counts[i]);
      }
    }
    if ((min_samples > labelled_char_counts[i]) && (tgt_char_counts[i] > 0)) {
      min_samples = labelled_char_counts[i];
      *min_uch_id = i;
    }
  }

  while (applybox_rebalance && rebalance_needed) {
    block_it.set_to_list (block_list);
    for (block_it.mark_cycle_pt ();
    !block_it.cycled_list (); block_it.forward ()) {
      row_it.set_to_list (block_it.data ()->row_list ());
      for (row_it.mark_cycle_pt ();
      !row_it.cycled_list (); row_it.forward ()) {
        row = row_it.data ();
        word_it.set_to_list (row->word_list ());
        for (word_it.mark_cycle_pt ();
        !word_it.cycled_list (); word_it.forward ()) {
          word = word_it.data ();
          left = word->bounding_box ().left ();
          if (*word->text () != '\0')
            uch_id = unicharset_boxes.unichar_to_id(word->text ());
          else
            uch_id = -1;
          at_dupe_of_prev_word = ((row == prev_row) &&
            (left = prev_left) &&
            (uch_id == prev_uch_id));
          if ((uch_id != -1) &&
            (labelled_char_counts[uch_id] > 1) &&
            (tgt_char_counts[uch_id] > labelled_char_counts[uch_id]) &&
          (!at_dupe_of_prev_word)) {
            /* Duplicate the word to rebalance the labelled samples */
            if (applybox_debug > 9) {
              tprintf ("Duping \"%s\" from ", unicharset_boxes.id_to_unichar(uch_id));
              word->bounding_box ().print ();
            }
            duplicate_word = new WERD;
            *duplicate_word = *word;
            word_it.add_after_then_move (duplicate_word);
            rebalance_count++;
            labelled_char_counts[uch_id]++;
          }
          prev_row = row;
          prev_left = left;
          prev_uch_id = uch_id;
        }
      }
    }
    rebalance_needed = FALSE;
    for (i = 0; i < unicharset_boxes.size(); i++) {
      if ((tgt_char_counts[i] > labelled_char_counts[i]) &&
      (labelled_char_counts[i] > 1)) {
        rebalance_needed = TRUE;
        break;
      }
    }
  }

  /* Now final check - count labelled blobs */
  final_labelled_blob_count = 0;
  block_it.set_to_list (block_list);
  for (block_it.mark_cycle_pt ();
  !block_it.cycled_list (); block_it.forward ()) {
    row_it.set_to_list (block_it.data ()->row_list ());
    for (row_it.mark_cycle_pt (); !row_it.cycled_list (); row_it.forward ()) {
      row = row_it.data ();
      word_it.set_to_list (row->word_list ());
      word_it.sort (word_comparator);
      for (word_it.mark_cycle_pt ();
      !word_it.cycled_list (); word_it.forward ()) {
        word = word_it.data ();
        if ((strlen (word->text ()) > 0) &&
          (word->gblob_list ()->length () == 1))
          final_labelled_blob_count++;
      }
    }
  }
}


void report_failed_box(inT16 boxfile_lineno,
                       inT16 boxfile_charno,
                       TBOX box,
                       const char *box_ch,
                       const char *err_msg) {
  if (applybox_debug > 4)
    tprintf ("APPLY_BOXES: boxfile %1d/%1d/%s ((%1d,%1d),(%1d,%1d)): %s\n",
      boxfile_lineno,
      boxfile_charno,
      box_ch,
      box.left (), box.bottom (), box.right (), box.top (), err_msg);
}


void apply_box_training(BLOCK_LIST *block_list) {
  BLOCK_IT block_it(block_list);
  ROW_IT row_it;
  ROW *row;
  WERD_IT word_it;
  WERD *word;
  WERD *bln_word;
  WERD copy_outword;             // copy to denorm
  PBLOB_IT blob_it;
  DENORM denorm;
  inT16 count = 0;
  char unichar[UNICHAR_LEN + 1];

  unichar[UNICHAR_LEN] = '\0';
  tprintf ("Generating training data\n");
  for (block_it.mark_cycle_pt ();
  !block_it.cycled_list (); block_it.forward ()) {
    row_it.set_to_list (block_it.data ()->row_list ());
    for (row_it.mark_cycle_pt (); !row_it.cycled_list (); row_it.forward ()) {
      row = row_it.data ();
      word_it.set_to_list (row->word_list ());
      for (word_it.mark_cycle_pt ();
      !word_it.cycled_list (); word_it.forward ()) {
        word = word_it.data ();
        if ((strlen (word->text ()) > 0) &&
        (word->gblob_list ()->length () == 1)) {
          /* Here is a word with a single unichar label and a single blob so train on it */
          bln_word =
            make_bln_copy (word, row, row->x_height (), &denorm);
          blob_it.set_to_list (bln_word->blob_list ());
          strncpy(unichar, word->text (), UNICHAR_LEN);
          tess_training_tester (blob_it.data (),
                                 //single blob
            &denorm, TRUE,       //correct
            unichar,             //correct character
            strlen(unichar),     //character length
            NULL);
          copy_outword = *(bln_word);
          copy_outword.baseline_denormalise (&denorm);
          blob_it.set_to_list (copy_outword.blob_list ());
          delete bln_word;
          count++;
        }
      }
    }
  }
  tprintf ("Generated training data for %d blobs\n", count);
}


void apply_box_testing(BLOCK_LIST *block_list) {
  BLOCK_IT block_it(block_list);
  ROW_IT row_it;
  ROW *row;
  inT16 row_count = 0;
  WERD_IT word_it;
  WERD *word;
  WERD *bln_word;
  inT16 word_count = 0;
  PBLOB_IT blob_it;
  DENORM denorm;
  inT16 count = 0;
  char ch[2];
  WERD *outword;                 //bln best choice
  //segmentation
  WERD_CHOICE *best_choice;      //tess output
  WERD_CHOICE *raw_choice;       //top choice permuter
                                 //detailed results
  BLOB_CHOICE_LIST_CLIST blob_choices;
  inT16 char_count = 0;
  inT16 correct_count = 0;
  inT16 err_count = 0;
  inT16 rej_count = 0;
  #ifndef SECURE_NAMES
  WERDSTATS wordstats;           //As from newdiff
  #endif
  char tess_rej_str[3];
  char tess_long_str[3];

  ch[1] = '\0';
  strcpy (tess_rej_str, "|A");
  strcpy (tess_long_str, "|B");

  for (block_it.mark_cycle_pt ();
  !block_it.cycled_list (); block_it.forward ()) {
    row_it.set_to_list (block_it.data ()->row_list ());
    for (row_it.mark_cycle_pt (); !row_it.cycled_list (); row_it.forward ()) {
      row = row_it.data ();
      row_count++;
      word_count = 0;
      word_it.set_to_list (row->word_list ());
      for (word_it.mark_cycle_pt ();
      !word_it.cycled_list (); word_it.forward ()) {
        word = word_it.data ();
        word_count++;
        if ((strlen (word->text ()) == 1) &&
          !STRING (applybox_test_exclusions).contains (*word->text ())
        && (word->gblob_list ()->length () == 1)) {
          /* Here is a word with a single char label and a single blob so test it */
          bln_word =
            make_bln_copy (word, row, row->x_height (), &denorm);
          blob_it.set_to_list (bln_word->blob_list ());
          ch[0] = *word->text ();
          char_count++;
          best_choice = tess_segment_pass1 (bln_word,
            &denorm,
            tess_default_matcher,
            raw_choice,
            &blob_choices, outword);

          /*
            Test for TESS screw up on word. Recog_word has already ensured that the
            choice list, outword blob lists and best_choice string are the same
            length. A TESS screw up is indicated by a blank filled or 0 length string.
          */
          if ((best_choice->lengths ().length () == 0) ||
            (strspn (best_choice->string ().string (), " ") ==
          best_choice->string ().length ())) {
            rej_count++;
            tprintf ("%d:%d: \"%s\" -> TESS FAILED\n",
              row_count, word_count, ch);
            #ifndef SECURE_NAMES
            wordstats.word (tess_rej_str, 2, ch, 1);
            #endif
          }
          else {
            if ((best_choice->lengths ().length () !=
              outword->blob_list ()->length ()) ||
              (best_choice->lengths ().length () !=
            blob_choices.length ())) {
              tprintf
                ("ASSERT FAIL String:\"%s\"; Strlen=%d; #Blobs=%d; #Choices=%d\n",
                best_choice->string ().string (),
                best_choice->lengths ().length (),
                outword->blob_list ()->length (),
                blob_choices.length ());
            }
            ASSERT_HOST (best_choice->lengths ().length () ==
              outword->blob_list ()->length ());
            ASSERT_HOST (best_choice->lengths ().length () ==
              blob_choices.length ());
            fix_quotes (best_choice,
                                 //turn to double
              outword, &blob_choices);
            if (strcmp (best_choice->string ().string (), ch) != 0) {
              err_count++;
              tprintf ("%d:%d: \"%s\" -> \"%s\"\n",
                row_count, word_count, ch,
                best_choice->string ().string ());
            }
            else
              correct_count++;
            #ifndef SECURE_NAMES
            if (best_choice->string ().length () > 2)
              wordstats.word (tess_long_str, 2, ch, 1);
            else
              wordstats.word ((char *) best_choice->string ().
                string (),
                best_choice->string ().length (), ch,
                1);
            #endif
          }
          delete bln_word;
          delete outword;
          delete best_choice;
          delete raw_choice;
          blob_choices.deep_clear ();
          count++;
        }
      }
    }
  }
  #ifndef SECURE_NAMES
  wordstats.print (1, 100.0);
  wordstats.conf_matrix ();
  tprintf ("Tested %d chars: %d correct; %d rejected by tess; %d errs\n",
    char_count, correct_count, rej_count, err_count);
  #endif
}