/**
* dawg_permute_and_select
*
* Recursively explore all the possible character combinations in
* the given char_choices. Use go_deeper_dawg_fxn() to search all the
* dawgs in the dawgs_ vector in parallel and discard invalid words.
*
* Allocate and return a WERD_CHOICE with the best valid word found.
*/
WERD_CHOICE *Dict::dawg_permute_and_select(
const BLOB_CHOICE_LIST_VECTOR &char_choices, float rating_limit) {
WERD_CHOICE *best_choice = new WERD_CHOICE();
best_choice->make_bad();
best_choice->set_rating(rating_limit);
if (char_choices.length() == 0) return best_choice;
DawgInfoVector *active_dawgs = new DawgInfoVector[char_choices.length() + 1];
DawgInfoVector *constraints = new DawgInfoVector[char_choices.length() + 1];
init_active_dawgs(&(active_dawgs[0]));
init_constraints(&(constraints[0]));
DawgArgs dawg_args(&(active_dawgs[0]), &(constraints[0]),
&(active_dawgs[1]), &(constraints[1]),
(segment_penalty_dict_case_bad /
segment_penalty_dict_case_ok));
WERD_CHOICE word(MAX_WERD_LENGTH);
copy_hyphen_info(&word);
// Discard rating and certainty of the hyphen base (if any).
word.set_rating(0.0);
word.set_certainty(0.0);
if (word.length() + char_choices.length() > MAX_WERD_LENGTH) {
delete[] active_dawgs;
delete[] constraints;
return best_choice; // the word is too long to permute
}
float certainties[MAX_WERD_LENGTH];
this->go_deeper_fxn_ = &tesseract::Dict::go_deeper_dawg_fxn;
permute_choices(segment_dawg_debug ? "segment_dawg_debug" : NULL,
char_choices, 0, NULL, &word, certainties,
&rating_limit, best_choice, &dawg_args);
delete[] active_dawgs;
delete[] constraints;
return best_choice;
}
/**
* append_choices
*
* Checks to see whether or not the next choice is worth appending to
* the word being generated. If so then keeps going deeper into the word.
*
* This function assumes that Dict::go_deeper_fxn_ is set.
*/
void Dict::append_choices(
const char *debug,
const BLOB_CHOICE_LIST_VECTOR &char_choices,
const BLOB_CHOICE &blob_choice,
int char_choice_index,
const CHAR_FRAGMENT_INFO *prev_char_frag_info,
WERD_CHOICE *word,
float certainties[],
float *limit,
WERD_CHOICE *best_choice,
int *attempts_left,
void *more_args) {
int word_ending =
(char_choice_index == char_choices.length() - 1) ? true : false;
// Deal with fragments.
CHAR_FRAGMENT_INFO char_frag_info;
if (!fragment_state_okay(blob_choice.unichar_id(), blob_choice.rating(),
blob_choice.certainty(), prev_char_frag_info, debug,
word_ending, &char_frag_info)) {
return; // blob_choice must be an invalid fragment
}
// Search the next letter if this character is a fragment.
if (char_frag_info.unichar_id == INVALID_UNICHAR_ID) {
permute_choices(debug, char_choices, char_choice_index + 1,
&char_frag_info, word, certainties, limit,
best_choice, attempts_left, more_args);
return;
}
// Add the next unichar.
float old_rating = word->rating();
float old_certainty = word->certainty();
uint8_t old_permuter = word->permuter();
certainties[word->length()] = char_frag_info.certainty;
word->append_unichar_id_space_allocated(
char_frag_info.unichar_id, char_frag_info.num_fragments,
char_frag_info.rating, char_frag_info.certainty);
// Explore the next unichar.
(this->*go_deeper_fxn_)(debug, char_choices, char_choice_index,
&char_frag_info, word_ending, word, certainties,
limit, best_choice, attempts_left, more_args);
// Remove the unichar we added to explore other choices in it's place.
word->remove_last_unichar_id();
word->set_rating(old_rating);
word->set_certainty(old_certainty);
word->set_permuter(old_permuter);
}
/**
* dawg_permute_and_select
*
* Recursively explore all the possible character combinations in
* the given char_choices. Use go_deeper_dawg_fxn() to search all the
* dawgs in the dawgs_ vector in parallel and discard invalid words.
*
* Allocate and return a WERD_CHOICE with the best valid word found.
*/
WERD_CHOICE *Dict::dawg_permute_and_select(
const BLOB_CHOICE_LIST_VECTOR &char_choices, float rating_limit) {
WERD_CHOICE *best_choice = new WERD_CHOICE(&getUnicharset());
best_choice->make_bad();
best_choice->set_rating(rating_limit);
if (char_choices.length() == 0 || char_choices.length() > MAX_WERD_LENGTH)
return best_choice;
DawgPositionVector *active_dawgs =
new DawgPositionVector[char_choices.length() + 1];
init_active_dawgs(&(active_dawgs[0]), true);
DawgArgs dawg_args(&(active_dawgs[0]), &(active_dawgs[1]), NO_PERM);
WERD_CHOICE word(&getUnicharset(), MAX_WERD_LENGTH);
float certainties[MAX_WERD_LENGTH];
this->go_deeper_fxn_ = &tesseract::Dict::go_deeper_dawg_fxn;
int attempts_left = max_permuter_attempts;
permute_choices((dawg_debug_level) ? "permute_dawg_debug" : nullptr,
char_choices, 0, nullptr, &word, certainties, &rating_limit, best_choice,
&attempts_left, &dawg_args);
delete[] active_dawgs;
return best_choice;
}
/**
* @name go_deeper_dawg_fxn
*
* If the choice being composed so far could be a dictionary word
* keep exploring choices.
*/
void Dict::go_deeper_dawg_fxn(
const char *debug, const BLOB_CHOICE_LIST_VECTOR &char_choices,
int char_choice_index, const CHAR_FRAGMENT_INFO *prev_char_frag_info,
bool word_ending, WERD_CHOICE *word, float certainties[], float *limit,
WERD_CHOICE *best_choice, int *attempts_left, void *void_more_args) {
DawgArgs *more_args = static_cast<DawgArgs *>(void_more_args);
word_ending = (char_choice_index == char_choices.size()-1);
int word_index = word->length() - 1;
if (best_choice->rating() < *limit) return;
// Look up char in DAWG
// If the current unichar is an ngram first try calling
// letter_is_okay() for each unigram it contains separately.
UNICHAR_ID orig_uch_id = word->unichar_id(word_index);
bool checked_unigrams = false;
if (getUnicharset().get_isngram(orig_uch_id)) {
if (dawg_debug_level) {
tprintf("checking unigrams in an ngram %s\n",
getUnicharset().debug_str(orig_uch_id).string());
}
int num_unigrams = 0;
word->remove_last_unichar_id();
GenericVector<UNICHAR_ID> encoding;
const char *ngram_str = getUnicharset().id_to_unichar(orig_uch_id);
// Since the string came out of the unicharset, failure is impossible.
ASSERT_HOST(getUnicharset().encode_string(ngram_str, true, &encoding, nullptr,
nullptr));
bool unigrams_ok = true;
// Construct DawgArgs that reflect the current state.
DawgPositionVector unigram_active_dawgs = *(more_args->active_dawgs);
DawgPositionVector unigram_updated_dawgs;
DawgArgs unigram_dawg_args(&unigram_active_dawgs,
&unigram_updated_dawgs,
more_args->permuter);
// Check unigrams in the ngram with letter_is_okay().
for (int i = 0; unigrams_ok && i < encoding.size(); ++i) {
UNICHAR_ID uch_id = encoding[i];
ASSERT_HOST(uch_id != INVALID_UNICHAR_ID);
++num_unigrams;
word->append_unichar_id(uch_id, 1, 0.0, 0.0);
unigrams_ok = (this->*letter_is_okay_)(
&unigram_dawg_args,
word->unichar_id(word_index+num_unigrams-1),
word_ending && i == encoding.size() - 1);
(*unigram_dawg_args.active_dawgs) = *(unigram_dawg_args.updated_dawgs);
if (dawg_debug_level) {
tprintf("unigram %s is %s\n",
getUnicharset().debug_str(uch_id).string(),
unigrams_ok ? "OK" : "not OK");
}
}
// Restore the word and copy the updated dawg state if needed.
while (num_unigrams-- > 0) word->remove_last_unichar_id();
word->append_unichar_id_space_allocated(orig_uch_id, 1, 0.0, 0.0);
if (unigrams_ok) {
checked_unigrams = true;
more_args->permuter = unigram_dawg_args.permuter;
*(more_args->updated_dawgs) = *(unigram_dawg_args.updated_dawgs);
}
}
// Check which dawgs from the dawgs_ vector contain the word
// up to and including the current unichar.
if (checked_unigrams || (this->*letter_is_okay_)(
more_args, word->unichar_id(word_index), word_ending)) {
// Add a new word choice
if (word_ending) {
if (dawg_debug_level) {
tprintf("found word = %s\n", word->debug_string().string());
}
if (strcmp(output_ambig_words_file.string(), "") != 0) {
if (output_ambig_words_file_ == nullptr) {
output_ambig_words_file_ =
fopen(output_ambig_words_file.string(), "wb+");
if (output_ambig_words_file_ == nullptr) {
tprintf("Failed to open output_ambig_words_file %s\n",
output_ambig_words_file.string());
exit(1);
}
STRING word_str;
word->string_and_lengths(&word_str, nullptr);
word_str += " ";
fprintf(output_ambig_words_file_, "%s", word_str.string());
}
STRING word_str;
word->string_and_lengths(&word_str, nullptr);
word_str += " ";
fprintf(output_ambig_words_file_, "%s", word_str.string());
}
WERD_CHOICE *adjusted_word = word;
adjusted_word->set_permuter(more_args->permuter);
update_best_choice(*adjusted_word, best_choice);
} else { // search the next letter
// Make updated_* point to the next entries in the DawgPositionVector
// arrays (that were originally created in dawg_permute_and_select)
++(more_args->updated_dawgs);
// Make active_dawgs and constraints point to the updated ones.
++(more_args->active_dawgs);
permute_choices(debug, char_choices, char_choice_index + 1,
prev_char_frag_info, word, certainties, limit,
best_choice, attempts_left, more_args);
// Restore previous state to explore another letter in this position.
--(more_args->updated_dawgs);
--(more_args->active_dawgs);
}
} else {
if (dawg_debug_level) {
tprintf("last unichar not OK at index %d in %s\n",
word_index, word->debug_string().string());
}
}
}
/**
* @name go_deeper_dawg_fxn
*
* If the choice being composed so far could be a dictionary word
* keep exploring choices.
*
* There are two modes for deciding whether to go deeper: regular dawg
* permuter mode and the special ambigs mode. If *limit is <= 0.0 the
* function switches to the ambigs mode (this is the case when
* dawg_permute_and_select() function is called from NoDangerousAmbigs()) and
* only searches for the first choice that has a rating better than *limit
* (in this case ratings are fake, since the real ratings can not be < 0).
* Modification of the hyphen state is turned off in the ambigs mode.
* When in the regular dawg permuter mode, the function explores all the
* possible words and chooses the one with the best rating. The letters with
* ratings that are far worse than the ones seen so far are pruned out.
*/
void Dict::go_deeper_dawg_fxn(
const char *debug, const BLOB_CHOICE_LIST_VECTOR &char_choices,
int char_choice_index,
const CHAR_FRAGMENT_INFO *prev_char_frag_info,
bool word_ending, WERD_CHOICE *word, float certainties[],
float *limit, WERD_CHOICE *best_choice, void *void_more_args) {
DawgArgs *more_args = reinterpret_cast<DawgArgs*>(void_more_args);
int word_index = word->length() - 1;
bool ambigs_mode = (*limit <= 0.0);
if (ambigs_mode) {
if (best_choice->rating() < *limit) return;
} else {
// Prune bad subwords
if (more_args->rating_array[word_index] == NO_RATING) {
more_args->rating_array[word_index] = word->rating();
} else {
float permdawg_limit = more_args->rating_array[word_index] *
more_args->rating_margin + kPermDawgRatingPad;
if (permdawg_limit < word->rating()) {
if (segment_dawg_debug) {
tprintf("early pruned word rating=%4.2f,"
" permdawg_limit=%4.2f, word=%s\n", word->rating(),
permdawg_limit, word->debug_string(getUnicharset()).string());
}
return;
}
}
}
// Deal with hyphens
if (word_ending && has_hyphen_end(*word) && !ambigs_mode) {
if (segment_dawg_debug)
tprintf("new hyphen choice = %s\n",
word->debug_string(getUnicharset()).string());
word->set_permuter(more_args->permuter);
adjust_word(word, certainties);
set_hyphen_word(*word, *(more_args->active_dawgs),
*(more_args->constraints));
update_best_choice(*word, best_choice);
} else { // Look up char in DAWG
// TODO(daria): update the rest of the code that specifies alternative
// letter_is_okay_ functions (e.g. TessCharNgram class) to work with
// multi-byte unichars and/or unichar ids.
// If the current unichar is an ngram first try calling
// letter_is_okay() for each unigram it contains separately.
UNICHAR_ID orig_uch_id = word->unichar_id(word_index);
bool checked_unigrams = false;
if (getUnicharset().get_isngram(orig_uch_id)) {
if (segment_dawg_debug) {
tprintf("checking unigrams in an ngram %s\n",
getUnicharset().debug_str(orig_uch_id).string());
}
int orig_num_fragments = word->fragment_length(word_index);
int num_unigrams = 0;
word->remove_last_unichar_id();
const char *ngram_str = getUnicharset().id_to_unichar(orig_uch_id);
const char *ngram_str_end = ngram_str + strlen(ngram_str);
const char *ngram_ptr = ngram_str;
bool unigrams_ok = true;
// Construct DawgArgs that reflect the current state.
DawgInfoVector unigram_active_dawgs = *(more_args->active_dawgs);
DawgInfoVector unigram_constraints = *(more_args->constraints);
DawgInfoVector unigram_updated_active_dawgs;
DawgInfoVector unigram_updated_constraints;
DawgArgs unigram_dawg_args(&unigram_active_dawgs, &unigram_constraints,
&unigram_updated_active_dawgs,
&unigram_updated_constraints, 0.0);
unigram_dawg_args.permuter = more_args->permuter;
// Check unigrams in the ngram with letter_is_okay().
while (unigrams_ok && ngram_ptr < ngram_str_end) {
int step = getUnicharset().step(ngram_ptr);
UNICHAR_ID uch_id = (step <= 0) ? INVALID_UNICHAR_ID :
getUnicharset().unichar_to_id(ngram_ptr, step);
ngram_ptr += step;
++num_unigrams;
word->append_unichar_id(uch_id, 1, 0.0, 0.0);
unigrams_ok = unigrams_ok && (this->*letter_is_okay_)(
&unigram_dawg_args, word_index+num_unigrams-1, word,
word_ending && (ngram_ptr == ngram_str_end));
(*unigram_dawg_args.active_dawgs) =
*(unigram_dawg_args.updated_active_dawgs);
(*unigram_dawg_args.constraints) =
*(unigram_dawg_args.updated_constraints);
if (segment_dawg_debug) {
tprintf("unigram %s is %s\n",
getUnicharset().debug_str(uch_id).string(),
unigrams_ok ? "OK" : "not OK");
}
}
// Restore the word and copy the updated dawg state if needed.
while (num_unigrams-- > 0) word->remove_last_unichar_id();
word->append_unichar_id_space_allocated(
orig_uch_id, orig_num_fragments, 0.0, 0.0);
if (unigrams_ok) {
checked_unigrams = true;
more_args->permuter = unigram_dawg_args.permuter;
*(more_args->updated_active_dawgs) =
*(unigram_dawg_args.updated_active_dawgs);
*(more_args->updated_constraints) =
*(unigram_dawg_args.updated_constraints);
}
}
// Check which dawgs from dawgs_ vector contain the word
// up to and including the current unichar.
if (checked_unigrams ||
(this->*letter_is_okay_)(more_args, word_index, word, word_ending)) {
// Add a new word choice
if (word_ending) {
if (segment_dawg_debug) {
tprintf("found word = %s\n",
word->debug_string(getUnicharset()).string());
}
WERD_CHOICE *adjusted_word = word;
WERD_CHOICE hyphen_tail_word;
if (!ambigs_mode && hyphen_base_size() > 0) {
hyphen_tail_word = *word;
remove_hyphen_head(&hyphen_tail_word);
adjusted_word = &hyphen_tail_word;
}
adjusted_word->set_permuter(more_args->permuter);
if (!ambigs_mode) {
adjust_word(adjusted_word, &certainties[hyphen_base_size()]);
}
update_best_choice(*adjusted_word, best_choice);
} else { // search the next letter
// Make updated_* point to the next entries in the DawgInfoVector
// arrays (that were originally created in dawg_permute_and_select)
++(more_args->updated_active_dawgs);
++(more_args->updated_constraints);
// Make active_dawgs and constraints point to the updated ones.
++(more_args->active_dawgs);
++(more_args->constraints);
permute_choices(debug, char_choices, char_choice_index + 1,
prev_char_frag_info, word, certainties, limit,
best_choice, more_args);
// Restore previous state to explore another letter in this position.
--(more_args->updated_active_dawgs);
--(more_args->updated_constraints);
--(more_args->active_dawgs);
--(more_args->constraints);
}
} else {
if (segment_dawg_debug) {
tprintf("last unichar not OK at index %d in %s\n",
word_index, word->debug_string(getUnicharset()).string());
}
}
}
}
