tesseract/wordrec/associate.cpp

///////////////////////////////////////////////////////////////////////
// File:        associate.cpp
// Description: Functions for scoring segmentation paths according to
//              their character widths, gap widths and seam cuts.
// Author:      Daria Antonova
// Created:     Mon Mar 8 11:26:43 PDT 2010
//
// (C) Copyright 2010, Google Inc.
// 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 <stdio.h>
#ifdef __UNIX__
#include <assert.h>
#endif
#include <math.h>

#include "associate.h"
#include "normalis.h"
#include "pageres.h"

namespace tesseract {

const float AssociateUtils::kMaxFixedPitchCharAspectRatio = 2.0f;
const float AssociateUtils::kMinGap = 0.03f;

void AssociateUtils::ComputeStats(int col, int row,
                                  const AssociateStats *parent_stats,
                                  int parent_path_length,
                                  bool fixed_pitch,
                                  float max_char_wh_ratio,
                                  WERD_RES *word_res,
                                  bool debug,
                                  AssociateStats *stats) {
  stats->Clear();

  ASSERT_HOST(word_res != nullptr);
  if (word_res->blob_widths.empty()) {
    return;
  }
  if (debug) {
    tprintf("AssociateUtils::ComputeStats() for col=%d, row=%d%s\n",
            col, row, fixed_pitch ? " (fixed pitch)" : "");
  }
  float normalizing_height = kBlnXHeight;
  ROW* blob_row = word_res->blob_row;
  // TODO(rays/daria) Can unicharset.script_has_xheight be useful here?
  if (fixed_pitch && blob_row != nullptr) {
    // For fixed pitch language like CJK, we use the full text height
    // as the normalizing factor so we are not dependent on xheight
    // calculation.
    if (blob_row->body_size() > 0.0f) {
      normalizing_height = word_res->denorm.y_scale() * blob_row->body_size();
    } else {
      normalizing_height = word_res->denorm.y_scale() *
          (blob_row->x_height() + blob_row->ascenders());
    }
    if (debug) {
      tprintf("normalizing height = %g (scale %g xheight %g ascenders %g)\n",
              normalizing_height, word_res->denorm.y_scale(),
              blob_row->x_height(), blob_row->ascenders());
    }
  }
  float wh_ratio = word_res->GetBlobsWidth(col, row) / normalizing_height;
  if (wh_ratio > max_char_wh_ratio) stats->bad_shape = true;
  // Compute the gap sum for this shape. If there are only negative or only
  // positive gaps, record their sum in stats->gap_sum. However, if there is
  // a mixture, record only the sum of the positive gaps.
  // TODO(antonova): explain fragment.
  int negative_gap_sum = 0;
  for (int c = col; c < row; ++c) {
    int gap = word_res->GetBlobsGap(c);
    (gap > 0) ? stats->gap_sum += gap : negative_gap_sum += gap;
  }
  if (stats->gap_sum == 0) stats->gap_sum = negative_gap_sum;
  if (debug) {
    tprintf("wh_ratio=%g (max_char_wh_ratio=%g) gap_sum=%d %s\n",
            wh_ratio, max_char_wh_ratio, stats->gap_sum,
            stats->bad_shape ? "bad_shape" : "");
  }
  // Compute shape_cost (for fixed pitch mode).
  if (fixed_pitch) {
    bool end_row = (row == (word_res->ratings->dimension() - 1));

    // Ensure that the blob has gaps on the left and the right sides
    // (except for beginning and ending punctuation) and that there is
    // no cutting through ink at the blob boundaries.
    if (col > 0) {
      float left_gap = word_res->GetBlobsGap(col - 1) / normalizing_height;
      SEAM *left_seam = word_res->seam_array[col - 1];
      if ((!end_row && left_gap < kMinGap) || left_seam->priority() > 0.0f) {
        stats->bad_shape = true;
      }
      if (debug) {
        tprintf("left_gap %g, left_seam %g %s\n", left_gap,
                left_seam->priority(), stats->bad_shape ? "bad_shape" : "");
      }
    }
    float right_gap = 0.0f;
    if (!end_row) {
      right_gap = word_res->GetBlobsGap(row) / normalizing_height;
      SEAM *right_seam = word_res->seam_array[row];
      if (right_gap < kMinGap || right_seam->priority() > 0.0f) {
        stats->bad_shape = true;
        if (right_gap < kMinGap) stats->bad_fixed_pitch_right_gap = true;
      }
      if (debug) {
        tprintf("right_gap %g right_seam %g %s\n", right_gap,
                right_seam->priority(), stats->bad_shape ? "bad_shape" : "");
      }
    }

    // Impose additional segmentation penalties if blob widths or gaps
    // distribution don't fit a fixed-pitch model.
    // Since we only know the widths and gaps of the path explored so far,
    // the means and variances are computed for the path so far (not
    // considering characters to the right of the last character on the path).
    stats->full_wh_ratio = wh_ratio + right_gap;
    if (parent_stats != nullptr) {
      stats->full_wh_ratio_total =
        (parent_stats->full_wh_ratio_total + stats->full_wh_ratio);
      float mean =
        stats->full_wh_ratio_total / static_cast<float>(parent_path_length+1);
      stats->full_wh_ratio_var =
        parent_stats->full_wh_ratio_var + pow(mean-stats->full_wh_ratio, 2);
    } else {
      stats->full_wh_ratio_total = stats->full_wh_ratio;
    }
    if (debug) {
      tprintf("full_wh_ratio %g full_wh_ratio_total %g full_wh_ratio_var %g\n",
              stats->full_wh_ratio, stats->full_wh_ratio_total,
              stats->full_wh_ratio_var);
    }

    stats->shape_cost =
      FixedPitchWidthCost(wh_ratio, right_gap, end_row, max_char_wh_ratio);

    // For some reason Tesseract prefers to treat the whole CJ words
    // as one blob when the initial segmentation is particularly bad.
    // This hack is to avoid favoring such states.
    if (col == 0 && end_row && wh_ratio > max_char_wh_ratio) {
      stats->shape_cost += 10;
    }
    stats->shape_cost += stats->full_wh_ratio_var;
    if (debug) tprintf("shape_cost %g\n", stats->shape_cost);
  }
}

float AssociateUtils::FixedPitchWidthCost(float norm_width,
                                          float right_gap,
                                          bool end_pos,
                                          float max_char_wh_ratio) {
  float cost = 0.0f;
  if (norm_width > max_char_wh_ratio) cost += norm_width;
  if (norm_width > kMaxFixedPitchCharAspectRatio)
    cost += norm_width * norm_width;  // extra penalty for merging CJK chars
  // Penalize skinny blobs, except for punctuation in the last position.
  if (norm_width+right_gap < 0.5f && !end_pos) {
    cost += 1.0f - (norm_width + right_gap);
  }
  return cost;
}

}  // namespace tesseract