Mercurial > hgrepos > Python2 > PyMuPDF

comparison mupdf-source/thirdparty/tesseract/src/ccmain/linerec.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: linerec.cpp
3 // Description: Top-level line-based recognition module for Tesseract.
4 // Author: Ray Smith
5 //
6 // (C) Copyright 2013, 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 #include "tesseractclass.h"
19
20 #include <allheaders.h>
21 #include "boxread.h"
22 #include "imagedata.h" // for ImageData
23 #include "lstmrecognizer.h"
24 #include "pageres.h"
25 #include "recodebeam.h"
26 #include "tprintf.h"
27
28 #include <algorithm>
29
30 namespace tesseract {
31
32 // Scale factor to make certainty more comparable to Tesseract.
33 const float kCertaintyScale = 7.0f;
34 // Worst acceptable certainty for a dictionary word.
35 const float kWorstDictCertainty = -25.0f;
36
37 // Generates training data for training a line recognizer, eg LSTM.
38 // Breaks the page into lines, according to the boxes, and writes them to a
39 // serialized DocumentData based on output_basename.
40 // Return true if successful, false if an error occurred.
41 bool Tesseract::TrainLineRecognizer(const char *input_imagename, const std::string &output_basename,
42 BLOCK_LIST *block_list) {
43 std::string lstmf_name = output_basename + ".lstmf";
44 DocumentData images(lstmf_name);
45 if (applybox_page > 0) {
46 // Load existing document for the previous pages.
47 if (!images.LoadDocument(lstmf_name.c_str(), 0, 0, nullptr)) {
48 tprintf("Failed to read training data from %s!\n", lstmf_name.c_str());
49 return false;
50 }
51 }
52 std::vector<TBOX> boxes;
53 std::vector<std::string> texts;
54 // Get the boxes for this page, if there are any.
55 if (!ReadAllBoxes(applybox_page, false, input_imagename, &boxes, &texts, nullptr, nullptr) ||
56 boxes.empty()) {
57 tprintf("Failed to read boxes from %s\n", input_imagename);
58 return false;
59 }
60 TrainFromBoxes(boxes, texts, block_list, &images);
61 if (images.PagesSize() == 0) {
62 tprintf("Failed to read pages from %s\n", input_imagename);
63 return false;
64 }
65 images.Shuffle();
66 if (!images.SaveDocument(lstmf_name.c_str(), nullptr)) {
67 tprintf("Failed to write training data to %s!\n", lstmf_name.c_str());
68 return false;
69 }
70 return true;
71 }
72
73 // Generates training data for training a line recognizer, eg LSTM.
74 // Breaks the boxes into lines, normalizes them, converts to ImageData and
75 // appends them to the given training_data.
76 void Tesseract::TrainFromBoxes(const std::vector<TBOX> &boxes, const std::vector<std::string> &texts,
77 BLOCK_LIST *block_list, DocumentData *training_data) {
78 auto box_count = boxes.size();
79 // Process all the text lines in this page, as defined by the boxes.
80 unsigned end_box = 0;
81 // Don't let \t, which marks newlines in the box file, get into the line
82 // content, as that makes the line unusable in training.
83 while (end_box < texts.size() && texts[end_box] == "\t") {
84 ++end_box;
85 }
86 for (auto start_box = end_box; start_box < box_count; start_box = end_box) {
87 // Find the textline of boxes starting at start and their bounding box.
88 TBOX line_box = boxes[start_box];
89 std::string line_str = texts[start_box];
90 for (end_box = start_box + 1; end_box < box_count && texts[end_box] != "\t"; ++end_box) {
91 line_box += boxes[end_box];
92 line_str += texts[end_box];
93 }
94 // Find the most overlapping block.
95 BLOCK *best_block = nullptr;
96 int best_overlap = 0;
97 BLOCK_IT b_it(block_list);
98 for (b_it.mark_cycle_pt(); !b_it.cycled_list(); b_it.forward()) {
99 BLOCK *block = b_it.data();
100 if (block->pdblk.poly_block() != nullptr && !block->pdblk.poly_block()->IsText()) {
101 continue; // Not a text block.
102 }
103 TBOX block_box = block->pdblk.bounding_box();
104 block_box.rotate(block->re_rotation());
105 if (block_box.major_overlap(line_box)) {
106 TBOX overlap_box = line_box.intersection(block_box);
107 if (overlap_box.area() > best_overlap) {
108 best_overlap = overlap_box.area();
109 best_block = block;
110 }
111 }
112 }
113 ImageData *imagedata = nullptr;
114 if (best_block == nullptr) {
115 tprintf("No block overlapping textline: %s\n", line_str.c_str());
116 } else {
117 imagedata = GetLineData(line_box, boxes, texts, start_box, end_box, *best_block);
118 }
119 if (imagedata != nullptr) {
120 training_data->AddPageToDocument(imagedata);
121 }
122 // Don't let \t, which marks newlines in the box file, get into the line
123 // content, as that makes the line unusable in training.
124 while (end_box < texts.size() && texts[end_box] == "\t") {
125 ++end_box;
126 }
127 }
128 }
129
130 // Returns an Imagedata containing the image of the given box,
131 // and ground truth boxes/truth text if available in the input.
132 // The image is not normalized in any way.
133 ImageData *Tesseract::GetLineData(const TBOX &line_box, const std::vector<TBOX> &boxes,
134 const std::vector<std::string> &texts, int start_box, int end_box,
135 const BLOCK &block) {
136 TBOX revised_box;
137 ImageData *image_data = GetRectImage(line_box, block, kImagePadding, &revised_box);
138 if (image_data == nullptr) {
139 return nullptr;
140 }
141 image_data->set_page_number(applybox_page);
142 // Copy the boxes and shift them so they are relative to the image.
143 FCOORD block_rotation(block.re_rotation().x(), -block.re_rotation().y());
144 ICOORD shift = -revised_box.botleft();
145 std::vector<TBOX> line_boxes;
146 std::vector<std::string> line_texts;
147 for (int b = start_box; b < end_box; ++b) {
148 TBOX box = boxes[b];
149 box.rotate(block_rotation);
150 box.move(shift);
151 line_boxes.push_back(box);
152 line_texts.push_back(texts[b]);
153 }
154 std::vector<int> page_numbers(line_boxes.size(), applybox_page);
155 image_data->AddBoxes(line_boxes, line_texts, page_numbers);
156 return image_data;
157 }
158
159 // Helper gets the image of a rectangle, using the block.re_rotation() if
160 // needed to get to the image, and rotating the result back to horizontal
161 // layout. (CJK characters will be on their left sides) The vertical text flag
162 // is set in the returned ImageData if the text was originally vertical, which
163 // can be used to invoke a different CJK recognition engine. The revised_box
164 // is also returned to enable calculation of output bounding boxes.
165 ImageData *Tesseract::GetRectImage(const TBOX &box, const BLOCK &block, int padding,
166 TBOX *revised_box) const {
167 TBOX wbox = box;
168 wbox.pad(padding, padding);
169 *revised_box = wbox;
170 // Number of clockwise 90 degree rotations needed to get back to tesseract
171 // coords from the clipped image.
172 int num_rotations = 0;
173 if (block.re_rotation().y() > 0.0f) {
174 num_rotations = 1;
175 } else if (block.re_rotation().x() < 0.0f) {
176 num_rotations = 2;
177 } else if (block.re_rotation().y() < 0.0f) {
178 num_rotations = 3;
179 }
180 // Handle two cases automatically: 1 the box came from the block, 2 the box
181 // came from a box file, and refers to the image, which the block may not.
182 if (block.pdblk.bounding_box().major_overlap(*revised_box)) {
183 revised_box->rotate(block.re_rotation());
184 }
185 // Now revised_box always refers to the image.
186 // BestPix is never colormapped, but may be of any depth.
187 Image pix = BestPix();
188 int width = pixGetWidth(pix);
189 int height = pixGetHeight(pix);
190 TBOX image_box(0, 0, width, height);
191 // Clip to image bounds;
192 *revised_box &= image_box;
193 if (revised_box->null_box()) {
194 return nullptr;
195 }
196 Box *clip_box = boxCreate(revised_box->left(), height - revised_box->top(), revised_box->width(),
197 revised_box->height());
198 Image box_pix = pixClipRectangle(pix, clip_box, nullptr);
199 boxDestroy(&clip_box);
200 if (box_pix == nullptr) {
201 return nullptr;
202 }
203 if (num_rotations > 0) {
204 Image rot_pix = pixRotateOrth(box_pix, num_rotations);
205 box_pix.destroy();
206 box_pix = rot_pix;
207 }
208 // Convert sub-8-bit images to 8 bit.
209 int depth = pixGetDepth(box_pix);
210 if (depth < 8) {
211 Image grey;
212 grey = pixConvertTo8(box_pix, false);
213 box_pix.destroy();
214 box_pix = grey;
215 }
216 bool vertical_text = false;
217 if (num_rotations > 0) {
218 // Rotated the clipped revised box back to internal coordinates.
219 FCOORD rotation(block.re_rotation().x(), -block.re_rotation().y());
220 revised_box->rotate(rotation);
221 if (num_rotations != 2) {
222 vertical_text = true;
223 }
224 }
225 return new ImageData(vertical_text, box_pix);
226 }
227
228 // Recognizes a word or group of words, converting to WERD_RES in *words.
229 // Analogous to classify_word_pass1, but can handle a group of words as well.
230 void Tesseract::LSTMRecognizeWord(const BLOCK &block, ROW *row, WERD_RES *word,
231 PointerVector<WERD_RES> *words) {
232 TBOX word_box = word->word->bounding_box();
233 // Get the word image - no frills.
234 if (tessedit_pageseg_mode == PSM_SINGLE_WORD || tessedit_pageseg_mode == PSM_RAW_LINE) {
235 // In single word mode, use the whole image without any other row/word
236 // interpretation.
237 word_box = TBOX(0, 0, ImageWidth(), ImageHeight());
238 } else {
239 float baseline = row->base_line((word_box.left() + word_box.right()) / 2);
240 if (baseline + row->descenders() < word_box.bottom()) {
241 word_box.set_bottom(baseline + row->descenders());
242 }
243 if (baseline + row->x_height() + row->ascenders() > word_box.top()) {
244 word_box.set_top(baseline + row->x_height() + row->ascenders());
245 }
246 }
247 ImageData *im_data = GetRectImage(word_box, block, kImagePadding, &word_box);
248 if (im_data == nullptr) {
249 return;
250 }
251
252 bool do_invert = tessedit_do_invert;
253 float threshold = do_invert ? double(invert_threshold) : 0.0f;
254 lstm_recognizer_->RecognizeLine(*im_data, threshold, classify_debug_level > 0,
255 kWorstDictCertainty / kCertaintyScale, word_box, words,
256 lstm_choice_mode, lstm_choice_iterations);
257 delete im_data;
258 SearchWords(words);
259 }
260
261 // Apply segmentation search to the given set of words, within the constraints
262 // of the existing ratings matrix. If there is already a best_choice on a word
263 // leaves it untouched and just sets the done/accepted etc flags.
264 void Tesseract::SearchWords(PointerVector<WERD_RES> *words) {
265 // Run the segmentation search on the network outputs and make a BoxWord
266 // for each of the output words.
267 // If we drop a word as junk, then there is always a space in front of the
268 // next.
269 const Dict *stopper_dict = lstm_recognizer_->GetDict();
270 if (stopper_dict == nullptr) {
271 stopper_dict = &getDict();
272 }
273 for (unsigned w = 0; w < words->size(); ++w) {
274 WERD_RES *word = (*words)[w];
275 if (word->best_choice == nullptr) {
276 // It is a dud.
277 word->SetupFake(lstm_recognizer_->GetUnicharset());
278 } else {
279 // Set the best state.
280 for (unsigned i = 0; i < word->best_choice->length(); ++i) {
281 int length = word->best_choice->state(i);
282 word->best_state.push_back(length);
283 }
284 word->reject_map.initialise(word->best_choice->length());
285 word->tess_failed = false;
286 word->tess_accepted = true;
287 word->tess_would_adapt = false;
288 word->done = true;
289 word->tesseract = this;
290 float word_certainty = std::min(word->space_certainty, word->best_choice->certainty());
291 word_certainty *= kCertaintyScale;
292 if (getDict().stopper_debug_level >= 1) {
293 tprintf("Best choice certainty=%g, space=%g, scaled=%g, final=%g\n",
294 word->best_choice->certainty(), word->space_certainty,
295 std::min(word->space_certainty, word->best_choice->certainty()) * kCertaintyScale,
296 word_certainty);
297 word->best_choice->print();
298 }
299 word->best_choice->set_certainty(word_certainty);
300
301 word->tess_accepted = stopper_dict->AcceptableResult(word);
302 }
303 }
304 }
305
306 } // namespace tesseract.