/**********************************************************************
* split_and_recog_word
*
* Split the word into 2 smaller pieces at the largest gap.
* Recognize the pieces and stick the results back together.
**********************************************************************/
void Tesseract::split_and_recog_word(WERD_RES *word) {
// Find the biggest blob gap in the chopped_word.
int bestgap = -MAX_INT32;
int split_index = 0;
for (int b = 1; b < word->chopped_word->NumBlobs(); ++b) {
TBOX prev_box = word->chopped_word->blobs[b - 1]->bounding_box();
TBOX blob_box = word->chopped_word->blobs[b]->bounding_box();
int gap = blob_box.left() - prev_box.right();
if (gap > bestgap) {
bestgap = gap;
split_index = b;
}
}
ASSERT_HOST(split_index > 0);
WERD_RES *word2 = NULL;
BlamerBundle *orig_bb = NULL;
split_word(word, split_index, &word2, &orig_bb);
// Recognize the first part of the word.
recog_word_recursive(word);
// Recognize the second part of the word.
recog_word_recursive(word2);
join_words(word, word2, orig_bb);
}
void Tesseract::recog_word(WERD_RES *word) {
if (wordrec_skip_no_truth_words && (word->blamer_bundle == NULL ||
word->blamer_bundle->incorrect_result_reason() == IRR_NO_TRUTH)) {
if (classify_debug_level) tprintf("No truth for word - skipping\n");
word->tess_failed = true;
return;
}
ASSERT_HOST(!word->chopped_word->blobs.empty());
recog_word_recursive(word);
word->SetupBoxWord();
if (word->best_choice->length() != word->box_word->length()) {
tprintf("recog_word ASSERT FAIL String:\"%s\"; "
"Strlen=%d; #Blobs=%d\n",
word->best_choice->debug_string().string(),
word->best_choice->length(), word->box_word->length());
}
ASSERT_HOST(word->best_choice->length() == word->box_word->length());
// Check that the ratings matrix size matches the sum of all the
// segmentation states.
if (!word->StatesAllValid()) {
tprintf("Not all words have valid states relative to ratings matrix!!");
word->DebugWordChoices(true, NULL);
ASSERT_HOST(word->StatesAllValid());
}
if (tessedit_override_permuter) {
/* Override the permuter type if a straight dictionary check disagrees. */
uinT8 perm_type = word->best_choice->permuter();
if ((perm_type != SYSTEM_DAWG_PERM) &&
(perm_type != FREQ_DAWG_PERM) && (perm_type != USER_DAWG_PERM)) {
uinT8 real_dict_perm_type = dict_word(*word->best_choice);
if (((real_dict_perm_type == SYSTEM_DAWG_PERM) ||
(real_dict_perm_type == FREQ_DAWG_PERM) ||
(real_dict_perm_type == USER_DAWG_PERM)) &&
(alpha_count(word->best_choice->unichar_string().string(),
word->best_choice->unichar_lengths().string()) > 0)) {
word->best_choice->set_permuter(real_dict_perm_type); // use dict perm
}
}
if (tessedit_rejection_debug &&
perm_type != word->best_choice->permuter()) {
tprintf("Permuter Type Flipped from %d to %d\n",
perm_type, word->best_choice->permuter());
}
}
// Factored out from control.cpp
ASSERT_HOST((word->best_choice == NULL) == (word->raw_choice == NULL));
if (word->best_choice == NULL || word->best_choice->length() == 0 ||
static_cast<int>(strspn(word->best_choice->unichar_string().string(),
" ")) == word->best_choice->length()) {
word->tess_failed = true;
word->reject_map.initialise(word->box_word->length());
word->reject_map.rej_word_tess_failure();
} else {
word->tess_failed = false;
}
}
/**********************************************************************
* recog_word
*
* Convert the word to tess form and pass it to the tess segmenter.
* Convert the output back to editor form.
**********************************************************************/
WERD_CHOICE *recog_word( //recog one owrd
WERD *word, //word to do
DENORM *denorm, //de-normaliser
POLY_MATCHER matcher, //matcher function
POLY_TESTER tester, //tester function
POLY_TESTER trainer, //trainer function
BOOL8 testing, //true if answer driven
WERD_CHOICE *&raw_choice, //raw result //list of blob lists
BLOB_CHOICE_LIST_CLIST *blob_choices,
WERD *&outword //bln word output
) {
WERD_CHOICE *word_choice;
uinT8 perm_type;
uinT8 real_dict_perm_type;
if (word->blob_list ()->empty ()) {
char empty_lengths[] = {0};
word_choice = new WERD_CHOICE ("", empty_lengths,
10.0f, -1.0f, TOP_CHOICE_PERM);
raw_choice = new WERD_CHOICE ("", empty_lengths,
10.0f, -1.0f, TOP_CHOICE_PERM);
outword = word->poly_copy (denorm->row ()->x_height ());
}
else
word_choice = recog_word_recursive (word, denorm, matcher, tester,
trainer, testing, raw_choice,
blob_choices, outword);
if ((word_choice->lengths ().length () !=
outword->blob_list ()->length ()) ||
(word_choice->lengths ().length () != blob_choices->length ())) {
tprintf
("recog_word ASSERT FAIL String:\"%s\"; Strlen=%d; #Blobs=%d; #Choices=%d\n",
word_choice->string ().string (), word_choice->lengths ().length (),
outword->blob_list ()->length (), blob_choices->length ());
}
ASSERT_HOST (word_choice->lengths ().length () ==
outword->blob_list ()->length ());
ASSERT_HOST (word_choice->lengths ().length () == blob_choices->length ());
/* Copy any reject blobs into the outword */
outword->rej_blob_list ()->deep_copy (word->rej_blob_list ());
if (tessedit_override_permuter) {
/* Override the permuter type if a straight dictionary check disagrees. */
perm_type = word_choice->permuter ();
if ((perm_type != SYSTEM_DAWG_PERM) &&
(perm_type != FREQ_DAWG_PERM) && (perm_type != USER_DAWG_PERM)) {
real_dict_perm_type = dict_word (word_choice->string ().string ());
if (((real_dict_perm_type == SYSTEM_DAWG_PERM) ||
(real_dict_perm_type == FREQ_DAWG_PERM) ||
(real_dict_perm_type == USER_DAWG_PERM)) &&
(alpha_count (word_choice->string ().string (),
word_choice->lengths ().string ()) > 0))
word_choice->set_permuter (real_dict_perm_type);
//Use dict perm
}
if (tessedit_rejection_debug && perm_type != word_choice->permuter ()) {
tprintf ("Permuter Type Flipped from %d to %d\n",
perm_type, word_choice->permuter ());
}
}
assert ((word_choice == NULL) == (raw_choice == NULL));
return word_choice;
}
WERD_CHOICE *split_and_recog_word( //recog one owrd
WERD *word, //word to do
DENORM *denorm, //de-normaliser
POLY_MATCHER matcher, //matcher function
POLY_TESTER tester, //tester function
POLY_TESTER trainer, //trainer function
BOOL8 testing, //true if answer driven
WERD_CHOICE *&raw_choice, //raw result //list of blob lists
BLOB_CHOICE_LIST_CLIST *blob_choices,
WERD *&outword //bln word output
) {
// inT32 outword1_len;
// inT32 outword2_len;
WERD *first_word; //poly copy of word
WERD *second_word; //fabricated word
WERD *outword2; //2nd output word
PBLOB *blob;
WERD_CHOICE *result; //resturn value
WERD_CHOICE *result2; //output of 2nd word
WERD_CHOICE *raw_choice2; //raw version of 2nd
float gap; //blob gap
float bestgap; //biggest gap
PBLOB_LIST new_blobs; //list of gathered blobs
PBLOB_IT blob_it;
//iterator
PBLOB_IT new_blob_it = &new_blobs;
first_word = word->poly_copy (denorm->row ()->x_height ());
blob_it.set_to_list (first_word->blob_list ());
bestgap = -MAX_INT32;
while (!blob_it.at_last ()) {
blob = blob_it.data ();
//gap to next
gap = blob_it.data_relative (1)->bounding_box ().left () - blob->bounding_box ().right ();
blob_it.forward ();
if (gap > bestgap) {
bestgap = gap; //find biggest
new_blob_it = blob_it; //save position
}
}
//take 2nd half
new_blobs.assign_to_sublist (&new_blob_it, &blob_it);
//make it a word
second_word = new WERD (&new_blobs, 1, NULL);
ASSERT_HOST (word->blob_list ()->length () ==
first_word->blob_list ()->length () +
second_word->blob_list ()->length ());
result = recog_word_recursive (first_word, denorm, matcher,
tester, trainer, testing, raw_choice,
blob_choices, outword);
delete first_word; //done that one
result2 = recog_word_recursive (second_word, denorm, matcher,
tester, trainer, testing, raw_choice2,
blob_choices, outword2);
delete second_word; //done that too
*result += *result2; //combine ratings
delete result2;
*raw_choice += *raw_choice2;
delete raw_choice2; //finished with it
// outword1_len= outword->blob_list()->length();
// outword2_len= outword2->blob_list()->length();
outword->join_on (outword2); //join words
delete outword2;
// if ( outword->blob_list()->length() != outword1_len + outword2_len )
// tprintf( "Split&Recog: part1len=%d; part2len=%d; combinedlen=%d\n",
// outword1_len, outword2_len, outword->blob_list()->length() );
// ASSERT_HOST( outword->blob_list()->length() == outword1_len + outword2_len );
return result;
}
/**
* Try splitting off the given number of (chopped) blobs from the front and
* back of the given word and recognizing the pieces.
*
* @param[in] num_chopped_leading how many chopped blobs from the left
* end of the word to chop off and try recognizing as a
* superscript (or subscript)
* @param[in] leading_certainty the (minimum) certainty had by the
* characters in the original leading section.
* @param[in] leading_pos "super" or "sub" (for debugging)
* @param[in] num_chopped_trailing how many chopped blobs from the right
* end of the word to chop off and try recognizing as a
* superscript (or subscript)
* @param[in] trailing_certainty the (minimum) certainty had by the
* characters in the original trailing section.
* @param[in] trailing_pos "super" or "sub" (for debugging)
* @param[in] word the word to try to chop up.
* @param[out] is_good do we believe our result?
* @param[out] retry_rebuild_leading, retry_rebuild_trailing
* If non-zero, and !is_good, then the caller may have luck trying
* to split the returned word with this number of (rebuilt) leading
* and trailing blobs / unichars.
* @return A word which is the result of re-recognizing as asked.
*/
WERD_RES *Tesseract::TrySuperscriptSplits(
int num_chopped_leading, float leading_certainty, ScriptPos leading_pos,
int num_chopped_trailing, float trailing_certainty,
ScriptPos trailing_pos,
WERD_RES *word,
bool *is_good,
int *retry_rebuild_leading, int *retry_rebuild_trailing) {
int num_chopped = word->chopped_word->NumBlobs();
*retry_rebuild_leading = *retry_rebuild_trailing = 0;
// Chop apart the word into up to three pieces.
BlamerBundle *bb0 = NULL;
BlamerBundle *bb1 = NULL;
WERD_RES *prefix = NULL;
WERD_RES *core = NULL;
WERD_RES *suffix = NULL;
if (num_chopped_leading > 0) {
prefix = new WERD_RES(*word);
split_word(prefix, num_chopped_leading, &core, &bb0);
} else {
core = new WERD_RES(*word);
}
if (num_chopped_trailing > 0) {
int split_pt = num_chopped - num_chopped_trailing - num_chopped_leading;
split_word(core, split_pt, &suffix, &bb1);
}
// Recognize the pieces in turn.
int saved_cp_multiplier = classify_class_pruner_multiplier;
int saved_im_multiplier = classify_integer_matcher_multiplier;
if (prefix) {
// Turn off Tesseract's y-position penalties for the leading superscript.
classify_class_pruner_multiplier.set_value(0);
classify_integer_matcher_multiplier.set_value(0);
// Adjust our expectations about the baseline for this prefix.
if (superscript_debug >= 3) {
tprintf(" recognizing first %d chopped blobs\n", num_chopped_leading);
}
recog_word_recursive(prefix);
if (superscript_debug >= 2) {
tprintf(" The leading bits look like %s %s\n",
ScriptPosToString(leading_pos),
prefix->best_choice->unichar_string().string());
}
// Restore the normal y-position penalties.
classify_class_pruner_multiplier.set_value(saved_cp_multiplier);
classify_integer_matcher_multiplier.set_value(saved_im_multiplier);
}
if (superscript_debug >= 3) {
tprintf(" recognizing middle %d chopped blobs\n",
num_chopped - num_chopped_leading - num_chopped_trailing);
}
if (suffix) {
// Turn off Tesseract's y-position penalties for the trailing superscript.
classify_class_pruner_multiplier.set_value(0);
classify_integer_matcher_multiplier.set_value(0);
if (superscript_debug >= 3) {
tprintf(" recognizing last %d chopped blobs\n", num_chopped_trailing);
}
recog_word_recursive(suffix);
if (superscript_debug >= 2) {
tprintf(" The trailing bits look like %s %s\n",
ScriptPosToString(trailing_pos),
suffix->best_choice->unichar_string().string());
}
// Restore the normal y-position penalties.
classify_class_pruner_multiplier.set_value(saved_cp_multiplier);
classify_integer_matcher_multiplier.set_value(saved_im_multiplier);
}
// Evaluate whether we think the results are believably better
// than what we already had.
bool good_prefix = !prefix || BelievableSuperscript(
superscript_debug >= 1, *prefix,
superscript_bettered_certainty * leading_certainty,
retry_rebuild_leading, NULL);
bool good_suffix = !suffix || BelievableSuperscript(
superscript_debug >= 1, *suffix,
superscript_bettered_certainty * trailing_certainty,
NULL, retry_rebuild_trailing);
*is_good = good_prefix && good_suffix;
if (!*is_good && !*retry_rebuild_leading && !*retry_rebuild_trailing) {
// None of it is any good. Quit now.
delete core;
delete prefix;
delete suffix;
return NULL;
}
recog_word_recursive(core);
// Now paste the results together into core.
if (suffix) {
suffix->SetAllScriptPositions(trailing_pos);
join_words(core, suffix, bb1);
}
if (prefix) {
prefix->SetAllScriptPositions(leading_pos);
join_words(prefix, core, bb0);
core = prefix;
prefix = NULL;
}
if (superscript_debug >= 1) {
tprintf("%s superscript fix: %s\n", *is_good ? "ACCEPT" : "REJECT",
core->best_choice->unichar_string().string());
}
return core;
}