bool Dict::AcceptableChoice(const WERD_CHOICE& best_choice, XHeightConsistencyEnum xheight_consistency) { float CertaintyThreshold = stopper_nondict_certainty_base; int WordSize; if (stopper_no_acceptable_choices) return false; if (best_choice.length() == 0) return false; bool no_dang_ambigs = !best_choice.dangerous_ambig_found(); bool is_valid_word = valid_word_permuter(best_choice.permuter(), false); bool is_case_ok = case_ok(best_choice, getUnicharset()); if (stopper_debug_level >= 1) { const char *xht = "UNKNOWN"; switch (xheight_consistency) { case XH_GOOD: xht = "NORMAL"; break; case XH_SUBNORMAL: xht = "SUBNORMAL"; break; case XH_INCONSISTENT: xht = "INCONSISTENT"; break; default: xht = "UNKNOWN"; } tprintf("\nStopper: %s (word=%c, case=%c, xht_ok=%s=[%g,%g])\n", best_choice.unichar_string().string(), (is_valid_word ? 'y' : 'n'), (is_case_ok ? 'y' : 'n'), xht, best_choice.min_x_height(), best_choice.max_x_height()); } // Do not accept invalid words in PASS1. if (reject_offset_ <= 0.0f && !is_valid_word) return false; if (is_valid_word && is_case_ok) { WordSize = LengthOfShortestAlphaRun(best_choice); WordSize -= stopper_smallword_size; if (WordSize < 0) WordSize = 0; CertaintyThreshold += WordSize * stopper_certainty_per_char; } if (stopper_debug_level >= 1) tprintf("Stopper: Rating = %4.1f, Certainty = %4.1f, Threshold = %4.1f\n", best_choice.rating(), best_choice.certainty(), CertaintyThreshold); if (no_dang_ambigs && best_choice.certainty() > CertaintyThreshold && xheight_consistency < XH_INCONSISTENT && UniformCertainties(best_choice)) { return true; } else { if (stopper_debug_level >= 1) { tprintf("AcceptableChoice() returned false" " (no_dang_ambig:%d cert:%.4g thresh:%g uniform:%d)\n", no_dang_ambigs, best_choice.certainty(), CertaintyThreshold, UniformCertainties(best_choice)); } return false; } }
int Dict::UniformCertainties(const WERD_CHOICE& word) { float Certainty; float WorstCertainty = MAX_FLOAT32; float CertaintyThreshold; FLOAT64 TotalCertainty; FLOAT64 TotalCertaintySquared; FLOAT64 Variance; FLOAT32 Mean, StdDev; int word_length = word.length(); if (word_length < 3) return true; TotalCertainty = TotalCertaintySquared = 0.0; for (int i = 0; i < word_length; ++i) { Certainty = word.certainty(i); TotalCertainty += Certainty; TotalCertaintySquared += Certainty * Certainty; if (Certainty < WorstCertainty) WorstCertainty = Certainty; } // Subtract off worst certainty from statistics. word_length--; TotalCertainty -= WorstCertainty; TotalCertaintySquared -= WorstCertainty * WorstCertainty; Mean = TotalCertainty / word_length; Variance = ((word_length * TotalCertaintySquared - TotalCertainty * TotalCertainty) / (word_length * (word_length - 1))); if (Variance < 0.0) Variance = 0.0; StdDev = sqrt(Variance); CertaintyThreshold = Mean - stopper_allowable_character_badness * StdDev; if (CertaintyThreshold > stopper_nondict_certainty_base) CertaintyThreshold = stopper_nondict_certainty_base; if (word.certainty() < CertaintyThreshold) { if (stopper_debug_level >= 1) tprintf("Stopper: Non-uniform certainty = %4.1f" " (m=%4.1f, s=%4.1f, t=%4.1f)\n", word.certainty(), Mean, StdDev, CertaintyThreshold); return false; } else { return true; } }
bool Dict::AcceptableResult(const WERD_CHOICE &BestChoice) { float CertaintyThreshold = stopper_nondict_certainty_base - reject_offset_; int WordSize; if (stopper_debug_level >= 1) { tprintf("\nRejecter: %s (word=%c, case=%c, unambig=%c)\n", BestChoice.debug_string(getUnicharset()).string(), (valid_word(BestChoice) ? 'y' : 'n'), (case_ok(BestChoice, getUnicharset()) ? 'y' : 'n'), ((list_rest (best_choices_) != NIL_LIST) ? 'n' : 'y')); } if (BestChoice.length() == 0 || CurrentWordAmbig()) return false; if (BestChoice.fragment_mark()) { if (stopper_debug_level >= 1) { cprintf("AcceptableResult(): a choice with fragments beats BestChoice\n"); } return false; } if (valid_word(BestChoice) && case_ok(BestChoice, getUnicharset())) { WordSize = LengthOfShortestAlphaRun(BestChoice); WordSize -= stopper_smallword_size; if (WordSize < 0) WordSize = 0; CertaintyThreshold += WordSize * stopper_certainty_per_char; } if (stopper_debug_level >= 1) cprintf ("Rejecter: Certainty = %4.1f, Threshold = %4.1f ", BestChoice.certainty(), CertaintyThreshold); if (BestChoice.certainty() > CertaintyThreshold && !stopper_no_acceptable_choices) { if (stopper_debug_level >= 1) cprintf("ACCEPTED\n"); return true; } else { if (stopper_debug_level >= 1) cprintf("REJECTED\n"); return false; } }
// Returns an array of all word confidences, terminated by -1. int* TessBaseAPI::AllTextConfidences(PAGE_RES* page_res) { if (!page_res) return NULL; int n_word = 0; PAGE_RES_IT res_it(page_res); for (res_it.restart_page(); res_it.word () != NULL; res_it.forward()) n_word++; int* conf = new int[n_word+1]; n_word = 0; for (res_it.restart_page(); res_it.word () != NULL; res_it.forward()) { WERD_RES *word = res_it.word(); WERD_CHOICE* choice = word->best_choice; int w_conf = static_cast<int>(100 + 5 * choice->certainty()); // This is the eq for converting Tesseract confidence to 1..100 if (w_conf < 0) w_conf = 0; if (w_conf > 100) w_conf = 100; conf[n_word++] = w_conf; } conf[n_word] = -1; return conf; }