Mercurial > hgrepos > Python2 > PyMuPDF

comparison mupdf-source/thirdparty/tesseract/src/ccmain/paragraphs_internal.h @ 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: paragraphs_internal.h
3 * Description: Paragraph Detection internal data structures.
4 * Author: David Eger
5 *
6 * (C) Copyright 2011, 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 #ifndef TESSERACT_CCMAIN_PARAGRAPHS_INTERNAL_H_
20 #define TESSERACT_CCMAIN_PARAGRAPHS_INTERNAL_H_
21
22 #include <tesseract/publictypes.h> // for ParagraphJustification
23 #include "paragraphs.h"
24
25 // NO CODE OUTSIDE OF paragraphs.cpp AND TESTS SHOULD NEED TO ACCESS
26 // DATA STRUCTURES OR FUNCTIONS IN THIS FILE.
27
28 namespace tesseract {
29
30 class UNICHARSET;
31 class WERD_CHOICE;
32
33 // Return whether the given word is likely to be a list item start word.
34 TESS_API
35 bool AsciiLikelyListItem(const std::string &word);
36
37 // Set right word attributes given either a unicharset and werd or a utf8
38 // string.
39 TESS_API
40 void RightWordAttributes(const UNICHARSET *unicharset, const WERD_CHOICE *werd, const std::string &utf8,
41 bool *is_list, bool *starts_idea, bool *ends_idea);
42
43 // Set left word attributes given either a unicharset and werd or a utf8 string.
44 TESS_API
45 void LeftWordAttributes(const UNICHARSET *unicharset, const WERD_CHOICE *werd, const std::string &utf8,
46 bool *is_list, bool *starts_idea, bool *ends_idea);
47
48 enum LineType {
49 LT_START = 'S', // First line of a paragraph.
50 LT_BODY = 'C', // Continuation line of a paragraph.
51 LT_UNKNOWN = 'U', // No clues.
52 LT_MULTIPLE = 'M', // Matches for both LT_START and LT_BODY.
53 };
54
55 // The first paragraph in a page of body text is often un-indented.
56 // This is a typographic convention which is common to indicate either that:
57 // (1) The paragraph is the continuation of a previous paragraph, or
58 // (2) The paragraph is the first paragraph in a chapter.
59 //
60 // I refer to such paragraphs as "crown"s, and the output of the paragraph
61 // detection algorithm attempts to give them the same paragraph model as
62 // the rest of the body text.
63 //
64 // Nonetheless, while building hypotheses, it is useful to mark the lines
65 // of crown paragraphs temporarily as crowns, either aligned left or right.
66 extern const ParagraphModel *kCrownLeft;
67 extern const ParagraphModel *kCrownRight;
68
69 inline bool StrongModel(const ParagraphModel *model) {
70 return model != nullptr && model != kCrownLeft && model != kCrownRight;
71 }
72
73 struct LineHypothesis {
74 LineHypothesis() : ty(LT_UNKNOWN), model(nullptr) {}
75 LineHypothesis(LineType line_type, const ParagraphModel *m) : ty(line_type), model(m) {}
76 LineHypothesis(const LineHypothesis &other) = default;
77
78 // Copy assignment operator.
79 LineHypothesis &operator=(const LineHypothesis &other) = default;
80
81 bool operator==(const LineHypothesis &other) const {
82 return ty == other.ty && model == other.model;
83 }
84
85 LineType ty;
86 const ParagraphModel *model;
87 };
88
89 class ParagraphTheory; // Forward Declaration
90
91 using SetOfModels = std::vector<const ParagraphModel *>;
92
93 // Row Scratch Registers are data generated by the paragraph detection
94 // algorithm based on a RowInfo input.
95 class RowScratchRegisters {
96 public:
97 // We presume row will outlive us.
98 void Init(const RowInfo &row);
99
100 LineType GetLineType() const;
101
102 LineType GetLineType(const ParagraphModel *model) const;
103
104 // Mark this as a start line type, sans model. This is useful for the
105 // initial marking of probable body lines or paragraph start lines.
106 void SetStartLine();
107
108 // Mark this as a body line type, sans model. This is useful for the
109 // initial marking of probably body lines or paragraph start lines.
110 void SetBodyLine();
111
112 // Record that this row fits as a paragraph start line in the given model,
113 void AddStartLine(const ParagraphModel *model);
114 // Record that this row fits as a paragraph body line in the given model,
115 void AddBodyLine(const ParagraphModel *model);
116
117 // Clear all hypotheses about this line.
118 void SetUnknown() {
119 hypotheses_.clear();
120 }
121
122 // Append all hypotheses of strong models that match this row as a start.
123 void StartHypotheses(SetOfModels *models) const;
124
125 // Append all hypotheses of strong models matching this row.
126 void StrongHypotheses(SetOfModels *models) const;
127
128 // Append all hypotheses for this row.
129 void NonNullHypotheses(SetOfModels *models) const;
130
131 // Discard any hypotheses whose model is not in the given list.
132 void DiscardNonMatchingHypotheses(const SetOfModels &models);
133
134 // If we have only one hypothesis and that is that this line is a paragraph
135 // start line of a certain model, return that model. Else return nullptr.
136 const ParagraphModel *UniqueStartHypothesis() const;
137
138 // If we have only one hypothesis and that is that this line is a paragraph
139 // body line of a certain model, return that model. Else return nullptr.
140 const ParagraphModel *UniqueBodyHypothesis() const;
141
142 // Return the indentation for the side opposite of the aligned side.
143 int OffsideIndent(tesseract::ParagraphJustification just) const {
144 switch (just) {
145 case tesseract::JUSTIFICATION_RIGHT:
146 return lindent_;
147 case tesseract::JUSTIFICATION_LEFT:
148 return rindent_;
149 default:
150 return lindent_ > rindent_ ? lindent_ : rindent_;
151 }
152 }
153
154 // Return the indentation for the side the text is aligned to.
155 int AlignsideIndent(tesseract::ParagraphJustification just) const {
156 switch (just) {
157 case tesseract::JUSTIFICATION_RIGHT:
158 return rindent_;
159 case tesseract::JUSTIFICATION_LEFT:
160 return lindent_;
161 default:
162 return lindent_ > rindent_ ? lindent_ : rindent_;
163 }
164 }
165
166 // Append header fields to a vector of row headings.
167 static void AppendDebugHeaderFields(std::vector<std::string> &header);
168
169 // Append data for this row to a vector of debug strings.
170 void AppendDebugInfo(const ParagraphTheory &theory, std::vector<std::string> &dbg) const;
171
172 const RowInfo *ri_;
173
174 // These four constants form a horizontal box model for the white space
175 // on the edges of each line. At each point in the algorithm, the following
176 // shall hold:
177 // ri_->pix_ldistance = lmargin_ + lindent_
178 // ri_->pix_rdistance = rindent_ + rmargin_
179 int lmargin_;
180 int lindent_;
181 int rindent_;
182 int rmargin_;
183
184 private:
185 // Hypotheses of either LT_START or LT_BODY
186 std::vector<LineHypothesis> hypotheses_;
187 };
188
189 // A collection of convenience functions for wrapping the set of
190 // Paragraph Models we believe correctly model the paragraphs in the image.
191 class ParagraphTheory {
192 public:
193 // We presume models will outlive us, and that models will take ownership
194 // of any ParagraphModel *'s we add.
195 explicit ParagraphTheory(std::vector<ParagraphModel *> *models) : models_(models) {}
196 std::vector<ParagraphModel *> &models() {
197 return *models_;
198 }
199 const std::vector<ParagraphModel *> &models() const {
200 return *models_;
201 }
202
203 // Return an existing model if one that is Comparable() can be found.
204 // Else, allocate a new copy of model to save and return a pointer to it.
205 const ParagraphModel *AddModel(const ParagraphModel &model);
206
207 // Discard any models we've made that are not in the list of used models.
208 void DiscardUnusedModels(const SetOfModels &used_models);
209
210 // Return the set of all non-centered models.
211 void NonCenteredModels(SetOfModels *models);
212
213 // If any of the non-centered paragraph models we know about fit
214 // rows[start, end), return it. Else nullptr.
215 const ParagraphModel *Fits(const std::vector<RowScratchRegisters> *rows, int start,
216 int end) const;
217
218 int IndexOf(const ParagraphModel *model) const;
219
220 private:
221 std::vector<ParagraphModel *> *models_;
222 std::vector<ParagraphModel *> models_we_added_;
223 };
224
225 bool ValidFirstLine(const std::vector<RowScratchRegisters> *rows, int row,
226 const ParagraphModel *model);
227 bool ValidBodyLine(const std::vector<RowScratchRegisters> *rows, int row,
228 const ParagraphModel *model);
229 bool CrownCompatible(const std::vector<RowScratchRegisters> *rows, int a, int b,
230 const ParagraphModel *model);
231
232 // A class for smearing Paragraph Model hypotheses to surrounding rows.
233 // The idea here is that StrongEvidenceClassify first marks only exceedingly
234 // obvious start and body rows and constructs models of them. Thereafter,
235 // we may have left over unmarked lines (mostly end-of-paragraph lines) which
236 // were too short to have much confidence about, but which fit the models we've
237 // constructed perfectly and which we ought to mark. This class is used to
238 // "smear" our models over the text.
239 class ParagraphModelSmearer {
240 public:
241 ParagraphModelSmearer(std::vector<RowScratchRegisters> *rows, int row_start, int row_end,
242 ParagraphTheory *theory);
243
244 // Smear forward paragraph models from existing row markings to subsequent
245 // text lines if they fit, and mark any thereafter still unmodeled rows
246 // with any model in the theory that fits them.
247 void Smear();
248
249 private:
250 // Record in open_models_ for rows [start_row, end_row) the list of models
251 // currently open at each row.
252 // A model is still open in a row if some previous row has said model as a
253 // start hypothesis, and all rows since (including this row) would fit as
254 // either a body or start line in that model.
255 void CalculateOpenModels(int row_start, int row_end);
256
257 SetOfModels &OpenModels(int row) {
258 return open_models_[row - row_start_ + 1];
259 }
260
261 ParagraphTheory *theory_;
262 std::vector<RowScratchRegisters> *rows_;
263 int row_start_;
264 int row_end_;
265
266 // open_models_ corresponds to rows[start_row_ - 1, end_row_]
267 //
268 // open_models_: Contains models which there was an active (open) paragraph
269 // as of the previous line and for which the left and right
270 // indents admit the possibility that this text line continues
271 // to fit the same model.
272 // TODO(eger): Think about whether we can get rid of "Open" models and just
273 // use the current hypotheses on RowScratchRegisters.
274 std::vector<SetOfModels> open_models_;
275 };
276
277 // Clear all hypotheses about lines [start, end) and reset the margins to the
278 // percentile (0..100) value of the left and right row edges for this run of
279 // rows.
280 void RecomputeMarginsAndClearHypotheses(std::vector<RowScratchRegisters> *rows, int start,
281 int end, int percentile);
282
283 // Return the median inter-word space in rows[row_start, row_end).
284 int InterwordSpace(const std::vector<RowScratchRegisters> &rows, int row_start, int row_end);
285
286 // Return whether the first word on the after line can fit in the space at
287 // the end of the before line (knowing which way the text is aligned and read).
288 bool FirstWordWouldHaveFit(const RowScratchRegisters &before, const RowScratchRegisters &after,
289 tesseract::ParagraphJustification justification);
290
291 // Return whether the first word on the after line can fit in the space at
292 // the end of the before line (not knowing the text alignment).
293 bool FirstWordWouldHaveFit(const RowScratchRegisters &before, const RowScratchRegisters &after);
294
295 // Do rows[start, end) form a single instance of the given paragraph model?
296 bool RowsFitModel(const std::vector<RowScratchRegisters> *rows, int start, int end,
297 const ParagraphModel *model);
298
299 // Given a set of row_owners pointing to PARAs or nullptr (no paragraph known),
300 // normalize each row_owner to point to an actual PARA, and output the
301 // paragraphs in order onto paragraphs.
302 void CanonicalizeDetectionResults(std::vector<PARA *> *row_owners, PARA_LIST *paragraphs);
303
304 } // namespace tesseract
305
306 #endif // TESSERACT_CCMAIN_PARAGRAPHS_INTERNAL_H_