Mercurial > hgrepos > Python2 > PyMuPDF

comparison mupdf-source/thirdparty/tesseract/src/wordrec/associate.cpp @ 2:b50eed0cc0ef upstream

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
ADD: MuPDF v1.26.7: the MuPDF source as downloaded by a default build of PyMuPDF 1.26.4. The directory name has changed: no version number in the expanded directory now.
author Franz Glasner <fzglas.hg@dom66.de>
date Mon, 15 Sep 2025 11:43:07 +0200
parents
children
comparison
equal deleted inserted replaced
1:1d09e1dec1d9 2:b50eed0cc0ef
1 ///////////////////////////////////////////////////////////////////////
2 // File: associate.cpp
3 // Description: Functions for scoring segmentation paths according to
4 // their character widths, gap widths and seam cuts.
5 // Author: Daria Antonova
6 // Created: Mon Mar 8 11:26:43 PDT 2010
7 //
8 // (C) Copyright 2010, Google Inc.
9 // Licensed under the Apache License, Version 2.0 (the "License");
10 // you may not use this file except in compliance with the License.
11 // You may obtain a copy of the License at
12 // http://www.apache.org/licenses/LICENSE-2.0
13 // Unless required by applicable law or agreed to in writing, software
14 // distributed under the License is distributed on an "AS IS" BASIS,
15 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
16 // See the License for the specific language governing permissions and
17 // limitations under the License.
18 //
19 ///////////////////////////////////////////////////////////////////////
20
21 #include <cmath>
22 #include <cstdio>
23
24 #include "associate.h"
25 #include "normalis.h"
26 #include "pageres.h"
27
28 namespace tesseract {
29
30 const float AssociateUtils::kMaxFixedPitchCharAspectRatio = 2.0f;
31 const float AssociateUtils::kMinGap = 0.03f;
32
33 void AssociateUtils::ComputeStats(int col, int row, const AssociateStats *parent_stats,
34 int parent_path_length, bool fixed_pitch, float max_char_wh_ratio,
35 WERD_RES *word_res, bool debug, AssociateStats *stats) {
36 stats->Clear();
37
38 ASSERT_HOST(word_res != nullptr);
39 if (word_res->blob_widths.empty()) {
40 return;
41 }
42 if (debug) {
43 tprintf("AssociateUtils::ComputeStats() for col=%d, row=%d%s\n", col, row,
44 fixed_pitch ? " (fixed pitch)" : "");
45 }
46 float normalizing_height = kBlnXHeight;
47 ROW *blob_row = word_res->blob_row;
48 // TODO(rays/daria) Can unicharset.script_has_xheight be useful here?
49 if (fixed_pitch && blob_row != nullptr) {
50 // For fixed pitch language like CJK, we use the full text height
51 // as the normalizing factor so we are not dependent on xheight
52 // calculation.
53 if (blob_row->body_size() > 0.0f) {
54 normalizing_height = word_res->denorm.y_scale() * blob_row->body_size();
55 } else {
56 normalizing_height =
57 word_res->denorm.y_scale() * (blob_row->x_height() + blob_row->ascenders());
58 }
59 if (debug) {
60 tprintf("normalizing height = %g (scale %g xheight %g ascenders %g)\n", normalizing_height,
61 word_res->denorm.y_scale(), blob_row->x_height(), blob_row->ascenders());
62 }
63 }
64 float wh_ratio = word_res->GetBlobsWidth(col, row) / normalizing_height;
65 if (wh_ratio > max_char_wh_ratio) {
66 stats->bad_shape = true;
67 }
68 // Compute the gap sum for this shape. If there are only negative or only
69 // positive gaps, record their sum in stats->gap_sum. However, if there is
70 // a mixture, record only the sum of the positive gaps.
71 // TODO(antonova): explain fragment.
72 int negative_gap_sum = 0;
73 for (int c = col; c < row; ++c) {
74 int gap = word_res->GetBlobsGap(c);
75 (gap > 0) ? stats->gap_sum += gap : negative_gap_sum += gap;
76 }
77 if (stats->gap_sum == 0) {
78 stats->gap_sum = negative_gap_sum;
79 }
80 if (debug) {
81 tprintf("wh_ratio=%g (max_char_wh_ratio=%g) gap_sum=%d %s\n", wh_ratio, max_char_wh_ratio,
82 stats->gap_sum, stats->bad_shape ? "bad_shape" : "");
83 }
84 // Compute shape_cost (for fixed pitch mode).
85 if (fixed_pitch) {
86 bool end_row = (row == (word_res->ratings->dimension() - 1));
87
88 // Ensure that the blob has gaps on the left and the right sides
89 // (except for beginning and ending punctuation) and that there is
90 // no cutting through ink at the blob boundaries.
91 if (col > 0) {
92 float left_gap = word_res->GetBlobsGap(col - 1) / normalizing_height;
93 SEAM *left_seam = word_res->seam_array[col - 1];
94 if ((!end_row && left_gap < kMinGap) || left_seam->priority() > 0.0f) {
95 stats->bad_shape = true;
96 }
97 if (debug) {
98 tprintf("left_gap %g, left_seam %g %s\n", left_gap, left_seam->priority(),
99 stats->bad_shape ? "bad_shape" : "");
100 }
101 }
102 float right_gap = 0.0f;
103 if (!end_row) {
104 right_gap = word_res->GetBlobsGap(row) / normalizing_height;
105 SEAM *right_seam = word_res->seam_array[row];
106 if (right_gap < kMinGap || right_seam->priority() > 0.0f) {
107 stats->bad_shape = true;
108 if (right_gap < kMinGap) {
109 stats->bad_fixed_pitch_right_gap = true;
110 }
111 }
112 if (debug) {
113 tprintf("right_gap %g right_seam %g %s\n", right_gap, right_seam->priority(),
114 stats->bad_shape ? "bad_shape" : "");
115 }
116 }
117
118 // Impose additional segmentation penalties if blob widths or gaps
119 // distribution don't fit a fixed-pitch model.
120 // Since we only know the widths and gaps of the path explored so far,
121 // the means and variances are computed for the path so far (not
122 // considering characters to the right of the last character on the path).
123 stats->full_wh_ratio = wh_ratio + right_gap;
124 if (parent_stats != nullptr) {
125 stats->full_wh_ratio_total = (parent_stats->full_wh_ratio_total + stats->full_wh_ratio);
126 float mean = stats->full_wh_ratio_total / static_cast<float>(parent_path_length + 1);
127 stats->full_wh_ratio_var =
128 parent_stats->full_wh_ratio_var + pow(mean - stats->full_wh_ratio, 2);
129 } else {
130 stats->full_wh_ratio_total = stats->full_wh_ratio;
131 }
132 if (debug) {
133 tprintf("full_wh_ratio %g full_wh_ratio_total %g full_wh_ratio_var %g\n",
134 stats->full_wh_ratio, stats->full_wh_ratio_total, stats->full_wh_ratio_var);
135 }
136
137 stats->shape_cost = FixedPitchWidthCost(wh_ratio, right_gap, end_row, max_char_wh_ratio);
138
139 // For some reason Tesseract prefers to treat the whole CJ words
140 // as one blob when the initial segmentation is particularly bad.
141 // This hack is to avoid favoring such states.
142 if (col == 0 && end_row && wh_ratio > max_char_wh_ratio) {
143 stats->shape_cost += 10;
144 }
145 stats->shape_cost += stats->full_wh_ratio_var;
146 if (debug) {
147 tprintf("shape_cost %g\n", stats->shape_cost);
148 }
149 }
150 }
151
152 float AssociateUtils::FixedPitchWidthCost(float norm_width, float right_gap, bool end_pos,
153 float max_char_wh_ratio) {
154 float cost = 0.0f;
155 if (norm_width > max_char_wh_ratio) {
156 cost += norm_width;
157 }
158 if (norm_width > kMaxFixedPitchCharAspectRatio) {
159 cost += norm_width * norm_width; // extra penalty for merging CJK chars
160 }
161 // Penalize skinny blobs, except for punctuation in the last position.
162 if (norm_width + right_gap < 0.5f && !end_pos) {
163 cost += 1.0f - (norm_width + right_gap);
164 }
165 return cost;
166 }
167
168 } // namespace tesseract