Mercurial > hgrepos > Python2 > PyMuPDF

comparison mupdf-source/thirdparty/tesseract/src/wordrec/segsearch.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: segsearch.cpp
3 // Description: Segmentation search functions.
4 // Author: Daria Antonova
5 //
6 // (C) Copyright 2009, Google Inc.
7 // Licensed under the Apache License, Version 2.0 (the "License");
8 // you may not use this file except in compliance with the License.
9 // You may obtain a copy of the License at
10 // http://www.apache.org/licenses/LICENSE-2.0
11 // Unless required by applicable law or agreed to in writing, software
12 // distributed under the License is distributed on an "AS IS" BASIS,
13 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14 // See the License for the specific language governing permissions and
15 // limitations under the License.
16 //
17 ///////////////////////////////////////////////////////////////////////
18
19 #include <cstdint> // for INT32_MAX
20 #include "blamer.h" // for BlamerBundle
21 #include "errcode.h" // for ASSERT_HOST
22 #include "lm_pain_points.h" // for LMPainPoints, LM_PPTYPE_SHAPE, LMPainPoi...
23 #include "lm_state.h" // for BestChoiceBundle, ViterbiStateEntry
24 #include "matrix.h" // for MATRIX_COORD, MATRIX
25 #include "pageres.h" // for WERD_RES
26 #include "params.h" // for BoolParam, IntParam, DoubleParam
27 #include "ratngs.h" // for BLOB_CHOICE_LIST, BLOB_CHOICE_IT
28 #include "tprintf.h" // for tprintf
29 #include "wordrec.h" // for Wordrec, SegSearchPending (ptr only)
30
31 namespace tesseract {
32
33 void Wordrec::SegSearch(WERD_RES *word_res, BestChoiceBundle *best_choice_bundle,
34 BlamerBundle *blamer_bundle) {
35 LMPainPoints pain_points(segsearch_max_pain_points, segsearch_max_char_wh_ratio,
36 assume_fixed_pitch_char_segment, &getDict(), segsearch_debug_level);
37 // Compute scaling factor that will help us recover blob outline length
38 // from classifier rating and certainty for the blob.
39 float rating_cert_scale = -1.0 * getDict().certainty_scale / rating_scale;
40 std::vector<SegSearchPending> pending;
41 InitialSegSearch(word_res, &pain_points, &pending, best_choice_bundle, blamer_bundle);
42
43 if (!SegSearchDone(0)) { // find a better choice
44 if (chop_enable && word_res->chopped_word != nullptr) {
45 improve_by_chopping(rating_cert_scale, word_res, best_choice_bundle, blamer_bundle,
46 &pain_points, &pending);
47 }
48 if (chop_debug) {
49 SEAM::PrintSeams("Final seam list:", word_res->seam_array);
50 }
51
52 if (blamer_bundle != nullptr && !blamer_bundle->ChoiceIsCorrect(word_res->best_choice)) {
53 blamer_bundle->SetChopperBlame(word_res, wordrec_debug_blamer);
54 }
55 }
56 // Keep trying to find a better path by fixing the "pain points".
57
58 MATRIX_COORD pain_point;
59 float pain_point_priority;
60 int num_futile_classifications = 0;
61 std::string blamer_debug;
62 while (wordrec_enable_assoc &&
63 (!SegSearchDone(num_futile_classifications) ||
64 (blamer_bundle != nullptr && blamer_bundle->GuidedSegsearchStillGoing()))) {
65 // Get the next valid "pain point".
66 bool found_nothing = true;
67 LMPainPointsType pp_type;
68 while ((pp_type = pain_points.Deque(&pain_point, &pain_point_priority)) != LM_PPTYPE_NUM) {
69 if (!pain_point.Valid(*word_res->ratings)) {
70 word_res->ratings->IncreaseBandSize(pain_point.row - pain_point.col + 1);
71 }
72 if (pain_point.Valid(*word_res->ratings) &&
73 !word_res->ratings->Classified(pain_point.col, pain_point.row, getDict().WildcardID())) {
74 found_nothing = false;
75 break;
76 }
77 }
78 if (found_nothing) {
79 if (segsearch_debug_level > 0) {
80 tprintf("Pain points queue is empty\n");
81 }
82 break;
83 }
84 ProcessSegSearchPainPoint(pain_point_priority, pain_point,
85 LMPainPoints::PainPointDescription(pp_type), &pending, word_res,
86 &pain_points, blamer_bundle);
87
88 UpdateSegSearchNodes(rating_cert_scale, pain_point.col, &pending, word_res, &pain_points,
89 best_choice_bundle, blamer_bundle);
90 if (!best_choice_bundle->updated) {
91 ++num_futile_classifications;
92 }
93
94 if (segsearch_debug_level > 0) {
95 tprintf("num_futile_classifications %d\n", num_futile_classifications);
96 }
97
98 best_choice_bundle->updated = false; // reset updated
99
100 // See if it's time to terminate SegSearch or time for starting a guided
101 // search for the true path to find the blame for the incorrect best_choice.
102 if (SegSearchDone(num_futile_classifications) && blamer_bundle != nullptr &&
103 blamer_bundle->GuidedSegsearchNeeded(word_res->best_choice)) {
104 InitBlamerForSegSearch(word_res, &pain_points, blamer_bundle, blamer_debug);
105 }
106 } // end while loop exploring alternative paths
107 if (blamer_bundle != nullptr) {
108 blamer_bundle->FinishSegSearch(word_res->best_choice, wordrec_debug_blamer, blamer_debug);
109 }
110
111 if (segsearch_debug_level > 0) {
112 tprintf("Done with SegSearch (AcceptableChoiceFound: %d)\n",
113 language_model_->AcceptableChoiceFound());
114 }
115 }
116
117 // Setup and run just the initial segsearch on an established matrix,
118 // without doing any additional chopping or joining.
119 // (Internal factored version that can be used as part of the main SegSearch.)
120 void Wordrec::InitialSegSearch(WERD_RES *word_res, LMPainPoints *pain_points,
121 std::vector<SegSearchPending> *pending,
122 BestChoiceBundle *best_choice_bundle, BlamerBundle *blamer_bundle) {
123 if (segsearch_debug_level > 0) {
124 tprintf("Starting SegSearch on ratings matrix%s:\n",
125 wordrec_enable_assoc ? " (with assoc)" : "");
126 word_res->ratings->print(getDict().getUnicharset());
127 }
128
129 pain_points->GenerateInitial(word_res);
130
131 // Compute scaling factor that will help us recover blob outline length
132 // from classifier rating and certainty for the blob.
133 float rating_cert_scale = -1.0 * getDict().certainty_scale / rating_scale;
134
135 language_model_->InitForWord(prev_word_best_choice_, assume_fixed_pitch_char_segment,
136 segsearch_max_char_wh_ratio, rating_cert_scale);
137
138 // Initialize blamer-related information: map character boxes recorded in
139 // blamer_bundle->norm_truth_word to the corresponding i,j indices in the
140 // ratings matrix. We expect this step to succeed, since when running the
141 // chopper we checked that the correct chops are present.
142 if (blamer_bundle != nullptr) {
143 blamer_bundle->SetupCorrectSegmentation(word_res->chopped_word, wordrec_debug_blamer);
144 }
145
146 // pending[col] tells whether there is update work to do to combine
147 // best_choice_bundle->beam[col - 1] with some BLOB_CHOICEs in matrix[col, *].
148 // As the language model state is updated, pending entries are modified to
149 // minimize duplication of work. It is important that during the update the
150 // children are considered in the non-decreasing order of their column, since
151 // this guarantees that all the parents would be up to date before an update
152 // of a child is done.
153 pending->clear();
154 pending->resize(word_res->ratings->dimension(), SegSearchPending());
155
156 // Search the ratings matrix for the initial best path.
157 (*pending)[0].SetColumnClassified();
158 UpdateSegSearchNodes(rating_cert_scale, 0, pending, word_res, pain_points, best_choice_bundle,
159 blamer_bundle);
160 }
161
162 void Wordrec::UpdateSegSearchNodes(float rating_cert_scale, int starting_col,
163 std::vector<SegSearchPending> *pending, WERD_RES *word_res,
164 LMPainPoints *pain_points, BestChoiceBundle *best_choice_bundle,
165 BlamerBundle *blamer_bundle) {
166 MATRIX *ratings = word_res->ratings;
167 ASSERT_HOST(static_cast<unsigned>(ratings->dimension()) == pending->size());
168 ASSERT_HOST(static_cast<unsigned>(ratings->dimension()) == best_choice_bundle->beam.size());
169 for (int col = starting_col; col < ratings->dimension(); ++col) {
170 if (!(*pending)[col].WorkToDo()) {
171 continue;
172 }
173 int first_row = col;
174 int last_row = std::min(ratings->dimension() - 1, col + ratings->bandwidth() - 1);
175 if ((*pending)[col].SingleRow() >= 0) {
176 first_row = last_row = (*pending)[col].SingleRow();
177 }
178 if (segsearch_debug_level > 0) {
179 tprintf("\n\nUpdateSegSearchNodes: col=%d, rows=[%d,%d], alljust=%d\n", col, first_row,
180 last_row, (*pending)[col].IsRowJustClassified(INT32_MAX));
181 }
182 // Iterate over the pending list for this column.
183 for (int row = first_row; row <= last_row; ++row) {
184 // Update language model state of this child+parent pair.
185 BLOB_CHOICE_LIST *current_node = ratings->get(col, row);
186 LanguageModelState *parent_node = col == 0 ? nullptr : best_choice_bundle->beam[col - 1];
187 if (current_node != nullptr &&
188 language_model_->UpdateState((*pending)[col].IsRowJustClassified(row), col, row,
189 current_node, parent_node, pain_points, word_res,
190 best_choice_bundle, blamer_bundle) &&
191 row + 1 < ratings->dimension()) {
192 // Since the language model state of this entry changed, process all
193 // the child column.
194 (*pending)[row + 1].RevisitWholeColumn();
195 if (segsearch_debug_level > 0) {
196 tprintf("Added child col=%d to pending\n", row + 1);
197 }
198 } // end if UpdateState.
199 } // end for row.
200 } // end for col.
201 if (best_choice_bundle->best_vse != nullptr) {
202 ASSERT_HOST(word_res->StatesAllValid());
203 if (best_choice_bundle->best_vse->updated) {
204 pain_points->GenerateFromPath(rating_cert_scale, best_choice_bundle->best_vse, word_res);
205 if (!best_choice_bundle->fixpt.empty()) {
206 pain_points->GenerateFromAmbigs(best_choice_bundle->fixpt, best_choice_bundle->best_vse,
207 word_res);
208 }
209 }
210 }
211 // The segsearch is completed. Reset all updated flags on all VSEs and reset
212 // all pendings.
213 for (unsigned col = 0; col < pending->size(); ++col) {
214 (*pending)[col].Clear();
215 ViterbiStateEntry_IT vse_it(&best_choice_bundle->beam[col]->viterbi_state_entries);
216 for (vse_it.mark_cycle_pt(); !vse_it.cycled_list(); vse_it.forward()) {
217 vse_it.data()->updated = false;
218 }
219 }
220 }
221
222 void Wordrec::ProcessSegSearchPainPoint(float pain_point_priority, const MATRIX_COORD &pain_point,
223 const char *pain_point_type,
224 std::vector<SegSearchPending> *pending,
225 WERD_RES *word_res, LMPainPoints *pain_points,
226 BlamerBundle *blamer_bundle) {
227 if (segsearch_debug_level > 0) {
228 tprintf("Classifying pain point %s priority=%.4f, col=%d, row=%d\n", pain_point_type,
229 pain_point_priority, pain_point.col, pain_point.row);
230 }
231 ASSERT_HOST(pain_points != nullptr);
232 MATRIX *ratings = word_res->ratings;
233 // Classify blob [pain_point.col pain_point.row]
234 if (!pain_point.Valid(*ratings)) {
235 ratings->IncreaseBandSize(pain_point.row + 1 - pain_point.col);
236 }
237 ASSERT_HOST(pain_point.Valid(*ratings));
238 BLOB_CHOICE_LIST *classified =
239 classify_piece(word_res->seam_array, pain_point.col, pain_point.row, pain_point_type,
240 word_res->chopped_word, blamer_bundle);
241 BLOB_CHOICE_LIST *lst = ratings->get(pain_point.col, pain_point.row);
242 if (lst == nullptr) {
243 ratings->put(pain_point.col, pain_point.row, classified);
244 } else {
245 // We cannot delete old BLOB_CHOICEs, since they might contain
246 // ViterbiStateEntries that are parents of other "active" entries.
247 // Thus if the matrix cell already contains classifications we add
248 // the new ones to the beginning of the list.
249 BLOB_CHOICE_IT it(lst);
250 it.add_list_before(classified);
251 delete classified; // safe to delete, since empty after add_list_before()
252 classified = nullptr;
253 }
254
255 if (segsearch_debug_level > 0) {
256 print_ratings_list("Updated ratings matrix with a new entry:",
257 ratings->get(pain_point.col, pain_point.row), getDict().getUnicharset());
258 ratings->print(getDict().getUnicharset());
259 }
260
261 // Insert initial "pain points" to join the newly classified blob
262 // with its left and right neighbors.
263 if (classified != nullptr && !classified->empty()) {
264 if (pain_point.col > 0) {
265 pain_points->GeneratePainPoint(pain_point.col - 1, pain_point.row, LM_PPTYPE_SHAPE, 0.0, true,
266 segsearch_max_char_wh_ratio, word_res);
267 }
268 if (pain_point.row + 1 < ratings->dimension()) {
269 pain_points->GeneratePainPoint(pain_point.col, pain_point.row + 1, LM_PPTYPE_SHAPE, 0.0, true,
270 segsearch_max_char_wh_ratio, word_res);
271 }
272 }
273 (*pending)[pain_point.col].SetBlobClassified(pain_point.row);
274 }
275
276 // Resets enough of the results so that the Viterbi search is re-run.
277 // Needed when the n-gram model is enabled, as the multi-length comparison
278 // implementation will re-value existing paths to worse values.
279 void Wordrec::ResetNGramSearch(WERD_RES *word_res, BestChoiceBundle *best_choice_bundle,
280 std::vector<SegSearchPending> &pending) {
281 // TODO(rays) More refactoring required here.
282 // Delete existing viterbi states.
283 for (auto &col : best_choice_bundle->beam) {
284 col->Clear();
285 }
286 // Reset best_choice_bundle.
287 word_res->ClearWordChoices();
288 best_choice_bundle->best_vse = nullptr;
289 // Clear out all existing pendings and add a new one for the first column.
290 pending[0].SetColumnClassified();
291 for (auto &data : pending) {
292 data.Clear();
293 }
294 }
295
296 void Wordrec::InitBlamerForSegSearch(WERD_RES *word_res, LMPainPoints *pain_points,
297 BlamerBundle *blamer_bundle, std::string &blamer_debug) {
298 pain_points->Clear(); // Clear pain points heap.
299 blamer_bundle->InitForSegSearch(word_res->best_choice, word_res->ratings, getDict().WildcardID(),
300 wordrec_debug_blamer, blamer_debug, pain_points,
301 segsearch_max_char_wh_ratio, word_res);
302 }
303
304 } // namespace tesseract