/********************************************************************** * File: fixxht.cpp (Formerly fixxht.c) * Description: Improve x_ht and look out for case inconsistencies * Author: Phil Cheatle * Created: Thu Aug 5 14:11:08 BST 1993 * * (C) Copyright 1992, Hewlett-Packard Ltd. ** Licensed under the Apache License, Version 2.0 (the "License"); ** you may not use this file except in compliance with the License. ** You may obtain a copy of the License at ** http://www.apache.org/licenses/LICENSE-2.0 ** Unless required by applicable law or agreed to in writing, software ** distributed under the License is distributed on an "AS IS" BASIS, ** WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. ** See the License for the specific language governing permissions and ** limitations under the License. * **********************************************************************/ #include #include #include #include "params.h" #include "float2int.h" #include "tesseractclass.h" namespace tesseract { // Fixxht overview. // Premise: Initial estimate of x-height is adequate most of the time, but // occasionally it is incorrect. Most notable causes of failure are: // 1. Small caps, where the top of the caps is the same as the body text // xheight. For small caps words the xheight needs to be reduced to correctly // recognize the caps in the small caps word. // 2. All xheight lines, such as summer. Here the initial estimate will have // guessed that the blob tops are caps and will have placed the xheight too low. // 3. Noise/logos beside words, or changes in font size on a line. Such // things can blow the statistics and cause an incorrect estimate. // 4. Incorrect baseline. Can happen when 2 columns are incorrectly merged. // In this case the x-height is often still correct. // // Algorithm. // Compare the vertical position (top only) of alphnumerics in a word with // the range of positions in training data (in the unicharset). // See CountMisfitTops. If any characters disagree sufficiently with the // initial xheight estimate, then recalculate the xheight, re-run OCR on // the word, and if the number of vertical misfits goes down, along with // either the word rating or certainty, then keep the new xheight. // The new xheight is calculated as follows:ComputeCompatibleXHeight // For each alphanumeric character that has a vertically misplaced top // (a misfit), yet its bottom is within the acceptable range (ie it is not // likely a sub-or super-script) calculate the range of acceptable xheight // positions from its range of tops, and give each value in the range a // number of votes equal to the distance of its top from its acceptance range. // The x-height position with the median of the votes becomes the new // x-height. This assumes that most characters will be correctly recognized // even if the x-height is incorrect. This is not a terrible assumption, but // it is not great. An improvement would be to use a classifier that does // not care about vertical position or scaling at all. // Separately collect stats on shifted baselines and apply the same logic to // computing a best-fit shift to fix the error. If the baseline needs to be // shifted, but the x-height is OK, returns the original x-height along with // the baseline shift to indicate that recognition needs to re-run. // If the max-min top of a unicharset char is bigger than kMaxCharTopRange // then the char top cannot be used to judge misfits or suggest a new top. const int kMaxCharTopRange = 48; // Returns the number of misfit blob tops in this word. int Tesseract::CountMisfitTops(WERD_RES *word_res) { int bad_blobs = 0; int num_blobs = word_res->rebuild_word->NumBlobs(); for (int blob_id = 0; blob_id < num_blobs; ++blob_id) { TBLOB* blob = word_res->rebuild_word->blobs[blob_id]; UNICHAR_ID class_id = word_res->best_choice->unichar_id(blob_id); if (unicharset.get_isalpha(class_id) || unicharset.get_isdigit(class_id)) { int top = blob->bounding_box().top(); if (top >= INT_FEAT_RANGE) top = INT_FEAT_RANGE - 1; int min_bottom, max_bottom, min_top, max_top; unicharset.get_top_bottom(class_id, &min_bottom, &max_bottom, &min_top, &max_top); if (max_top - min_top > kMaxCharTopRange) continue; bool bad = top < min_top - x_ht_acceptance_tolerance || top > max_top + x_ht_acceptance_tolerance; if (bad) ++bad_blobs; if (debug_x_ht_level >= 1) { tprintf("Class %s is %s with top %d vs limits of %d->%d, +/-%d\n", unicharset.id_to_unichar(class_id), bad ? "Misfit" : "OK", top, min_top, max_top, static_cast(x_ht_acceptance_tolerance)); } } } return bad_blobs; } // Returns a new x-height maximally compatible with the result in word_res. // See comment above for overall algorithm. float Tesseract::ComputeCompatibleXheight(WERD_RES *word_res, float* baseline_shift) { STATS top_stats(0, UINT8_MAX); STATS shift_stats(-UINT8_MAX, UINT8_MAX); int bottom_shift = 0; int num_blobs = word_res->rebuild_word->NumBlobs(); do { top_stats.clear(); shift_stats.clear(); for (int blob_id = 0; blob_id < num_blobs; ++blob_id) { TBLOB* blob = word_res->rebuild_word->blobs[blob_id]; UNICHAR_ID class_id = word_res->best_choice->unichar_id(blob_id); if (unicharset.get_isalpha(class_id) || unicharset.get_isdigit(class_id)) { int top = blob->bounding_box().top() + bottom_shift; // Clip the top to the limit of normalized feature space. if (top >= INT_FEAT_RANGE) top = INT_FEAT_RANGE - 1; int bottom = blob->bounding_box().bottom() + bottom_shift; int min_bottom, max_bottom, min_top, max_top; unicharset.get_top_bottom(class_id, &min_bottom, &max_bottom, &min_top, &max_top); // Chars with a wild top range would mess up the result so ignore them. if (max_top - min_top > kMaxCharTopRange) continue; int misfit_dist = std::max((min_top - x_ht_acceptance_tolerance) - top, top - (max_top + x_ht_acceptance_tolerance)); int height = top - kBlnBaselineOffset; if (debug_x_ht_level >= 2) { tprintf("Class %s: height=%d, bottom=%d,%d top=%d,%d, actual=%d,%d: ", unicharset.id_to_unichar(class_id), height, min_bottom, max_bottom, min_top, max_top, bottom, top); } // Use only chars that fit in the expected bottom range, and where // the range of tops is sensibly near the xheight. if (min_bottom <= bottom + x_ht_acceptance_tolerance && bottom - x_ht_acceptance_tolerance <= max_bottom && min_top > kBlnBaselineOffset && max_top - kBlnBaselineOffset >= kBlnXHeight && misfit_dist > 0) { // Compute the x-height position using proportionality between the // actual height and expected height. int min_xht = DivRounded(height * kBlnXHeight, max_top - kBlnBaselineOffset); int max_xht = DivRounded(height * kBlnXHeight, min_top - kBlnBaselineOffset); if (debug_x_ht_level >= 2) { tprintf(" xht range min=%d, max=%d\n", min_xht, max_xht); } // The range of expected heights gets a vote equal to the distance // of the actual top from the expected top. for (int y = min_xht; y <= max_xht; ++y) top_stats.add(y, misfit_dist); } else if ((min_bottom > bottom + x_ht_acceptance_tolerance || bottom - x_ht_acceptance_tolerance > max_bottom) && bottom_shift == 0) { // Get the range of required bottom shift. int min_shift = min_bottom - bottom; int max_shift = max_bottom - bottom; if (debug_x_ht_level >= 2) { tprintf(" bottom shift min=%d, max=%d\n", min_shift, max_shift); } // The range of expected shifts gets a vote equal to the min distance // of the actual bottom from the expected bottom, spread over the // range of its acceptance. int misfit_weight = abs(min_shift); if (max_shift > min_shift) misfit_weight /= max_shift - min_shift; for (int y = min_shift; y <= max_shift; ++y) shift_stats.add(y, misfit_weight); } else { if (bottom_shift == 0) { // Things with bottoms that are already ok need to say so, on the // 1st iteration only. shift_stats.add(0, kBlnBaselineOffset); } if (debug_x_ht_level >= 2) { tprintf(" already OK\n"); } } } } if (shift_stats.get_total() > top_stats.get_total()) { bottom_shift = IntCastRounded(shift_stats.median()); if (debug_x_ht_level >= 2) { tprintf("Applying bottom shift=%d\n", bottom_shift); } } } while (bottom_shift != 0 && top_stats.get_total() < shift_stats.get_total()); // Baseline shift is opposite sign to the bottom shift. *baseline_shift = -bottom_shift / word_res->denorm.y_scale(); if (debug_x_ht_level >= 2) { tprintf("baseline shift=%g\n", *baseline_shift); } if (top_stats.get_total() == 0) return bottom_shift != 0 ? word_res->x_height : 0.0f; // The new xheight is just the median vote, which is then scaled out // of BLN space back to pixel space to get the x-height in pixel space. float new_xht = top_stats.median(); if (debug_x_ht_level >= 2) { tprintf("Median xht=%f\n", new_xht); tprintf("Mode20:A: New x-height = %f (norm), %f (orig)\n", new_xht, new_xht / word_res->denorm.y_scale()); } // The xheight must change by at least x_ht_min_change to be used. if (fabs(new_xht - kBlnXHeight) >= x_ht_min_change) return new_xht / word_res->denorm.y_scale(); else return bottom_shift != 0 ? word_res->x_height : 0.0f; } } // namespace tesseract