1 ///////////////////////////////////////////////////////////////////////
2 // File: strokewidth.cpp
3 // Description: Subclass of BBGrid to find uniformity of strokewidth.
4 // Author: Ray Smith
5 //
6 // (C) Copyright 2008, 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 #ifdef HAVE_CONFIG_H
20 # include "config_auto.h"
21 #endif
22
23 #include "strokewidth.h"
24
25 #include <algorithm>
26 #include <cmath>
27
28 #include "blobbox.h"
29 #include "colpartition.h"
30 #include "colpartitiongrid.h"
31 #include "helpers.h" // for IntCastRounded
32 #include "imagefind.h"
33 #include "linlsq.h"
34 #include "statistc.h"
35 #include "tabfind.h"
36 #include "textlineprojection.h"
37 #include "tordmain.h" // For SetBlobStrokeWidth.
38
39 namespace tesseract {
40
41 #ifndef GRAPHICS_DISABLED
42 static INT_VAR(textord_tabfind_show_strokewidths, 0, "Show stroke widths (ScrollView)");
43 #else
44 static INT_VAR(textord_tabfind_show_strokewidths, 0, "Show stroke widths");
45 #endif
46 static BOOL_VAR(textord_tabfind_only_strokewidths, false, "Only run stroke widths");
47
48 /** Allowed proportional change in stroke width to be the same font. */
49 const double kStrokeWidthFractionTolerance = 0.125;
50 /**
51 * Allowed constant change in stroke width to be the same font.
52 * Really 1.5 pixels.
53 */
54 const double kStrokeWidthTolerance = 1.5;
55 // Same but for CJK we are a bit more generous.
56 const double kStrokeWidthFractionCJK = 0.25;
57 const double kStrokeWidthCJK = 2.0;
58 // Radius in grid cells of search for broken CJK. Doesn't need to be very
59 // large as the grid size should be about the size of a character anyway.
60 const int kCJKRadius = 2;
61 // Max distance fraction of size to join close but broken CJK characters.
62 const double kCJKBrokenDistanceFraction = 0.25;
63 // Max number of components in a broken CJK character.
64 const int kCJKMaxComponents = 8;
65 // Max aspect ratio of CJK broken characters when put back together.
66 const double kCJKAspectRatio = 1.25;
67 // Max increase in aspect ratio of CJK broken characters when merged.
68 const double kCJKAspectRatioIncrease = 1.0625;
69 // Max multiple of the grid size that will be used in computing median CJKsize.
70 const int kMaxCJKSizeRatio = 5;
71 // Min fraction of blobs broken CJK to iterate and run it again.
72 const double kBrokenCJKIterationFraction = 0.125;
73 // Multiple of gridsize as x-padding for a search box for diacritic base
74 // characters.
75 const double kDiacriticXPadRatio = 7.0;
76 // Multiple of gridsize as y-padding for a search box for diacritic base
77 // characters.
78 const double kDiacriticYPadRatio = 1.75;
79 // Min multiple of diacritic height that a neighbour must be to be a
80 // convincing base character.
81 const double kMinDiacriticSizeRatio = 1.0625;
82 // Max multiple of a textline's median height as a threshold for the sum of
83 // a diacritic's farthest x and y distances (gap + size).
84 const double kMaxDiacriticDistanceRatio = 1.25;
85 // Max x-gap between a diacritic and its base char as a fraction of the height
86 // of the base char (allowing other blobs to fill the gap.)
87 const double kMaxDiacriticGapToBaseCharHeight = 1.0;
88 // Ratio between longest side of a line and longest side of a character.
89 // (neighbor_min > blob_min * kLineTrapShortest &&
90 // neighbor_max < blob_max / kLineTrapLongest)
91 // => neighbor is a grapheme and blob is a line.
92 const int kLineTrapLongest = 4;
93 // Ratio between shortest side of a line and shortest side of a character.
94 const int kLineTrapShortest = 2;
95 // Max aspect ratio of the total box before CountNeighbourGaps
96 // decides immediately based on the aspect ratio.
97 const int kMostlyOneDirRatio = 3;
98 // Aspect ratio for a blob to be considered as line residue.
99 const double kLineResidueAspectRatio = 8.0;
100 // Padding ratio for line residue search box.
101 const int kLineResiduePadRatio = 3;
102 // Min multiple of neighbour size for a line residue to be genuine.
103 const double kLineResidueSizeRatio = 1.75;
104 // Aspect ratio filter for OSD.
105 const float kSizeRatioToReject = 2.0;
106 // Expansion factor for search box for good neighbours.
107 const double kNeighbourSearchFactor = 2.5;
108 // Factor of increase of overlap when adding diacritics to make an image noisy.
109 const double kNoiseOverlapGrowthFactor = 4.0;
110 // Fraction of the image size to add overlap when adding diacritics for an
111 // image to qualify as noisy.
112 const double kNoiseOverlapAreaFactor = 1.0 / 512;
113
StrokeWidth(int gridsize,const ICOORD & bleft,const ICOORD & tright)114 StrokeWidth::StrokeWidth(int gridsize, const ICOORD &bleft, const ICOORD &tright)
115 : BlobGrid(gridsize, bleft, tright)
116 , nontext_map_(nullptr)
117 , projection_(nullptr)
118 , denorm_(nullptr)
119 , grid_box_(bleft, tright)
120 , rerotation_(1.0f, 0.0f) {
121 }
122
~StrokeWidth()123 StrokeWidth::~StrokeWidth() {
124 #ifndef GRAPHICS_DISABLED
125 if (widths_win_ != nullptr) {
126 delete widths_win_->AwaitEvent(SVET_DESTROY);
127 if (textord_tabfind_only_strokewidths) {
128 exit(0);
129 }
130 delete widths_win_;
131 }
132 delete leaders_win_;
133 delete initial_widths_win_;
134 delete chains_win_;
135 delete textlines_win_;
136 delete smoothed_win_;
137 delete diacritics_win_;
138 #endif
139 }
140
141 // Sets the neighbours member of the medium-sized blobs in the block.
142 // Searches on 4 sides of each blob for similar-sized, similar-strokewidth
143 // blobs and sets pointers to the good neighbours.
SetNeighboursOnMediumBlobs(TO_BLOCK * block)144 void StrokeWidth::SetNeighboursOnMediumBlobs(TO_BLOCK *block) {
145 // Run a preliminary strokewidth neighbour detection on the medium blobs.
146 InsertBlobList(&block->blobs);
147 BLOBNBOX_IT blob_it(&block->blobs);
148 for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) {
149 SetNeighbours(false, false, blob_it.data());
150 }
151 Clear();
152 }
153
154 // Sets the neighbour/textline writing direction members of the medium
155 // and large blobs with optional repair of broken CJK characters first.
156 // Repair of broken CJK is needed here because broken CJK characters
157 // can fool the textline direction detection algorithm.
FindTextlineDirectionAndFixBrokenCJK(PageSegMode pageseg_mode,bool cjk_merge,TO_BLOCK * input_block)158 void StrokeWidth::FindTextlineDirectionAndFixBrokenCJK(PageSegMode pageseg_mode, bool cjk_merge,
159 TO_BLOCK *input_block) {
160 // Setup the grid with the remaining (non-noise) blobs.
161 InsertBlobs(input_block);
162 // Repair broken CJK characters if needed.
163 while (cjk_merge && FixBrokenCJK(input_block)) {
164 ;
165 }
166 // Grade blobs by inspection of neighbours.
167 FindTextlineFlowDirection(pageseg_mode, false);
168 // Clear the grid ready for rotation or leader finding.
169 Clear();
170 }
171
172 // Helper to collect and count horizontal and vertical blobs from a list.
CollectHorizVertBlobs(BLOBNBOX_LIST * input_blobs,int * num_vertical_blobs,int * num_horizontal_blobs,BLOBNBOX_CLIST * vertical_blobs,BLOBNBOX_CLIST * horizontal_blobs,BLOBNBOX_CLIST * nondescript_blobs)173 static void CollectHorizVertBlobs(BLOBNBOX_LIST *input_blobs, int *num_vertical_blobs,
174 int *num_horizontal_blobs, BLOBNBOX_CLIST *vertical_blobs,
175 BLOBNBOX_CLIST *horizontal_blobs,
176 BLOBNBOX_CLIST *nondescript_blobs) {
177 BLOBNBOX_C_IT v_it(vertical_blobs);
178 BLOBNBOX_C_IT h_it(horizontal_blobs);
179 BLOBNBOX_C_IT n_it(nondescript_blobs);
180 BLOBNBOX_IT blob_it(input_blobs);
181 for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) {
182 BLOBNBOX *blob = blob_it.data();
183 const TBOX &box = blob->bounding_box();
184 float y_x = static_cast<float>(box.height()) / box.width();
185 float x_y = 1.0f / y_x;
186 // Select a >= 1.0 ratio
187 float ratio = x_y > y_x ? x_y : y_x;
188 // If the aspect ratio is small and we want them for osd, save the blob.
189 bool ok_blob = ratio <= kSizeRatioToReject;
190 if (blob->UniquelyVertical()) {
191 ++*num_vertical_blobs;
192 if (ok_blob) {
193 v_it.add_after_then_move(blob);
194 }
195 } else if (blob->UniquelyHorizontal()) {
196 ++*num_horizontal_blobs;
197 if (ok_blob) {
198 h_it.add_after_then_move(blob);
199 }
200 } else if (ok_blob) {
201 n_it.add_after_then_move(blob);
202 }
203 }
204 }
205
206 // Types all the blobs as vertical or horizontal text or unknown and
207 // returns true if the majority are vertical.
208 // If the blobs are rotated, it is necessary to call CorrectForRotation
209 // after rotating everything, otherwise the work done here will be enough.
210 // If osd_blobs is not null, a list of blobs from the dominant textline
211 // direction are returned for use in orientation and script detection.
TestVerticalTextDirection(double find_vertical_text_ratio,TO_BLOCK * block,BLOBNBOX_CLIST * osd_blobs)212 bool StrokeWidth::TestVerticalTextDirection(double find_vertical_text_ratio, TO_BLOCK *block,
213 BLOBNBOX_CLIST *osd_blobs) {
214 int vertical_boxes = 0;
215 int horizontal_boxes = 0;
216 // Count vertical normal and large blobs.
217 BLOBNBOX_CLIST vertical_blobs;
218 BLOBNBOX_CLIST horizontal_blobs;
219 BLOBNBOX_CLIST nondescript_blobs;
220 CollectHorizVertBlobs(&block->blobs, &vertical_boxes, &horizontal_boxes, &vertical_blobs,
221 &horizontal_blobs, &nondescript_blobs);
222 CollectHorizVertBlobs(&block->large_blobs, &vertical_boxes, &horizontal_boxes, &vertical_blobs,
223 &horizontal_blobs, &nondescript_blobs);
224 if (textord_debug_tabfind) {
225 tprintf("TextDir hbox=%d vs vbox=%d, %dH, %dV, %dN osd blobs\n", horizontal_boxes,
226 vertical_boxes, horizontal_blobs.length(), vertical_blobs.length(),
227 nondescript_blobs.length());
228 }
229 if (osd_blobs != nullptr && vertical_boxes == 0 && horizontal_boxes == 0) {
230 // Only nondescript blobs available, so return those.
231 BLOBNBOX_C_IT osd_it(osd_blobs);
232 osd_it.add_list_after(&nondescript_blobs);
233 return false;
234 }
235 int min_vert_boxes =
236 static_cast<int>((vertical_boxes + horizontal_boxes) * find_vertical_text_ratio);
237 if (vertical_boxes >= min_vert_boxes) {
238 if (osd_blobs != nullptr) {
239 BLOBNBOX_C_IT osd_it(osd_blobs);
240 osd_it.add_list_after(&vertical_blobs);
241 }
242 return true;
243 } else {
244 if (osd_blobs != nullptr) {
245 BLOBNBOX_C_IT osd_it(osd_blobs);
246 osd_it.add_list_after(&horizontal_blobs);
247 }
248 return false;
249 }
250 }
251
252 // Corrects the data structures for the given rotation.
CorrectForRotation(const FCOORD & rotation,ColPartitionGrid * part_grid)253 void StrokeWidth::CorrectForRotation(const FCOORD &rotation, ColPartitionGrid *part_grid) {
254 Init(part_grid->gridsize(), part_grid->bleft(), part_grid->tright());
255 grid_box_ = TBOX(bleft(), tright());
256 rerotation_.set_x(rotation.x());
257 rerotation_.set_y(-rotation.y());
258 }
259
260 // Finds leader partitions and inserts them into the given part_grid.
FindLeaderPartitions(TO_BLOCK * block,ColPartitionGrid * part_grid)261 void StrokeWidth::FindLeaderPartitions(TO_BLOCK *block, ColPartitionGrid *part_grid) {
262 Clear();
263 // Find and isolate leaders in the noise list.
264 ColPartition_LIST leader_parts;
265 FindLeadersAndMarkNoise(block, &leader_parts);
266 // Setup the strokewidth grid with the block's remaining (non-noise) blobs.
267 InsertBlobList(&block->blobs);
268 // Mark blobs that have leader neighbours.
269 for (ColPartition_IT it(&leader_parts); !it.empty(); it.forward()) {
270 ColPartition *part = it.extract();
271 part->ClaimBoxes();
272 MarkLeaderNeighbours(part, LR_LEFT);
273 MarkLeaderNeighbours(part, LR_RIGHT);
274 part_grid->InsertBBox(true, true, part);
275 }
276 }
277
278 // Finds and marks noise those blobs that look like bits of vertical lines
279 // that would otherwise screw up layout analysis.
RemoveLineResidue(ColPartition_LIST * big_part_list)280 void StrokeWidth::RemoveLineResidue(ColPartition_LIST *big_part_list) {
281 BlobGridSearch gsearch(this);
282 BLOBNBOX *bbox;
283 // For every vertical line-like bbox in the grid, search its neighbours
284 // to find the tallest, and if the original box is taller by sufficient
285 // margin, then call it line residue and delete it.
286 gsearch.StartFullSearch();
287 while ((bbox = gsearch.NextFullSearch()) != nullptr) {
288 TBOX box = bbox->bounding_box();
289 if (box.height() < box.width() * kLineResidueAspectRatio) {
290 continue;
291 }
292 // Set up a rectangle search around the blob to find the size of its
293 // neighbours.
294 int padding = box.height() * kLineResiduePadRatio;
295 TBOX search_box = box;
296 search_box.pad(padding, padding);
297 bool debug = AlignedBlob::WithinTestRegion(2, box.left(), box.bottom());
298 // Find the largest object in the search box not equal to bbox.
299 BlobGridSearch rsearch(this);
300 int max_height = 0;
301 BLOBNBOX *n;
302 rsearch.StartRectSearch(search_box);
303 while ((n = rsearch.NextRectSearch()) != nullptr) {
304 if (n == bbox) {
305 continue;
306 }
307 TBOX nbox = n->bounding_box();
308 if (nbox.height() > max_height) {
309 max_height = nbox.height();
310 }
311 }
312 if (debug) {
313 tprintf("Max neighbour size=%d for candidate line box at:", max_height);
314 box.print();
315 }
316 if (max_height * kLineResidueSizeRatio < box.height()) {
317 #ifndef GRAPHICS_DISABLED
318 if (leaders_win_ != nullptr) {
319 // We are debugging, so display deleted in pink blobs in the same
320 // window that we use to display leader detection.
321 leaders_win_->Pen(ScrollView::PINK);
322 leaders_win_->Rectangle(box.left(), box.bottom(), box.right(), box.top());
323 }
324 #endif // !GRAPHICS_DISABLED
325 ColPartition::MakeBigPartition(bbox, big_part_list);
326 }
327 }
328 }
329
330 // Types all the blobs as vertical text or horizontal text or unknown and
331 // puts them into initial ColPartitions in the supplied part_grid.
332 // rerotation determines how to get back to the image coordinates from the
333 // blob coordinates (since they may have been rotated for vertical text).
334 // block is the single block for the whole page or rectangle to be OCRed.
335 // nontext_pix (full-size), is a binary mask used to prevent merges across
336 // photo/text boundaries. It is not kept beyond this function.
337 // denorm provides a mapping back to the image from the current blob
338 // coordinate space.
339 // projection provides a measure of textline density over the image and
340 // provides functions to assist with diacritic detection. It should be a
341 // pointer to a new TextlineProjection, and will be setup here.
342 // part_grid is the output grid of textline partitions.
343 // Large blobs that cause overlap are put in separate partitions and added
344 // to the big_parts list.
GradeBlobsIntoPartitions(PageSegMode pageseg_mode,const FCOORD & rerotation,TO_BLOCK * block,Image nontext_pix,const DENORM * denorm,bool cjk_script,TextlineProjection * projection,BLOBNBOX_LIST * diacritic_blobs,ColPartitionGrid * part_grid,ColPartition_LIST * big_parts)345 void StrokeWidth::GradeBlobsIntoPartitions(PageSegMode pageseg_mode, const FCOORD &rerotation,
346 TO_BLOCK *block, Image nontext_pix, const DENORM *denorm,
347 bool cjk_script, TextlineProjection *projection,
348 BLOBNBOX_LIST *diacritic_blobs,
349 ColPartitionGrid *part_grid,
350 ColPartition_LIST *big_parts) {
351 nontext_map_ = nontext_pix;
352 projection_ = projection;
353 denorm_ = denorm;
354 // Clear and re Insert to take advantage of the tab stops in the blobs.
355 Clear();
356 // Setup the strokewidth grid with the remaining non-noise, non-leader blobs.
357 InsertBlobs(block);
358
359 // Run FixBrokenCJK() again if the page is CJK.
360 if (cjk_script) {
361 FixBrokenCJK(block);
362 }
363 FindTextlineFlowDirection(pageseg_mode, false);
364 projection_->ConstructProjection(block, rerotation, nontext_map_);
365 #ifndef GRAPHICS_DISABLED
366 if (textord_tabfind_show_strokewidths) {
367 ScrollView *line_blobs_win = MakeWindow(0, 0, "Initial textline Blobs");
368 projection_->PlotGradedBlobs(&block->blobs, line_blobs_win);
369 projection_->PlotGradedBlobs(&block->small_blobs, line_blobs_win);
370 }
371 #endif
372 projection_->MoveNonTextlineBlobs(&block->blobs, &block->noise_blobs);
373 projection_->MoveNonTextlineBlobs(&block->small_blobs, &block->noise_blobs);
374 // Clear and re Insert to take advantage of the removed diacritics.
375 Clear();
376 InsertBlobs(block);
377 FCOORD skew;
378 FindTextlineFlowDirection(pageseg_mode, true);
379 PartitionFindResult r = FindInitialPartitions(pageseg_mode, rerotation, true, block,
380 diacritic_blobs, part_grid, big_parts, &skew);
381 if (r == PFR_NOISE) {
382 tprintf("Detected %d diacritics\n", diacritic_blobs->length());
383 // Noise was found, and removed.
384 Clear();
385 InsertBlobs(block);
386 FindTextlineFlowDirection(pageseg_mode, true);
387 r = FindInitialPartitions(pageseg_mode, rerotation, false, block, diacritic_blobs, part_grid,
388 big_parts, &skew);
389 }
390 nontext_map_ = nullptr;
391 projection_ = nullptr;
392 denorm_ = nullptr;
393 }
394
PrintBoxWidths(BLOBNBOX * neighbour)395 static void PrintBoxWidths(BLOBNBOX *neighbour) {
396 const TBOX &nbox = neighbour->bounding_box();
397 tprintf("Box (%d,%d)->(%d,%d): h-width=%.1f, v-width=%.1f p-width=%1.f\n", nbox.left(),
398 nbox.bottom(), nbox.right(), nbox.top(), neighbour->horz_stroke_width(),
399 neighbour->vert_stroke_width(),
400 2.0 * neighbour->cblob()->area() / neighbour->cblob()->perimeter());
401 }
402
403 /** Handles a click event in a display window. */
HandleClick(int x,int y)404 void StrokeWidth::HandleClick(int x, int y) {
405 BBGrid<BLOBNBOX, BLOBNBOX_CLIST, BLOBNBOX_C_IT>::HandleClick(x, y);
406 // Run a radial search for blobs that overlap.
407 BlobGridSearch radsearch(this);
408 radsearch.StartRadSearch(x, y, 1);
409 BLOBNBOX *neighbour;
410 FCOORD click(static_cast<float>(x), static_cast<float>(y));
411 while ((neighbour = radsearch.NextRadSearch()) != nullptr) {
412 TBOX nbox = neighbour->bounding_box();
413 if (nbox.contains(click) && neighbour->cblob() != nullptr) {
414 PrintBoxWidths(neighbour);
415 if (neighbour->neighbour(BND_LEFT) != nullptr) {
416 PrintBoxWidths(neighbour->neighbour(BND_LEFT));
417 }
418 if (neighbour->neighbour(BND_RIGHT) != nullptr) {
419 PrintBoxWidths(neighbour->neighbour(BND_RIGHT));
420 }
421 if (neighbour->neighbour(BND_ABOVE) != nullptr) {
422 PrintBoxWidths(neighbour->neighbour(BND_ABOVE));
423 }
424 if (neighbour->neighbour(BND_BELOW) != nullptr) {
425 PrintBoxWidths(neighbour->neighbour(BND_BELOW));
426 }
427 int gaps[BND_COUNT];
428 neighbour->NeighbourGaps(gaps);
429 tprintf(
430 "Left gap=%d, right=%d, above=%d, below=%d, horz=%d, vert=%d\n"
431 "Good= %d %d %d %d\n",
432 gaps[BND_LEFT], gaps[BND_RIGHT], gaps[BND_ABOVE], gaps[BND_BELOW],
433 neighbour->horz_possible(), neighbour->vert_possible(),
434 neighbour->good_stroke_neighbour(BND_LEFT), neighbour->good_stroke_neighbour(BND_RIGHT),
435 neighbour->good_stroke_neighbour(BND_ABOVE), neighbour->good_stroke_neighbour(BND_BELOW));
436 break;
437 }
438 }
439 }
440
441 // Detects and marks leader dots/dashes.
442 // Leaders are horizontal chains of small or noise blobs that look
443 // monospace according to ColPartition::MarkAsLeaderIfMonospaced().
444 // Detected leaders become the only occupants of the block->small_blobs list.
445 // Non-leader small blobs get moved to the blobs list.
446 // Non-leader noise blobs remain singletons in the noise list.
447 // All small and noise blobs in high density regions are marked BTFT_NONTEXT.
448 // block is the single block for the whole page or rectangle to be OCRed.
449 // leader_parts is the output.
FindLeadersAndMarkNoise(TO_BLOCK * block,ColPartition_LIST * leader_parts)450 void StrokeWidth::FindLeadersAndMarkNoise(TO_BLOCK *block, ColPartition_LIST *leader_parts) {
451 InsertBlobList(&block->small_blobs);
452 InsertBlobList(&block->noise_blobs);
453 BlobGridSearch gsearch(this);
454 BLOBNBOX *bbox;
455 // For every bbox in the grid, set its neighbours.
456 gsearch.StartFullSearch();
457 while ((bbox = gsearch.NextFullSearch()) != nullptr) {
458 SetNeighbours(true, false, bbox);
459 }
460 ColPartition_IT part_it(leader_parts);
461 gsearch.StartFullSearch();
462 while ((bbox = gsearch.NextFullSearch()) != nullptr) {
463 if (bbox->flow() == BTFT_NONE) {
464 if (bbox->neighbour(BND_RIGHT) == nullptr && bbox->neighbour(BND_LEFT) == nullptr) {
465 continue;
466 }
467 // Put all the linked blobs into a ColPartition.
468 auto *part = new ColPartition(BRT_UNKNOWN, ICOORD(0, 1));
469 BLOBNBOX *blob;
470 for (blob = bbox; blob != nullptr && blob->flow() == BTFT_NONE;
471 blob = blob->neighbour(BND_RIGHT)) {
472 part->AddBox(blob);
473 }
474 for (blob = bbox->neighbour(BND_LEFT); blob != nullptr && blob->flow() == BTFT_NONE;
475 blob = blob->neighbour(BND_LEFT)) {
476 part->AddBox(blob);
477 }
478 if (part->MarkAsLeaderIfMonospaced()) {
479 part_it.add_after_then_move(part);
480 } else {
481 delete part;
482 }
483 }
484 }
485 #ifndef GRAPHICS_DISABLED
486 if (textord_tabfind_show_strokewidths) {
487 leaders_win_ = DisplayGoodBlobs("LeaderNeighbours", 0, 0);
488 }
489 #endif
490 // Move any non-leaders from the small to the blobs list, as they are
491 // most likely dashes or broken characters.
492 BLOBNBOX_IT blob_it(&block->blobs);
493 BLOBNBOX_IT small_it(&block->small_blobs);
494 for (small_it.mark_cycle_pt(); !small_it.cycled_list(); small_it.forward()) {
495 BLOBNBOX *blob = small_it.data();
496 if (blob->flow() != BTFT_LEADER) {
497 if (blob->flow() == BTFT_NEIGHBOURS) {
498 blob->set_flow(BTFT_NONE);
499 }
500 blob->ClearNeighbours();
501 blob_it.add_to_end(small_it.extract());
502 }
503 }
504 // Move leaders from the noise list to the small list, leaving the small
505 // list exclusively leaders, so they don't get processed further,
506 // and the remaining small blobs all in the noise list.
507 BLOBNBOX_IT noise_it(&block->noise_blobs);
508 for (noise_it.mark_cycle_pt(); !noise_it.cycled_list(); noise_it.forward()) {
509 BLOBNBOX *blob = noise_it.data();
510 if (blob->flow() == BTFT_LEADER || blob->joined_to_prev()) {
511 small_it.add_to_end(noise_it.extract());
512 } else if (blob->flow() == BTFT_NEIGHBOURS) {
513 blob->set_flow(BTFT_NONE);
514 blob->ClearNeighbours();
515 }
516 }
517 // Clear the grid as we don't want the small stuff hanging around in it.
518 Clear();
519 }
520
521 /** Inserts the block blobs (normal and large) into this grid.
522 * Blobs remain owned by the block. */
InsertBlobs(TO_BLOCK * block)523 void StrokeWidth::InsertBlobs(TO_BLOCK *block) {
524 InsertBlobList(&block->blobs);
525 InsertBlobList(&block->large_blobs);
526 }
527
528 // Checks the left or right side of the given leader partition and sets the
529 // (opposite) leader_on_right or leader_on_left flags for blobs
530 // that are next to the given side of the given leader partition.
MarkLeaderNeighbours(const ColPartition * part,LeftOrRight side)531 void StrokeWidth::MarkLeaderNeighbours(const ColPartition *part, LeftOrRight side) {
532 const TBOX &part_box = part->bounding_box();
533 BlobGridSearch blobsearch(this);
534 // Search to the side of the leader for the nearest neighbour.
535 BLOBNBOX *best_blob = nullptr;
536 int best_gap = 0;
537 blobsearch.StartSideSearch(side == LR_LEFT ? part_box.left() : part_box.right(),
538 part_box.bottom(), part_box.top());
539 BLOBNBOX *blob;
540 while ((blob = blobsearch.NextSideSearch(side == LR_LEFT)) != nullptr) {
541 const TBOX &blob_box = blob->bounding_box();
542 if (!blob_box.y_overlap(part_box)) {
543 continue;
544 }
545 int x_gap = blob_box.x_gap(part_box);
546 if (x_gap > 2 * gridsize()) {
547 break;
548 } else if (best_blob == nullptr || x_gap < best_gap) {
549 best_blob = blob;
550 best_gap = x_gap;
551 }
552 }
553 if (best_blob != nullptr) {
554 if (side == LR_LEFT) {
555 best_blob->set_leader_on_right(true);
556 } else {
557 best_blob->set_leader_on_left(true);
558 }
559 #ifndef GRAPHICS_DISABLED
560 if (leaders_win_ != nullptr) {
561 leaders_win_->Pen(side == LR_LEFT ? ScrollView::RED : ScrollView::GREEN);
562 const TBOX &blob_box = best_blob->bounding_box();
563 leaders_win_->Rectangle(blob_box.left(), blob_box.bottom(), blob_box.right(), blob_box.top());
564 }
565 #endif // !GRAPHICS_DISABLED
566 }
567 }
568
569 // Helper to compute the UQ of the square-ish CJK characters.
UpperQuartileCJKSize(int gridsize,BLOBNBOX_LIST * blobs)570 static int UpperQuartileCJKSize(int gridsize, BLOBNBOX_LIST *blobs) {
571 STATS sizes(0, gridsize * kMaxCJKSizeRatio);
572 BLOBNBOX_IT it(blobs);
573 for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) {
574 BLOBNBOX *blob = it.data();
575 int width = blob->bounding_box().width();
576 int height = blob->bounding_box().height();
577 if (width <= height * kCJKAspectRatio && height < width * kCJKAspectRatio) {
578 sizes.add(height, 1);
579 }
580 }
581 return static_cast<int>(sizes.ile(0.75f) + 0.5);
582 }
583
584 // Fix broken CJK characters, using the fake joined blobs mechanism.
585 // Blobs are really merged, ie the master takes all the outlines and the
586 // others are deleted.
587 // Returns true if sufficient blobs are merged that it may be worth running
588 // again, due to a better estimate of character size.
FixBrokenCJK(TO_BLOCK * block)589 bool StrokeWidth::FixBrokenCJK(TO_BLOCK *block) {
590 BLOBNBOX_LIST *blobs = &block->blobs;
591 int median_height = UpperQuartileCJKSize(gridsize(), blobs);
592 int max_dist = static_cast<int>(median_height * kCJKBrokenDistanceFraction);
593 int max_height = static_cast<int>(median_height * kCJKAspectRatio);
594 int num_fixed = 0;
595 BLOBNBOX_IT blob_it(blobs);
596
597 for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) {
598 BLOBNBOX *blob = blob_it.data();
599 if (blob->cblob() == nullptr || blob->cblob()->out_list()->empty()) {
600 continue;
601 }
602 TBOX bbox = blob->bounding_box();
603 bool debug = AlignedBlob::WithinTestRegion(3, bbox.left(), bbox.bottom());
604 if (debug) {
605 tprintf("Checking for Broken CJK (max size=%d):", max_height);
606 bbox.print();
607 }
608 // Generate a list of blobs that overlap or are near enough to merge.
609 BLOBNBOX_CLIST overlapped_blobs;
610 AccumulateOverlaps(blob, debug, max_height, max_dist, &bbox, &overlapped_blobs);
611 if (!overlapped_blobs.empty()) {
612 // There are overlapping blobs, so qualify them as being satisfactory
613 // before removing them from the grid and replacing them with the union.
614 // The final box must be roughly square.
615 if (bbox.width() > bbox.height() * kCJKAspectRatio ||
616 bbox.height() > bbox.width() * kCJKAspectRatio) {
617 if (debug) {
618 tprintf("Bad final aspectratio:");
619 bbox.print();
620 }
621 continue;
622 }
623 // There can't be too many blobs to merge.
624 if (overlapped_blobs.length() >= kCJKMaxComponents) {
625 if (debug) {
626 tprintf("Too many neighbours: %d\n", overlapped_blobs.length());
627 }
628 continue;
629 }
630 // The strokewidths must match amongst the join candidates.
631 BLOBNBOX_C_IT n_it(&overlapped_blobs);
632 for (n_it.mark_cycle_pt(); !n_it.cycled_list(); n_it.forward()) {
633 BLOBNBOX *neighbour = nullptr;
634 neighbour = n_it.data();
635 if (!blob->MatchingStrokeWidth(*neighbour, kStrokeWidthFractionCJK, kStrokeWidthCJK)) {
636 break;
637 }
638 }
639 if (!n_it.cycled_list()) {
640 if (debug) {
641 tprintf("Bad stroke widths:");
642 PrintBoxWidths(blob);
643 }
644 continue; // Not good enough.
645 }
646
647 // Merge all the candidates into blob.
648 // We must remove blob from the grid and reinsert it after merging
649 // to maintain the integrity of the grid.
650 RemoveBBox(blob);
651 // Everything else will be calculated later.
652 for (n_it.mark_cycle_pt(); !n_it.cycled_list(); n_it.forward()) {
653 BLOBNBOX *neighbour = n_it.data();
654 RemoveBBox(neighbour);
655 // Mark empty blob for deletion.
656 neighbour->set_region_type(BRT_NOISE);
657 blob->really_merge(neighbour);
658 if (rerotation_.x() != 1.0f || rerotation_.y() != 0.0f) {
659 blob->rotate_box(rerotation_);
660 }
661 }
662 InsertBBox(true, true, blob);
663 ++num_fixed;
664 if (debug) {
665 tprintf("Done! Final box:");
666 bbox.print();
667 }
668 }
669 }
670 // Count remaining blobs.
671 int num_remaining = 0;
672 for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) {
673 BLOBNBOX *blob = blob_it.data();
674 if (blob->cblob() != nullptr && !blob->cblob()->out_list()->empty()) {
675 ++num_remaining;
676 }
677 }
678 // Permanently delete all the marked blobs after first removing all
679 // references in the neighbour members.
680 block->DeleteUnownedNoise();
681 return num_fixed > num_remaining * kBrokenCJKIterationFraction;
682 }
683
684 // Helper function to determine whether it is reasonable to merge the
685 // bbox and the nbox for repairing broken CJK.
686 // The distance apart must not exceed max_dist, the combined size must
687 // not exceed max_size, and the aspect ratio must either improve or at
688 // least not get worse by much.
AcceptableCJKMerge(const TBOX & bbox,const TBOX & nbox,bool debug,int max_size,int max_dist,int * x_gap,int * y_gap)689 static bool AcceptableCJKMerge(const TBOX &bbox, const TBOX &nbox, bool debug, int max_size,
690 int max_dist, int *x_gap, int *y_gap) {
691 *x_gap = bbox.x_gap(nbox);
692 *y_gap = bbox.y_gap(nbox);
693 TBOX merged(nbox);
694 merged += bbox;
695 if (debug) {
696 tprintf("gaps = %d, %d, merged_box:", *x_gap, *y_gap);
697 merged.print();
698 }
699 if (*x_gap <= max_dist && *y_gap <= max_dist && merged.width() <= max_size &&
700 merged.height() <= max_size) {
701 // Close enough to call overlapping. Check aspect ratios.
702 double old_ratio = static_cast<double>(bbox.width()) / bbox.height();
703 if (old_ratio < 1.0) {
704 old_ratio = 1.0 / old_ratio;
705 }
706 double new_ratio = static_cast<double>(merged.width()) / merged.height();
707 if (new_ratio < 1.0) {
708 new_ratio = 1.0 / new_ratio;
709 }
710 if (new_ratio <= old_ratio * kCJKAspectRatioIncrease) {
711 return true;
712 }
713 }
714 return false;
715 }
716
717 // Collect blobs that overlap or are within max_dist of the input bbox.
718 // Return them in the list of blobs and expand the bbox to be the union
719 // of all the boxes. not_this is excluded from the search, as are blobs
720 // that cause the merged box to exceed max_size in either dimension.
AccumulateOverlaps(const BLOBNBOX * not_this,bool debug,int max_size,int max_dist,TBOX * bbox,BLOBNBOX_CLIST * blobs)721 void StrokeWidth::AccumulateOverlaps(const BLOBNBOX *not_this, bool debug, int max_size,
722 int max_dist, TBOX *bbox, BLOBNBOX_CLIST *blobs) {
723 // While searching, nearests holds the nearest failed blob in each
724 // direction. When we have a nearest in each of the 4 directions, then
725 // the search is over, and at this point the final bbox must not overlap
726 // any of the nearests.
727 BLOBNBOX *nearests[BND_COUNT];
728 for (auto &nearest : nearests) {
729 nearest = nullptr;
730 }
731 int x = (bbox->left() + bbox->right()) / 2;
732 int y = (bbox->bottom() + bbox->top()) / 2;
733 // Run a radial search for blobs that overlap or are sufficiently close.
734 BlobGridSearch radsearch(this);
735 radsearch.StartRadSearch(x, y, kCJKRadius);
736 BLOBNBOX *neighbour;
737 while ((neighbour = radsearch.NextRadSearch()) != nullptr) {
738 if (neighbour == not_this) {
739 continue;
740 }
741 TBOX nbox = neighbour->bounding_box();
742 int x_gap, y_gap;
743 if (AcceptableCJKMerge(*bbox, nbox, debug, max_size, max_dist, &x_gap, &y_gap)) {
744 // Close enough to call overlapping. Merge boxes.
745 *bbox += nbox;
746 blobs->add_sorted(SortByBoxLeft<BLOBNBOX>, true, neighbour);
747 if (debug) {
748 tprintf("Added:");
749 nbox.print();
750 }
751 // Since we merged, search the nearests, as some might now me mergeable.
752 for (int dir = 0; dir < BND_COUNT; ++dir) {
753 if (nearests[dir] == nullptr) {
754 continue;
755 }
756 nbox = nearests[dir]->bounding_box();
757 if (AcceptableCJKMerge(*bbox, nbox, debug, max_size, max_dist, &x_gap, &y_gap)) {
758 // Close enough to call overlapping. Merge boxes.
759 *bbox += nbox;
760 blobs->add_sorted(SortByBoxLeft<BLOBNBOX>, true, nearests[dir]);
761 if (debug) {
762 tprintf("Added:");
763 nbox.print();
764 }
765 nearests[dir] = nullptr;
766 dir = -1; // Restart the search.
767 }
768 }
769 } else if (x_gap < 0 && x_gap <= y_gap) {
770 // A vertical neighbour. Record the nearest.
771 BlobNeighbourDir dir = nbox.top() > bbox->top() ? BND_ABOVE : BND_BELOW;
772 if (nearests[dir] == nullptr || y_gap < bbox->y_gap(nearests[dir]->bounding_box())) {
773 nearests[dir] = neighbour;
774 }
775 } else if (y_gap < 0 && y_gap <= x_gap) {
776 // A horizontal neighbour. Record the nearest.
777 BlobNeighbourDir dir = nbox.left() > bbox->left() ? BND_RIGHT : BND_LEFT;
778 if (nearests[dir] == nullptr || x_gap < bbox->x_gap(nearests[dir]->bounding_box())) {
779 nearests[dir] = neighbour;
780 }
781 }
782 // If all nearests are non-null, then we have finished.
783 if (nearests[BND_LEFT] && nearests[BND_RIGHT] && nearests[BND_ABOVE] && nearests[BND_BELOW]) {
784 break;
785 }
786 }
787 // Final overlap with a nearest is not allowed.
788 for (auto &nearest : nearests) {
789 if (nearest == nullptr) {
790 continue;
791 }
792 const TBOX &nbox = nearest->bounding_box();
793 if (debug) {
794 tprintf("Testing for overlap with:");
795 nbox.print();
796 }
797 if (bbox->overlap(nbox)) {
798 blobs->shallow_clear();
799 if (debug) {
800 tprintf("Final box overlaps nearest\n");
801 }
802 return;
803 }
804 }
805 }
806
807 // For each blob in this grid, Finds the textline direction to be horizontal
808 // or vertical according to distance to neighbours and 1st and 2nd order
809 // neighbours. Non-text tends to end up without a definite direction.
810 // Result is setting of the neighbours and vert_possible/horz_possible
811 // flags in the BLOBNBOXes currently in this grid.
812 // This function is called more than once if page orientation is uncertain,
813 // so display_if_debugging is true on the final call to display the results.
FindTextlineFlowDirection(PageSegMode pageseg_mode,bool display_if_debugging)814 void StrokeWidth::FindTextlineFlowDirection(PageSegMode pageseg_mode, bool display_if_debugging) {
815 BlobGridSearch gsearch(this);
816 BLOBNBOX *bbox;
817 // For every bbox in the grid, set its neighbours.
818 gsearch.StartFullSearch();
819 while ((bbox = gsearch.NextFullSearch()) != nullptr) {
820 SetNeighbours(false, display_if_debugging, bbox);
821 }
822 // Where vertical or horizontal wins by a big margin, clarify it.
823 gsearch.StartFullSearch();
824 while ((bbox = gsearch.NextFullSearch()) != nullptr) {
825 SimplifyObviousNeighbours(bbox);
826 }
827 // Now try to make the blobs only vertical or horizontal using neighbours.
828 gsearch.StartFullSearch();
829 while ((bbox = gsearch.NextFullSearch()) != nullptr) {
830 if (FindingVerticalOnly(pageseg_mode)) {
831 bbox->set_vert_possible(true);
832 bbox->set_horz_possible(false);
833 } else if (FindingHorizontalOnly(pageseg_mode)) {
834 bbox->set_vert_possible(false);
835 bbox->set_horz_possible(true);
836 } else {
837 SetNeighbourFlows(bbox);
838 }
839 }
840 #ifndef GRAPHICS_DISABLED
841 if ((textord_tabfind_show_strokewidths && display_if_debugging) ||
842 textord_tabfind_show_strokewidths > 1) {
843 initial_widths_win_ = DisplayGoodBlobs("InitialStrokewidths", 400, 0);
844 }
845 #endif
846 // Improve flow direction with neighbours.
847 gsearch.StartFullSearch();
848 while ((bbox = gsearch.NextFullSearch()) != nullptr) {
849 SmoothNeighbourTypes(pageseg_mode, false, bbox);
850 }
851 // Now allow reset of firm values to fix renegades.
852 gsearch.StartFullSearch();
853 while ((bbox = gsearch.NextFullSearch()) != nullptr) {
854 SmoothNeighbourTypes(pageseg_mode, true, bbox);
855 }
856 // Repeat.
857 gsearch.StartFullSearch();
858 while ((bbox = gsearch.NextFullSearch()) != nullptr) {
859 SmoothNeighbourTypes(pageseg_mode, true, bbox);
860 }
861 #ifndef GRAPHICS_DISABLED
862 if ((textord_tabfind_show_strokewidths && display_if_debugging) ||
863 textord_tabfind_show_strokewidths > 1) {
864 widths_win_ = DisplayGoodBlobs("ImprovedStrokewidths", 800, 0);
865 }
866 #endif
867 }
868
869 // Sets the neighbours and good_stroke_neighbours members of the blob by
870 // searching close on all 4 sides.
871 // When finding leader dots/dashes, there is a slightly different rule for
872 // what makes a good neighbour.
SetNeighbours(bool leaders,bool activate_line_trap,BLOBNBOX * blob)873 void StrokeWidth::SetNeighbours(bool leaders, bool activate_line_trap, BLOBNBOX *blob) {
874 int line_trap_count = 0;
875 for (int dir = 0; dir < BND_COUNT; ++dir) {
876 auto bnd = static_cast<BlobNeighbourDir>(dir);
877 line_trap_count += FindGoodNeighbour(bnd, leaders, blob);
878 }
879 if (line_trap_count > 0 && activate_line_trap) {
880 // It looks like a line so isolate it by clearing its neighbours.
881 blob->ClearNeighbours();
882 const TBOX &box = blob->bounding_box();
883 blob->set_region_type(box.width() > box.height() ? BRT_HLINE : BRT_VLINE);
884 }
885 }
886
887 // Sets the good_stroke_neighbours member of the blob if it has a
888 // GoodNeighbour on the given side.
889 // Also sets the neighbour in the blob, whether or not a good one is found.
890 // Returns the number of blobs in the nearby search area that would lead us to
891 // believe that this blob is a line separator.
892 // Leaders get extra special lenient treatment.
FindGoodNeighbour(BlobNeighbourDir dir,bool leaders,BLOBNBOX * blob)893 int StrokeWidth::FindGoodNeighbour(BlobNeighbourDir dir, bool leaders, BLOBNBOX *blob) {
894 // Search for neighbours that overlap vertically.
895 TBOX blob_box = blob->bounding_box();
896 bool debug = AlignedBlob::WithinTestRegion(2, blob_box.left(), blob_box.bottom());
897 if (debug) {
898 tprintf("FGN in dir %d for blob:", dir);
899 blob_box.print();
900 }
901 int top = blob_box.top();
902 int bottom = blob_box.bottom();
903 int left = blob_box.left();
904 int right = blob_box.right();
905 int width = right - left;
906 int height = top - bottom;
907
908 // A trap to detect lines tests for the min dimension of neighbours
909 // being larger than a multiple of the min dimension of the line
910 // and the larger dimension being smaller than a fraction of the max
911 // dimension of the line.
912 int line_trap_max = std::max(width, height) / kLineTrapLongest;
913 int line_trap_min = std::min(width, height) * kLineTrapShortest;
914 int line_trap_count = 0;
915
916 int min_good_overlap = (dir == BND_LEFT || dir == BND_RIGHT) ? height / 2 : width / 2;
917 int min_decent_overlap = (dir == BND_LEFT || dir == BND_RIGHT) ? height / 3 : width / 3;
918 if (leaders) {
919 min_good_overlap = min_decent_overlap = 1;
920 }
921
922 int search_pad =
923 static_cast<int>(sqrt(static_cast<double>(width * height)) * kNeighbourSearchFactor);
924 if (gridsize() > search_pad) {
925 search_pad = gridsize();
926 }
927 TBOX search_box = blob_box;
928 // Pad the search in the appropriate direction.
929 switch (dir) {
930 case BND_LEFT:
931 search_box.set_left(search_box.left() - search_pad);
932 break;
933 case BND_RIGHT:
934 search_box.set_right(search_box.right() + search_pad);
935 break;
936 case BND_BELOW:
937 search_box.set_bottom(search_box.bottom() - search_pad);
938 break;
939 case BND_ABOVE:
940 search_box.set_top(search_box.top() + search_pad);
941 break;
942 case BND_COUNT:
943 return 0;
944 }
945
946 BlobGridSearch rectsearch(this);
947 rectsearch.StartRectSearch(search_box);
948 BLOBNBOX *best_neighbour = nullptr;
949 double best_goodness = 0.0;
950 bool best_is_good = false;
951 BLOBNBOX *neighbour;
952 while ((neighbour = rectsearch.NextRectSearch()) != nullptr) {
953 TBOX nbox = neighbour->bounding_box();
954 if (neighbour == blob) {
955 continue;
956 }
957 int mid_x = (nbox.left() + nbox.right()) / 2;
958 if (mid_x < blob->left_rule() || mid_x > blob->right_rule()) {
959 continue; // In a different column.
960 }
961 if (debug) {
962 tprintf("Neighbour at:");
963 nbox.print();
964 }
965
966 // Last-minute line detector. There is a small upper limit to the line
967 // width accepted by the morphological line detector.
968 int n_width = nbox.width();
969 int n_height = nbox.height();
970 if (std::min(n_width, n_height) > line_trap_min &&
971 std::max(n_width, n_height) < line_trap_max) {
972 ++line_trap_count;
973 }
974 // Heavily joined text, such as Arabic may have very different sizes when
975 // looking at the maxes, but the heights may be almost identical, so check
976 // for a difference in height if looking sideways or width vertically.
977 if (TabFind::VeryDifferentSizes(std::max(n_width, n_height), std::max(width, height)) &&
978 (((dir == BND_LEFT || dir == BND_RIGHT) && TabFind::DifferentSizes(n_height, height)) ||
979 ((dir == BND_BELOW || dir == BND_ABOVE) && TabFind::DifferentSizes(n_width, width)))) {
980 if (debug) {
981 tprintf("Bad size\n");
982 }
983 continue; // Could be a different font size or non-text.
984 }
985 // Amount of vertical overlap between the blobs.
986 int overlap;
987 // If the overlap is along the short side of the neighbour, and it
988 // is fully overlapped, then perp_overlap holds the length of the long
989 // side of the neighbour. A measure to include hyphens and dashes as
990 // legitimate neighbours.
991 int perp_overlap;
992 int gap;
993 if (dir == BND_LEFT || dir == BND_RIGHT) {
994 overlap = std::min(static_cast<int>(nbox.top()), top) -
995 std::max(static_cast<int>(nbox.bottom()), bottom);
996 if (overlap == nbox.height() && nbox.width() > nbox.height()) {
997 perp_overlap = nbox.width();
998 } else {
999 perp_overlap = overlap;
1000 }
1001 gap = dir == BND_LEFT ? left - nbox.left() : nbox.right() - right;
1002 if (gap <= 0) {
1003 if (debug) {
1004 tprintf("On wrong side\n");
1005 }
1006 continue; // On the wrong side.
1007 }
1008 gap -= n_width;
1009 } else {
1010 overlap = std::min(static_cast<int>(nbox.right()), right) -
1011 std::max(static_cast<int>(nbox.left()), left);
1012 if (overlap == nbox.width() && nbox.height() > nbox.width()) {
1013 perp_overlap = nbox.height();
1014 } else {
1015 perp_overlap = overlap;
1016 }
1017 gap = dir == BND_BELOW ? bottom - nbox.bottom() : nbox.top() - top;
1018 if (gap <= 0) {
1019 if (debug) {
1020 tprintf("On wrong side\n");
1021 }
1022 continue; // On the wrong side.
1023 }
1024 gap -= n_height;
1025 }
1026 if (-gap > overlap) {
1027 if (debug) {
1028 tprintf("Overlaps wrong way\n");
1029 }
1030 continue; // Overlaps the wrong way.
1031 }
1032 if (perp_overlap < min_decent_overlap) {
1033 if (debug) {
1034 tprintf("Doesn't overlap enough\n");
1035 }
1036 continue; // Doesn't overlap enough.
1037 }
1038 bool bad_sizes =
1039 TabFind::DifferentSizes(height, n_height) && TabFind::DifferentSizes(width, n_width);
1040 bool is_good =
1041 overlap >= min_good_overlap && !bad_sizes &&
1042 blob->MatchingStrokeWidth(*neighbour, kStrokeWidthFractionTolerance, kStrokeWidthTolerance);
1043 // Best is a fuzzy combination of gap, overlap and is good.
1044 // Basically if you make one thing twice as good without making
1045 // anything else twice as bad, then it is better.
1046 if (gap < 1) {
1047 gap = 1;
1048 }
1049 double goodness = (1.0 + is_good) * overlap / gap;
1050 if (debug) {
1051 tprintf("goodness = %g vs best of %g, good=%d, overlap=%d, gap=%d\n", goodness, best_goodness,
1052 is_good, overlap, gap);
1053 }
1054 if (goodness > best_goodness) {
1055 best_neighbour = neighbour;
1056 best_goodness = goodness;
1057 best_is_good = is_good;
1058 }
1059 }
1060 blob->set_neighbour(dir, best_neighbour, best_is_good);
1061 return line_trap_count;
1062 }
1063
1064 // Helper to get a list of 1st-order neighbours.
ListNeighbours(const BLOBNBOX * blob,BLOBNBOX_CLIST * neighbours)1065 static void ListNeighbours(const BLOBNBOX *blob, BLOBNBOX_CLIST *neighbours) {
1066 for (int dir = 0; dir < BND_COUNT; ++dir) {
1067 auto bnd = static_cast<BlobNeighbourDir>(dir);
1068 BLOBNBOX *neighbour = blob->neighbour(bnd);
1069 if (neighbour != nullptr) {
1070 neighbours->add_sorted(SortByBoxLeft<BLOBNBOX>, true, neighbour);
1071 }
1072 }
1073 }
1074
1075 // Helper to get a list of 1st and 2nd order neighbours.
List2ndNeighbours(const BLOBNBOX * blob,BLOBNBOX_CLIST * neighbours)1076 static void List2ndNeighbours(const BLOBNBOX *blob, BLOBNBOX_CLIST *neighbours) {
1077 ListNeighbours(blob, neighbours);
1078 for (int dir = 0; dir < BND_COUNT; ++dir) {
1079 auto bnd = static_cast<BlobNeighbourDir>(dir);
1080 BLOBNBOX *neighbour = blob->neighbour(bnd);
1081 if (neighbour != nullptr) {
1082 ListNeighbours(neighbour, neighbours);
1083 }
1084 }
1085 }
1086
1087 // Helper to get a list of 1st, 2nd and 3rd order neighbours.
List3rdNeighbours(const BLOBNBOX * blob,BLOBNBOX_CLIST * neighbours)1088 static void List3rdNeighbours(const BLOBNBOX *blob, BLOBNBOX_CLIST *neighbours) {
1089 List2ndNeighbours(blob, neighbours);
1090 for (int dir = 0; dir < BND_COUNT; ++dir) {
1091 auto bnd = static_cast<BlobNeighbourDir>(dir);
1092 BLOBNBOX *neighbour = blob->neighbour(bnd);
1093 if (neighbour != nullptr) {
1094 List2ndNeighbours(neighbour, neighbours);
1095 }
1096 }
1097 }
1098
1099 // Helper to count the evidence for verticalness or horizontalness
1100 // in a list of neighbours.
CountNeighbourGaps(bool debug,BLOBNBOX_CLIST * neighbours,int * pure_h_count,int * pure_v_count)1101 static void CountNeighbourGaps(bool debug, BLOBNBOX_CLIST *neighbours, int *pure_h_count,
1102 int *pure_v_count) {
1103 if (neighbours->length() <= kMostlyOneDirRatio) {
1104 return;
1105 }
1106 BLOBNBOX_C_IT it(neighbours);
1107 for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) {
1108 BLOBNBOX *blob = it.data();
1109 int h_min, h_max, v_min, v_max;
1110 blob->MinMaxGapsClipped(&h_min, &h_max, &v_min, &v_max);
1111 if (debug) {
1112 tprintf("Hgaps [%d,%d], vgaps [%d,%d]:", h_min, h_max, v_min, v_max);
1113 }
1114 if (h_max < v_min || blob->leader_on_left() || blob->leader_on_right()) {
1115 // Horizontal gaps are clear winners. Count a pure horizontal.
1116 ++*pure_h_count;
1117 if (debug) {
1118 tprintf("Horz at:");
1119 }
1120 } else if (v_max < h_min) {
1121 // Vertical gaps are clear winners. Clear a pure vertical.
1122 ++*pure_v_count;
1123 if (debug) {
1124 tprintf("Vert at:");
1125 }
1126 } else {
1127 if (debug) {
1128 tprintf("Neither at:");
1129 }
1130 }
1131 if (debug) {
1132 blob->bounding_box().print();
1133 }
1134 }
1135 }
1136
1137 // Makes the blob to be only horizontal or vertical where evidence
1138 // is clear based on gaps of 2nd order neighbours, or definite individual
1139 // blobs.
SetNeighbourFlows(BLOBNBOX * blob)1140 void StrokeWidth::SetNeighbourFlows(BLOBNBOX *blob) {
1141 if (blob->DefiniteIndividualFlow()) {
1142 return;
1143 }
1144 bool debug =
1145 AlignedBlob::WithinTestRegion(2, blob->bounding_box().left(), blob->bounding_box().bottom());
1146 if (debug) {
1147 tprintf("SetNeighbourFlows (current flow=%d, type=%d) on:", blob->flow(), blob->region_type());
1148 blob->bounding_box().print();
1149 }
1150 BLOBNBOX_CLIST neighbours;
1151 List3rdNeighbours(blob, &neighbours);
1152 // The number of pure horizontal and vertical neighbours.
1153 int pure_h_count = 0;
1154 int pure_v_count = 0;
1155 CountNeighbourGaps(debug, &neighbours, &pure_h_count, &pure_v_count);
1156 if (debug) {
1157 HandleClick(blob->bounding_box().left() + 1, blob->bounding_box().bottom() + 1);
1158 tprintf("SetFlows: h_count=%d, v_count=%d\n", pure_h_count, pure_v_count);
1159 }
1160 if (!neighbours.empty()) {
1161 blob->set_vert_possible(true);
1162 blob->set_horz_possible(true);
1163 if (pure_h_count > 2 * pure_v_count) {
1164 // Horizontal gaps are clear winners. Clear vertical neighbours.
1165 blob->set_vert_possible(false);
1166 } else if (pure_v_count > 2 * pure_h_count) {
1167 // Vertical gaps are clear winners. Clear horizontal neighbours.
1168 blob->set_horz_possible(false);
1169 }
1170 } else {
1171 // Lonely blob. Can't tell its flow direction.
1172 blob->set_vert_possible(false);
1173 blob->set_horz_possible(false);
1174 }
1175 }
1176
1177 // Helper to count the number of horizontal and vertical blobs in a list.
CountNeighbourTypes(BLOBNBOX_CLIST * neighbours,int * pure_h_count,int * pure_v_count)1178 static void CountNeighbourTypes(BLOBNBOX_CLIST *neighbours, int *pure_h_count, int *pure_v_count) {
1179 BLOBNBOX_C_IT it(neighbours);
1180 for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) {
1181 BLOBNBOX *blob = it.data();
1182 if (blob->UniquelyHorizontal()) {
1183 ++*pure_h_count;
1184 }
1185 if (blob->UniquelyVertical()) {
1186 ++*pure_v_count;
1187 }
1188 }
1189 }
1190
1191 // Nullify the neighbours in the wrong directions where the direction
1192 // is clear-cut based on a distance margin. Good for isolating vertical
1193 // text from neighbouring horizontal text.
SimplifyObviousNeighbours(BLOBNBOX * blob)1194 void StrokeWidth::SimplifyObviousNeighbours(BLOBNBOX *blob) {
1195 // Case 1: We have text that is likely several characters, blurry and joined
1196 // together.
1197 if ((blob->bounding_box().width() > 3 * blob->area_stroke_width() &&
1198 blob->bounding_box().height() > 3 * blob->area_stroke_width())) {
1199 // The blob is complex (not stick-like).
1200 if (blob->bounding_box().width() > 4 * blob->bounding_box().height()) {
1201 // Horizontal conjoined text.
1202 blob->set_neighbour(BND_ABOVE, nullptr, false);
1203 blob->set_neighbour(BND_BELOW, nullptr, false);
1204 return;
1205 }
1206 if (blob->bounding_box().height() > 4 * blob->bounding_box().width()) {
1207 // Vertical conjoined text.
1208 blob->set_neighbour(BND_LEFT, nullptr, false);
1209 blob->set_neighbour(BND_RIGHT, nullptr, false);
1210 return;
1211 }
1212 }
1213
1214 // Case 2: This blob is likely a single character.
1215 int margin = gridsize() / 2;
1216 int h_min, h_max, v_min, v_max;
1217 blob->MinMaxGapsClipped(&h_min, &h_max, &v_min, &v_max);
1218 if ((h_max + margin < v_min && h_max < margin / 2) || blob->leader_on_left() ||
1219 blob->leader_on_right()) {
1220 // Horizontal gaps are clear winners. Clear vertical neighbours.
1221 blob->set_neighbour(BND_ABOVE, nullptr, false);
1222 blob->set_neighbour(BND_BELOW, nullptr, false);
1223 } else if (v_max + margin < h_min && v_max < margin / 2) {
1224 // Vertical gaps are clear winners. Clear horizontal neighbours.
1225 blob->set_neighbour(BND_LEFT, nullptr, false);
1226 blob->set_neighbour(BND_RIGHT, nullptr, false);
1227 }
1228 }
1229
1230 // Smoothes the vertical/horizontal type of the blob based on the
1231 // 2nd-order neighbours. If reset_all is true, then all blobs are
1232 // changed. Otherwise, only ambiguous blobs are processed.
SmoothNeighbourTypes(PageSegMode pageseg_mode,bool reset_all,BLOBNBOX * blob)1233 void StrokeWidth::SmoothNeighbourTypes(PageSegMode pageseg_mode, bool reset_all, BLOBNBOX *blob) {
1234 if ((blob->vert_possible() && blob->horz_possible()) || reset_all) {
1235 // There are both horizontal and vertical so try to fix it.
1236 BLOBNBOX_CLIST neighbours;
1237 List2ndNeighbours(blob, &neighbours);
1238 // The number of pure horizontal and vertical neighbours.
1239 int pure_h_count = 0;
1240 int pure_v_count = 0;
1241 CountNeighbourTypes(&neighbours, &pure_h_count, &pure_v_count);
1242 if (AlignedBlob::WithinTestRegion(2, blob->bounding_box().left(),
1243 blob->bounding_box().bottom())) {
1244 HandleClick(blob->bounding_box().left() + 1, blob->bounding_box().bottom() + 1);
1245 tprintf("pure_h=%d, pure_v=%d\n", pure_h_count, pure_v_count);
1246 }
1247 if (pure_h_count > pure_v_count && !FindingVerticalOnly(pageseg_mode)) {
1248 // Horizontal gaps are clear winners. Clear vertical neighbours.
1249 blob->set_vert_possible(false);
1250 blob->set_horz_possible(true);
1251 } else if (pure_v_count > pure_h_count && !FindingHorizontalOnly(pageseg_mode)) {
1252 // Vertical gaps are clear winners. Clear horizontal neighbours.
1253 blob->set_horz_possible(false);
1254 blob->set_vert_possible(true);
1255 }
1256 } else if (AlignedBlob::WithinTestRegion(2, blob->bounding_box().left(),
1257 blob->bounding_box().bottom())) {
1258 HandleClick(blob->bounding_box().left() + 1, blob->bounding_box().bottom() + 1);
1259 tprintf("Clean on pass 3!\n");
1260 }
1261 }
1262
1263 // Partition creation. Accumulates vertical and horizontal text chains,
1264 // puts the remaining blobs in as unknowns, and then merges/splits to
1265 // minimize overlap and smoothes the types with neighbours and the color
1266 // image if provided. rerotation is used to rotate the coordinate space
1267 // back to the nontext_map_ image.
1268 // If find_problems is true, detects possible noise pollution by the amount
1269 // of partition overlap that is created by the diacritics. If excessive, the
1270 // noise is separated out into diacritic blobs, and PFR_NOISE is returned.
1271 // [TODO(rays): if the partition overlap is caused by heavy skew, deskews
1272 // the components, saves the skew_angle and returns PFR_SKEW.] If the return
1273 // is not PFR_OK, the job is incomplete, and FindInitialPartitions must be
1274 // called again after cleaning up the partly done work.
FindInitialPartitions(PageSegMode pageseg_mode,const FCOORD & rerotation,bool find_problems,TO_BLOCK * block,BLOBNBOX_LIST * diacritic_blobs,ColPartitionGrid * part_grid,ColPartition_LIST * big_parts,FCOORD * skew_angle)1275 PartitionFindResult StrokeWidth::FindInitialPartitions(
1276 PageSegMode pageseg_mode, const FCOORD &rerotation, bool find_problems, TO_BLOCK *block,
1277 BLOBNBOX_LIST *diacritic_blobs, ColPartitionGrid *part_grid, ColPartition_LIST *big_parts,
1278 FCOORD *skew_angle) {
1279 if (!FindingHorizontalOnly(pageseg_mode)) {
1280 FindVerticalTextChains(part_grid);
1281 }
1282 if (!FindingVerticalOnly(pageseg_mode)) {
1283 FindHorizontalTextChains(part_grid);
1284 }
1285 #ifndef GRAPHICS_DISABLED
1286 if (textord_tabfind_show_strokewidths) {
1287 chains_win_ = MakeWindow(0, 400, "Initial text chains");
1288 part_grid->DisplayBoxes(chains_win_);
1289 projection_->DisplayProjection();
1290 }
1291 #endif
1292 if (find_problems) {
1293 // TODO(rays) Do something to find skew, set skew_angle and return if there
1294 // is some.
1295 }
1296 part_grid->SplitOverlappingPartitions(big_parts);
1297 EasyMerges(part_grid);
1298 RemoveLargeUnusedBlobs(block, part_grid, big_parts);
1299 TBOX grid_box(bleft(), tright());
1300 while (part_grid->GridSmoothNeighbours(BTFT_CHAIN, nontext_map_, grid_box, rerotation)) {
1301 ;
1302 }
1303 while (part_grid->GridSmoothNeighbours(BTFT_NEIGHBOURS, nontext_map_, grid_box, rerotation)) {
1304 ;
1305 }
1306 int pre_overlap = part_grid->ComputeTotalOverlap(nullptr);
1307 TestDiacritics(part_grid, block);
1308 MergeDiacritics(block, part_grid);
1309 if (find_problems && diacritic_blobs != nullptr &&
1310 DetectAndRemoveNoise(pre_overlap, grid_box, block, part_grid, diacritic_blobs)) {
1311 return PFR_NOISE;
1312 }
1313 #ifndef GRAPHICS_DISABLED
1314 if (textord_tabfind_show_strokewidths) {
1315 textlines_win_ = MakeWindow(400, 400, "GoodTextline blobs");
1316 part_grid->DisplayBoxes(textlines_win_);
1317 diacritics_win_ = DisplayDiacritics("Diacritics", 0, 0, block);
1318 }
1319 #endif
1320 PartitionRemainingBlobs(pageseg_mode, part_grid);
1321 part_grid->SplitOverlappingPartitions(big_parts);
1322 EasyMerges(part_grid);
1323 while (part_grid->GridSmoothNeighbours(BTFT_CHAIN, nontext_map_, grid_box, rerotation)) {
1324 ;
1325 }
1326 while (part_grid->GridSmoothNeighbours(BTFT_NEIGHBOURS, nontext_map_, grid_box, rerotation)) {
1327 ;
1328 }
1329 // Now eliminate strong stuff in a sea of the opposite.
1330 while (part_grid->GridSmoothNeighbours(BTFT_STRONG_CHAIN, nontext_map_, grid_box, rerotation)) {
1331 ;
1332 }
1333 #ifndef GRAPHICS_DISABLED
1334 if (textord_tabfind_show_strokewidths) {
1335 smoothed_win_ = MakeWindow(800, 400, "Smoothed blobs");
1336 part_grid->DisplayBoxes(smoothed_win_);
1337 }
1338 #endif
1339 return PFR_OK;
1340 }
1341
1342 // Detects noise by a significant increase in partition overlap from
1343 // pre_overlap to now, and removes noise from the union of all the overlapping
1344 // partitions, placing the blobs in diacritic_blobs. Returns true if any noise
1345 // was found and removed.
DetectAndRemoveNoise(int pre_overlap,const TBOX & grid_box,TO_BLOCK * block,ColPartitionGrid * part_grid,BLOBNBOX_LIST * diacritic_blobs)1346 bool StrokeWidth::DetectAndRemoveNoise(int pre_overlap, const TBOX &grid_box, TO_BLOCK *block,
1347 ColPartitionGrid *part_grid,
1348 BLOBNBOX_LIST *diacritic_blobs) {
1349 ColPartitionGrid *noise_grid = nullptr;
1350 int post_overlap = part_grid->ComputeTotalOverlap(&noise_grid);
1351 if (pre_overlap == 0) {
1352 pre_overlap = 1;
1353 }
1354 BLOBNBOX_IT diacritic_it(diacritic_blobs);
1355 if (noise_grid != nullptr) {
1356 if (post_overlap > pre_overlap * kNoiseOverlapGrowthFactor &&
1357 post_overlap > grid_box.area() * kNoiseOverlapAreaFactor) {
1358 // This is noisy enough to fix.
1359 #ifndef GRAPHICS_DISABLED
1360 if (textord_tabfind_show_strokewidths) {
1361 ScrollView *noise_win = MakeWindow(1000, 500, "Noise Areas");
1362 noise_grid->DisplayBoxes(noise_win);
1363 }
1364 #endif
1365 part_grid->DeleteNonLeaderParts();
1366 BLOBNBOX_IT blob_it(&block->noise_blobs);
1367 ColPartitionGridSearch rsearch(noise_grid);
1368 for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) {
1369 BLOBNBOX *blob = blob_it.data();
1370 blob->ClearNeighbours();
1371 if (!blob->IsDiacritic() || blob->owner() != nullptr) {
1372 continue; // Not a noise candidate.
1373 }
1374 TBOX search_box(blob->bounding_box());
1375 search_box.pad(gridsize(), gridsize());
1376 rsearch.StartRectSearch(search_box);
1377 ColPartition *part = rsearch.NextRectSearch();
1378 if (part != nullptr) {
1379 // Consider blob as possible noise.
1380 blob->set_owns_cblob(true);
1381 blob->compute_bounding_box();
1382 diacritic_it.add_after_then_move(blob_it.extract());
1383 }
1384 }
1385 noise_grid->DeleteParts();
1386 delete noise_grid;
1387 return true;
1388 }
1389 noise_grid->DeleteParts();
1390 delete noise_grid;
1391 }
1392 return false;
1393 }
1394
1395 // Helper verifies that blob's neighbour in direction dir is good to add to a
1396 // vertical text chain by returning the neighbour if it is not null, not owned,
1397 // and not uniquely horizontal, as well as its neighbour in the opposite
1398 // direction is blob.
MutualUnusedVNeighbour(const BLOBNBOX * blob,BlobNeighbourDir dir)1399 static BLOBNBOX *MutualUnusedVNeighbour(const BLOBNBOX *blob, BlobNeighbourDir dir) {
1400 BLOBNBOX *next_blob = blob->neighbour(dir);
1401 if (next_blob == nullptr || next_blob->owner() != nullptr || next_blob->UniquelyHorizontal()) {
1402 return nullptr;
1403 }
1404 if (next_blob->neighbour(DirOtherWay(dir)) == blob) {
1405 return next_blob;
1406 }
1407 return nullptr;
1408 }
1409
1410 // Finds vertical chains of text-like blobs and puts them in ColPartitions.
FindVerticalTextChains(ColPartitionGrid * part_grid)1411 void StrokeWidth::FindVerticalTextChains(ColPartitionGrid *part_grid) {
1412 // A PageSegMode that forces vertical textlines with the current rotation.
1413 PageSegMode pageseg_mode =
1414 rerotation_.y() == 0.0f ? PSM_SINGLE_BLOCK_VERT_TEXT : PSM_SINGLE_COLUMN;
1415 BlobGridSearch gsearch(this);
1416 BLOBNBOX *bbox;
1417 gsearch.StartFullSearch();
1418 while ((bbox = gsearch.NextFullSearch()) != nullptr) {
1419 // Only process boxes that have no horizontal hope and have not yet
1420 // been included in a chain.
1421 BLOBNBOX *blob;
1422 if (bbox->owner() == nullptr && bbox->UniquelyVertical() &&
1423 (blob = MutualUnusedVNeighbour(bbox, BND_ABOVE)) != nullptr) {
1424 // Put all the linked blobs into a ColPartition.
1425 auto *part = new ColPartition(BRT_VERT_TEXT, ICOORD(0, 1));
1426 part->AddBox(bbox);
1427 while (blob != nullptr) {
1428 part->AddBox(blob);
1429 blob = MutualUnusedVNeighbour(blob, BND_ABOVE);
1430 }
1431 blob = MutualUnusedVNeighbour(bbox, BND_BELOW);
1432 while (blob != nullptr) {
1433 part->AddBox(blob);
1434 blob = MutualUnusedVNeighbour(blob, BND_BELOW);
1435 }
1436 CompletePartition(pageseg_mode, part, part_grid);
1437 }
1438 }
1439 }
1440
1441 // Helper verifies that blob's neighbour in direction dir is good to add to a
1442 // horizontal text chain by returning the neighbour if it is not null, not
1443 // owned, and not uniquely vertical, as well as its neighbour in the opposite
1444 // direction is blob.
MutualUnusedHNeighbour(const BLOBNBOX * blob,BlobNeighbourDir dir)1445 static BLOBNBOX *MutualUnusedHNeighbour(const BLOBNBOX *blob, BlobNeighbourDir dir) {
1446 BLOBNBOX *next_blob = blob->neighbour(dir);
1447 if (next_blob == nullptr || next_blob->owner() != nullptr || next_blob->UniquelyVertical()) {
1448 return nullptr;
1449 }
1450 if (next_blob->neighbour(DirOtherWay(dir)) == blob) {
1451 return next_blob;
1452 }
1453 return nullptr;
1454 }
1455
1456 // Finds horizontal chains of text-like blobs and puts them in ColPartitions.
FindHorizontalTextChains(ColPartitionGrid * part_grid)1457 void StrokeWidth::FindHorizontalTextChains(ColPartitionGrid *part_grid) {
1458 // A PageSegMode that forces horizontal textlines with the current rotation.
1459 PageSegMode pageseg_mode =
1460 rerotation_.y() == 0.0f ? PSM_SINGLE_COLUMN : PSM_SINGLE_BLOCK_VERT_TEXT;
1461 BlobGridSearch gsearch(this);
1462 BLOBNBOX *bbox;
1463 gsearch.StartFullSearch();
1464 while ((bbox = gsearch.NextFullSearch()) != nullptr) {
1465 BLOBNBOX *blob;
1466 if (bbox->owner() == nullptr && bbox->UniquelyHorizontal() &&
1467 (blob = MutualUnusedHNeighbour(bbox, BND_RIGHT)) != nullptr) {
1468 // Put all the linked blobs into a ColPartition.
1469 auto *part = new ColPartition(BRT_TEXT, ICOORD(0, 1));
1470 part->AddBox(bbox);
1471 while (blob != nullptr) {
1472 part->AddBox(blob);
1473 blob = MutualUnusedHNeighbour(blob, BND_RIGHT);
1474 }
1475 blob = MutualUnusedHNeighbour(bbox, BND_LEFT);
1476 while (blob != nullptr) {
1477 part->AddBox(blob);
1478 blob = MutualUnusedVNeighbour(blob, BND_LEFT);
1479 }
1480 CompletePartition(pageseg_mode, part, part_grid);
1481 }
1482 }
1483 }
1484
1485 // Finds diacritics and saves their base character in the blob.
1486 // The objective is to move all diacritics to the noise_blobs list, so
1487 // they don't mess up early textline finding/merging, or force splits
1488 // on textlines that overlap a bit. Blobs that become diacritics must be
1489 // either part of no ColPartition (nullptr owner) or in a small partition in
1490 // which ALL the blobs are diacritics, in which case the partition is
1491 // exploded (deleted) back to its blobs.
TestDiacritics(ColPartitionGrid * part_grid,TO_BLOCK * block)1492 void StrokeWidth::TestDiacritics(ColPartitionGrid *part_grid, TO_BLOCK *block) {
1493 BlobGrid small_grid(gridsize(), bleft(), tright());
1494 small_grid.InsertBlobList(&block->noise_blobs);
1495 small_grid.InsertBlobList(&block->blobs);
1496 int medium_diacritics = 0;
1497 int small_diacritics = 0;
1498 BLOBNBOX_IT small_it(&block->noise_blobs);
1499 for (small_it.mark_cycle_pt(); !small_it.cycled_list(); small_it.forward()) {
1500 BLOBNBOX *blob = small_it.data();
1501 if (blob->owner() == nullptr && !blob->IsDiacritic() && DiacriticBlob(&small_grid, blob)) {
1502 ++small_diacritics;
1503 }
1504 }
1505 BLOBNBOX_IT blob_it(&block->blobs);
1506 for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) {
1507 BLOBNBOX *blob = blob_it.data();
1508 if (blob->IsDiacritic()) {
1509 small_it.add_to_end(blob_it.extract());
1510 continue; // Already a diacritic.
1511 }
1512 ColPartition *part = blob->owner();
1513 if (part == nullptr && DiacriticBlob(&small_grid, blob)) {
1514 ++medium_diacritics;
1515 RemoveBBox(blob);
1516 small_it.add_to_end(blob_it.extract());
1517 } else if (part != nullptr && !part->block_owned() && part->boxes_count() < 3) {
1518 // We allow blobs in small partitions to become diacritics if ALL the
1519 // blobs in the partition qualify as we can then cleanly delete the
1520 // partition, turn all the blobs in it to diacritics and they can be
1521 // merged into the base character partition more easily than merging
1522 // the partitions.
1523 BLOBNBOX_C_IT box_it(part->boxes());
1524 for (box_it.mark_cycle_pt();
1525 !box_it.cycled_list() && DiacriticBlob(&small_grid, box_it.data()); box_it.forward()) {
1526 ;
1527 }
1528 if (box_it.cycled_list()) {
1529 // They are all good.
1530 while (!box_it.empty()) {
1531 // Liberate the blob from its partition so it can be treated
1532 // as a diacritic and merged explicitly with the base part.
1533 // The blob is really owned by the block. The partition "owner"
1534 // is nulled to allow the blob to get merged with its base character
1535 // partition.
1536 BLOBNBOX *box = box_it.extract();
1537 box->set_owner(nullptr);
1538 box_it.forward();
1539 ++medium_diacritics;
1540 // We remove the blob from the grid so it isn't found by subsequent
1541 // searches where we might not want to include diacritics.
1542 RemoveBBox(box);
1543 }
1544 // We only move the one blob to the small list here, but the others
1545 // all get moved by the test at the top of the loop.
1546 small_it.add_to_end(blob_it.extract());
1547 part_grid->RemoveBBox(part);
1548 delete part;
1549 }
1550 } else if (AlignedBlob::WithinTestRegion(2, blob->bounding_box().left(),
1551 blob->bounding_box().bottom())) {
1552 tprintf("Blob not available to be a diacritic at:");
1553 blob->bounding_box().print();
1554 }
1555 }
1556 if (textord_tabfind_show_strokewidths) {
1557 tprintf("Found %d small diacritics, %d medium\n", small_diacritics, medium_diacritics);
1558 }
1559 }
1560
1561 // Searches this grid for an appropriately close and sized neighbour of the
1562 // given [small] blob. If such a blob is found, the diacritic base is saved
1563 // in the blob and true is returned.
1564 // The small_grid is a secondary grid that contains the small/noise objects
1565 // that are not in this grid, but may be useful for determining a connection
1566 // between blob and its potential base character. (See DiacriticXGapFilled.)
DiacriticBlob(BlobGrid * small_grid,BLOBNBOX * blob)1567 bool StrokeWidth::DiacriticBlob(BlobGrid *small_grid, BLOBNBOX *blob) {
1568 if (BLOBNBOX::UnMergeableType(blob->region_type()) || blob->region_type() == BRT_VERT_TEXT) {
1569 return false;
1570 }
1571 TBOX small_box(blob->bounding_box());
1572 bool debug = AlignedBlob::WithinTestRegion(2, small_box.left(), small_box.bottom());
1573 if (debug) {
1574 tprintf("Testing blob for diacriticness at:");
1575 small_box.print();
1576 }
1577 int x = (small_box.left() + small_box.right()) / 2;
1578 int y = (small_box.bottom() + small_box.top()) / 2;
1579 int grid_x, grid_y;
1580 GridCoords(x, y, &grid_x, &grid_y);
1581 int height = small_box.height();
1582 // Setup a rectangle search to find its nearest base-character neighbour.
1583 // We keep 2 different best candidates:
1584 // best_x_overlap is a category of base characters that have an overlap in x
1585 // (like a acute) in which we look for the least y-gap, computed using the
1586 // projection to favor base characters in the same textline.
1587 // best_y_overlap is a category of base characters that have no x overlap,
1588 // (nominally a y-overlap is preferrecd but not essential) in which we
1589 // look for the least weighted sum of x-gap and y-gap, with x-gap getting
1590 // a lower weight to catch quotes at the end of a textline.
1591 // NOTE that x-gap and y-gap are measured from the nearest side of the base
1592 // character to the FARTHEST side of the diacritic to allow small diacritics
1593 // to be a reasonable distance away, but not big diacritics.
1594 BLOBNBOX *best_x_overlap = nullptr;
1595 BLOBNBOX *best_y_overlap = nullptr;
1596 int best_total_dist = 0;
1597 int best_y_gap = 0;
1598 TBOX best_xbox;
1599 // TODO(rays) the search box could be setup using the projection as a guide.
1600 TBOX search_box(small_box);
1601 int x_pad = IntCastRounded(gridsize() * kDiacriticXPadRatio);
1602 int y_pad = IntCastRounded(gridsize() * kDiacriticYPadRatio);
1603 search_box.pad(x_pad, y_pad);
1604 BlobGridSearch rsearch(this);
1605 rsearch.SetUniqueMode(true);
1606 int min_height = height * kMinDiacriticSizeRatio;
1607 rsearch.StartRectSearch(search_box);
1608 BLOBNBOX *neighbour;
1609 while ((neighbour = rsearch.NextRectSearch()) != nullptr) {
1610 if (BLOBNBOX::UnMergeableType(neighbour->region_type()) || neighbour == blob ||
1611 neighbour->owner() == blob->owner()) {
1612 continue;
1613 }
1614 TBOX nbox = neighbour->bounding_box();
1615 if (neighbour->owner() == nullptr || neighbour->owner()->IsVerticalType() ||
1616 (neighbour->flow() != BTFT_CHAIN && neighbour->flow() != BTFT_STRONG_CHAIN)) {
1617 if (debug) {
1618 tprintf("Neighbour not strong enough:");
1619 nbox.print();
1620 }
1621 continue; // Diacritics must be attached to strong text.
1622 }
1623 if (nbox.height() < min_height) {
1624 if (debug) {
1625 tprintf("Neighbour not big enough:");
1626 nbox.print();
1627 }
1628 continue; // Too small to be the base character.
1629 }
1630 int x_gap = small_box.x_gap(nbox);
1631 int y_gap = small_box.y_gap(nbox);
1632 int total_distance = projection_->DistanceOfBoxFromBox(small_box, nbox, true, denorm_, debug);
1633 if (debug) {
1634 tprintf("xgap=%d, y=%d, total dist=%d\n", x_gap, y_gap, total_distance);
1635 }
1636 if (total_distance > neighbour->owner()->median_height() * kMaxDiacriticDistanceRatio) {
1637 if (debug) {
1638 tprintf("Neighbour with median size %d too far away:", neighbour->owner()->median_height());
1639 neighbour->bounding_box().print();
1640 }
1641 continue; // Diacritics must not be too distant.
1642 }
1643 if (x_gap <= 0) {
1644 if (debug) {
1645 tprintf("Computing reduced box for :");
1646 nbox.print();
1647 }
1648 int left = small_box.left() - small_box.width();
1649 int right = small_box.right() + small_box.width();
1650 nbox = neighbour->BoundsWithinLimits(left, right);
1651 y_gap = small_box.y_gap(nbox);
1652 if (best_x_overlap == nullptr || y_gap < best_y_gap) {
1653 best_x_overlap = neighbour;
1654 best_xbox = nbox;
1655 best_y_gap = y_gap;
1656 if (debug) {
1657 tprintf("New best:");
1658 nbox.print();
1659 }
1660 } else if (debug) {
1661 tprintf("Shrunken box doesn't win:");
1662 nbox.print();
1663 }
1664 } else if (blob->ConfirmNoTabViolation(*neighbour)) {
1665 if (best_y_overlap == nullptr || total_distance < best_total_dist) {
1666 if (debug) {
1667 tprintf("New best y overlap:");
1668 nbox.print();
1669 }
1670 best_y_overlap = neighbour;
1671 best_total_dist = total_distance;
1672 } else if (debug) {
1673 tprintf("New y overlap box doesn't win:");
1674 nbox.print();
1675 }
1676 } else if (debug) {
1677 tprintf("Neighbour wrong side of a tab:");
1678 nbox.print();
1679 }
1680 }
1681 if (best_x_overlap != nullptr &&
1682 (best_y_overlap == nullptr || best_xbox.major_y_overlap(best_y_overlap->bounding_box()))) {
1683 blob->set_diacritic_box(best_xbox);
1684 blob->set_base_char_blob(best_x_overlap);
1685 if (debug) {
1686 tprintf("DiacriticBlob OK! (x-overlap:");
1687 small_box.print();
1688 best_xbox.print();
1689 }
1690 return true;
1691 }
1692 if (best_y_overlap != nullptr &&
1693 DiacriticXGapFilled(small_grid, small_box, best_y_overlap->bounding_box()) &&
1694 NoNoiseInBetween(small_box, best_y_overlap->bounding_box())) {
1695 blob->set_diacritic_box(best_y_overlap->bounding_box());
1696 blob->set_base_char_blob(best_y_overlap);
1697 if (debug) {
1698 tprintf("DiacriticBlob OK! (y-overlap:");
1699 small_box.print();
1700 best_y_overlap->bounding_box().print();
1701 }
1702 return true;
1703 }
1704 if (debug) {
1705 tprintf("DiacriticBlob fails:");
1706 small_box.print();
1707 tprintf("Best x+y gap = %d, y = %d\n", best_total_dist, best_y_gap);
1708 if (best_y_overlap != nullptr) {
1709 tprintf("XGapFilled=%d, NoiseBetween=%d\n",
1710 DiacriticXGapFilled(small_grid, small_box, best_y_overlap->bounding_box()),
1711 NoNoiseInBetween(small_box, best_y_overlap->bounding_box()));
1712 }
1713 }
1714 return false;
1715 }
1716
1717 // Returns true if there is no gap between the base char and the diacritic
1718 // bigger than a fraction of the height of the base char:
1719 // Eg: line end.....'
1720 // The quote is a long way from the end of the line, yet it needs to be a
1721 // diacritic. To determine that the quote is not part of an image, or
1722 // a different text block, we check for other marks in the gap between
1723 // the base char and the diacritic.
1724 // '<--Diacritic
1725 // |---------|
1726 // | |<-toobig-gap->
1727 // | Base |<ok gap>
1728 // |---------| x<-----Dot occupying gap
1729 // The grid is const really.
DiacriticXGapFilled(BlobGrid * grid,const TBOX & diacritic_box,const TBOX & base_box)1730 bool StrokeWidth::DiacriticXGapFilled(BlobGrid *grid, const TBOX &diacritic_box,
1731 const TBOX &base_box) {
1732 // Since most gaps are small, use an iterative algorithm to search the gap.
1733 int max_gap = IntCastRounded(base_box.height() * kMaxDiacriticGapToBaseCharHeight);
1734 TBOX occupied_box(base_box);
1735 int diacritic_gap;
1736 while ((diacritic_gap = diacritic_box.x_gap(occupied_box)) > max_gap) {
1737 TBOX search_box(occupied_box);
1738 if (diacritic_box.left() > search_box.right()) {
1739 // We are looking right.
1740 search_box.set_left(search_box.right());
1741 search_box.set_right(search_box.left() + max_gap);
1742 } else {
1743 // We are looking left.
1744 search_box.set_right(search_box.left());
1745 search_box.set_left(search_box.left() - max_gap);
1746 }
1747 BlobGridSearch rsearch(grid);
1748 rsearch.StartRectSearch(search_box);
1749 BLOBNBOX *neighbour;
1750 while ((neighbour = rsearch.NextRectSearch()) != nullptr) {
1751 const TBOX &nbox = neighbour->bounding_box();
1752 if (nbox.x_gap(diacritic_box) < diacritic_gap) {
1753 if (nbox.left() < occupied_box.left()) {
1754 occupied_box.set_left(nbox.left());
1755 }
1756 if (nbox.right() > occupied_box.right()) {
1757 occupied_box.set_right(nbox.right());
1758 }
1759 break;
1760 }
1761 }
1762 if (neighbour == nullptr) {
1763 return false; // Found a big gap.
1764 }
1765 }
1766 return true; // The gap was filled.
1767 }
1768
1769 // Merges diacritics with the ColPartition of the base character blob.
MergeDiacritics(TO_BLOCK * block,ColPartitionGrid * part_grid)1770 void StrokeWidth::MergeDiacritics(TO_BLOCK *block, ColPartitionGrid *part_grid) {
1771 BLOBNBOX_IT small_it(&block->noise_blobs);
1772 for (small_it.mark_cycle_pt(); !small_it.cycled_list(); small_it.forward()) {
1773 BLOBNBOX *blob = small_it.data();
1774 if (blob->base_char_blob() != nullptr) {
1775 ColPartition *part = blob->base_char_blob()->owner();
1776 // The base character must be owned by a partition and that partition
1777 // must not be on the big_parts list (not block owned).
1778 if (part != nullptr && !part->block_owned() && blob->owner() == nullptr &&
1779 blob->IsDiacritic()) {
1780 // The partition has to be removed from the grid and reinserted
1781 // because its bounding box may change.
1782 part_grid->RemoveBBox(part);
1783 part->AddBox(blob);
1784 blob->set_region_type(part->blob_type());
1785 blob->set_flow(part->flow());
1786 blob->set_owner(part);
1787 part_grid->InsertBBox(true, true, part);
1788 }
1789 // Set all base chars to nullptr before any blobs get deleted.
1790 blob->set_base_char_blob(nullptr);
1791 }
1792 }
1793 }
1794
1795 // Any blobs on the large_blobs list of block that are still unowned by a
1796 // ColPartition, are probably drop-cap or vertically touching so the blobs
1797 // are removed to the big_parts list and treated separately.
RemoveLargeUnusedBlobs(TO_BLOCK * block,ColPartitionGrid * part_grid,ColPartition_LIST * big_parts)1798 void StrokeWidth::RemoveLargeUnusedBlobs(TO_BLOCK *block, ColPartitionGrid *part_grid,
1799 ColPartition_LIST *big_parts) {
1800 BLOBNBOX_IT large_it(&block->large_blobs);
1801 for (large_it.mark_cycle_pt(); !large_it.cycled_list(); large_it.forward()) {
1802 BLOBNBOX *blob = large_it.data();
1803 ColPartition *big_part = blob->owner();
1804 if (big_part == nullptr) {
1805 // Large blobs should have gone into partitions by now if they are
1806 // genuine characters, so move any unowned ones out to the big parts
1807 // list. This will include drop caps and vertically touching characters.
1808 ColPartition::MakeBigPartition(blob, big_parts);
1809 }
1810 }
1811 }
1812
1813 // All remaining unused blobs are put in individual ColPartitions.
PartitionRemainingBlobs(PageSegMode pageseg_mode,ColPartitionGrid * part_grid)1814 void StrokeWidth::PartitionRemainingBlobs(PageSegMode pageseg_mode, ColPartitionGrid *part_grid) {
1815 BlobGridSearch gsearch(this);
1816 BLOBNBOX *bbox;
1817 int prev_grid_x = -1;
1818 int prev_grid_y = -1;
1819 BLOBNBOX_CLIST cell_list;
1820 BLOBNBOX_C_IT cell_it(&cell_list);
1821 bool cell_all_noise = true;
1822 gsearch.StartFullSearch();
1823 while ((bbox = gsearch.NextFullSearch()) != nullptr) {
1824 int grid_x = gsearch.GridX();
1825 int grid_y = gsearch.GridY();
1826 if (grid_x != prev_grid_x || grid_y != prev_grid_y) {
1827 // New cell. Process old cell.
1828 MakePartitionsFromCellList(pageseg_mode, cell_all_noise, part_grid, &cell_list);
1829 cell_it.set_to_list(&cell_list);
1830 prev_grid_x = grid_x;
1831 prev_grid_y = grid_y;
1832 cell_all_noise = true;
1833 }
1834 if (bbox->owner() == nullptr) {
1835 cell_it.add_to_end(bbox);
1836 if (bbox->flow() != BTFT_NONTEXT) {
1837 cell_all_noise = false;
1838 }
1839 } else {
1840 cell_all_noise = false;
1841 }
1842 }
1843 MakePartitionsFromCellList(pageseg_mode, cell_all_noise, part_grid, &cell_list);
1844 }
1845
1846 // If combine, put all blobs in the cell_list into a single partition, otherwise
1847 // put each one into its own partition.
MakePartitionsFromCellList(PageSegMode pageseg_mode,bool combine,ColPartitionGrid * part_grid,BLOBNBOX_CLIST * cell_list)1848 void StrokeWidth::MakePartitionsFromCellList(PageSegMode pageseg_mode, bool combine,
1849 ColPartitionGrid *part_grid,
1850 BLOBNBOX_CLIST *cell_list) {
1851 if (cell_list->empty()) {
1852 return;
1853 }
1854 BLOBNBOX_C_IT cell_it(cell_list);
1855 if (combine) {
1856 BLOBNBOX *bbox = cell_it.extract();
1857 auto *part = new ColPartition(bbox->region_type(), ICOORD(0, 1));
1858 part->AddBox(bbox);
1859 part->set_flow(bbox->flow());
1860 for (cell_it.forward(); !cell_it.empty(); cell_it.forward()) {
1861 part->AddBox(cell_it.extract());
1862 }
1863 CompletePartition(pageseg_mode, part, part_grid);
1864 } else {
1865 for (; !cell_it.empty(); cell_it.forward()) {
1866 BLOBNBOX *bbox = cell_it.extract();
1867 auto *part = new ColPartition(bbox->region_type(), ICOORD(0, 1));
1868 part->set_flow(bbox->flow());
1869 part->AddBox(bbox);
1870 CompletePartition(pageseg_mode, part, part_grid);
1871 }
1872 }
1873 }
1874
1875 // Helper function to finish setting up a ColPartition and insert into
1876 // part_grid.
CompletePartition(PageSegMode pageseg_mode,ColPartition * part,ColPartitionGrid * part_grid)1877 void StrokeWidth::CompletePartition(PageSegMode pageseg_mode, ColPartition *part,
1878 ColPartitionGrid *part_grid) {
1879 part->ComputeLimits();
1880 TBOX box = part->bounding_box();
1881 bool debug = AlignedBlob::WithinTestRegion(2, box.left(), box.bottom());
1882 int value = projection_->EvaluateColPartition(*part, denorm_, debug);
1883 // Override value if pageseg_mode disagrees.
1884 if (value > 0 && FindingVerticalOnly(pageseg_mode)) {
1885 value = part->boxes_count() == 1 ? 0 : -2;
1886 } else if (value < 0 && FindingHorizontalOnly(pageseg_mode)) {
1887 value = part->boxes_count() == 1 ? 0 : 2;
1888 }
1889 part->SetRegionAndFlowTypesFromProjectionValue(value);
1890 part->ClaimBoxes();
1891 part_grid->InsertBBox(true, true, part);
1892 }
1893
1894 // Merge partitions where the merge appears harmless.
1895 // As this
EasyMerges(ColPartitionGrid * part_grid)1896 void StrokeWidth::EasyMerges(ColPartitionGrid *part_grid) {
1897 using namespace std::placeholders; // for _1, _2
1898 part_grid->Merges(std::bind(&StrokeWidth::OrientationSearchBox, this, _1, _2),
1899 std::bind(&StrokeWidth::ConfirmEasyMerge, this, _1, _2));
1900 }
1901
1902 // Compute a search box based on the orientation of the partition.
1903 // Returns true if a suitable box can be calculated.
1904 // Callback for EasyMerges.
OrientationSearchBox(ColPartition * part,TBOX * box)1905 bool StrokeWidth::OrientationSearchBox(ColPartition *part, TBOX *box) {
1906 if (part->IsVerticalType()) {
1907 box->set_top(box->top() + box->width());
1908 box->set_bottom(box->bottom() - box->width());
1909 } else {
1910 box->set_left(box->left() - box->height());
1911 box->set_right(box->right() + box->height());
1912 }
1913 return true;
1914 }
1915
1916 // Merge confirmation callback for EasyMerges.
ConfirmEasyMerge(const ColPartition * p1,const ColPartition * p2)1917 bool StrokeWidth::ConfirmEasyMerge(const ColPartition *p1, const ColPartition *p2) {
1918 ASSERT_HOST(p1 != nullptr && p2 != nullptr);
1919 ASSERT_HOST(!p1->IsEmpty() && !p2->IsEmpty());
1920 if ((p1->flow() == BTFT_NONTEXT && p2->flow() >= BTFT_CHAIN) ||
1921 (p1->flow() >= BTFT_CHAIN && p2->flow() == BTFT_NONTEXT)) {
1922 return false; // Don't merge confirmed image with text.
1923 }
1924 if ((p1->IsVerticalType() || p2->IsVerticalType()) && p1->HCoreOverlap(*p2) <= 0 &&
1925 ((!p1->IsSingleton() && !p2->IsSingleton()) ||
1926 !p1->bounding_box().major_overlap(p2->bounding_box()))) {
1927 return false; // Overlap must be in the text line.
1928 }
1929 if ((p1->IsHorizontalType() || p2->IsHorizontalType()) && p1->VCoreOverlap(*p2) <= 0 &&
1930 ((!p1->IsSingleton() && !p2->IsSingleton()) ||
1931 (!p1->bounding_box().major_overlap(p2->bounding_box()) &&
1932 !p1->OKDiacriticMerge(*p2, false) && !p2->OKDiacriticMerge(*p1, false)))) {
1933 return false; // Overlap must be in the text line.
1934 }
1935 if (!p1->ConfirmNoTabViolation(*p2)) {
1936 return false;
1937 }
1938 if (p1->flow() <= BTFT_NONTEXT && p2->flow() <= BTFT_NONTEXT) {
1939 return true;
1940 }
1941 return NoNoiseInBetween(p1->bounding_box(), p2->bounding_box());
1942 }
1943
1944 // Returns true if there is no significant noise in between the boxes.
NoNoiseInBetween(const TBOX & box1,const TBOX & box2) const1945 bool StrokeWidth::NoNoiseInBetween(const TBOX &box1, const TBOX &box2) const {
1946 return ImageFind::BlankImageInBetween(box1, box2, grid_box_, rerotation_, nontext_map_);
1947 }
1948
1949 #ifndef GRAPHICS_DISABLED
1950
1951 /** Displays the blobs colored according to the number of good neighbours
1952 * and the vertical/horizontal flow.
1953 */
DisplayGoodBlobs(const char * window_name,int x,int y)1954 ScrollView *StrokeWidth::DisplayGoodBlobs(const char *window_name, int x, int y) {
1955 auto window = MakeWindow(x, y, window_name);
1956 // For every blob in the grid, display it.
1957 window->Brush(ScrollView::NONE);
1958
1959 // For every bbox in the grid, display it.
1960 BlobGridSearch gsearch(this);
1961 gsearch.StartFullSearch();
1962 BLOBNBOX *bbox;
1963 while ((bbox = gsearch.NextFullSearch()) != nullptr) {
1964 const TBOX &box = bbox->bounding_box();
1965 int left_x = box.left();
1966 int right_x = box.right();
1967 int top_y = box.top();
1968 int bottom_y = box.bottom();
1969 int goodness = bbox->GoodTextBlob();
1970 BlobRegionType blob_type = bbox->region_type();
1971 if (bbox->UniquelyVertical()) {
1972 blob_type = BRT_VERT_TEXT;
1973 }
1974 if (bbox->UniquelyHorizontal()) {
1975 blob_type = BRT_TEXT;
1976 }
1977 BlobTextFlowType flow = bbox->flow();
1978 if (flow == BTFT_NONE) {
1979 if (goodness == 0) {
1980 flow = BTFT_NEIGHBOURS;
1981 } else if (goodness == 1) {
1982 flow = BTFT_CHAIN;
1983 } else {
1984 flow = BTFT_STRONG_CHAIN;
1985 }
1986 }
1987 window->Pen(BLOBNBOX::TextlineColor(blob_type, flow));
1988 window->Rectangle(left_x, bottom_y, right_x, top_y);
1989 }
1990 window->Update();
1991 return window;
1992 }
1993
DrawDiacriticJoiner(const BLOBNBOX * blob,ScrollView * window)1994 static void DrawDiacriticJoiner(const BLOBNBOX *blob, ScrollView *window) {
1995 const TBOX &blob_box(blob->bounding_box());
1996 int top = std::max(static_cast<int>(blob_box.top()), blob->base_char_top());
1997 int bottom = std::min(static_cast<int>(blob_box.bottom()), blob->base_char_bottom());
1998 int x = (blob_box.left() + blob_box.right()) / 2;
1999 window->Line(x, top, x, bottom);
2000 }
2001
2002 // Displays blobs colored according to whether or not they are diacritics.
DisplayDiacritics(const char * window_name,int x,int y,TO_BLOCK * block)2003 ScrollView *StrokeWidth::DisplayDiacritics(const char *window_name, int x, int y, TO_BLOCK *block) {
2004 auto window = MakeWindow(x, y, window_name);
2005 // For every blob in the grid, display it.
2006 window->Brush(ScrollView::NONE);
2007
2008 BLOBNBOX_IT it(&block->blobs);
2009 for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) {
2010 BLOBNBOX *blob = it.data();
2011 if (blob->IsDiacritic()) {
2012 window->Pen(ScrollView::GREEN);
2013 DrawDiacriticJoiner(blob, window);
2014 } else {
2015 window->Pen(blob->BoxColor());
2016 }
2017 const TBOX &box = blob->bounding_box();
2018 window->Rectangle(box.left(), box.bottom(), box.right(), box.top());
2019 }
2020 it.set_to_list(&block->noise_blobs);
2021 for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) {
2022 BLOBNBOX *blob = it.data();
2023 if (blob->IsDiacritic()) {
2024 window->Pen(ScrollView::GREEN);
2025 DrawDiacriticJoiner(blob, window);
2026 } else {
2027 window->Pen(ScrollView::WHITE);
2028 }
2029 const TBOX &box = blob->bounding_box();
2030 window->Rectangle(box.left(), box.bottom(), box.right(), box.top());
2031 }
2032 window->Update();
2033 return window;
2034 }
2035
2036 #endif // !GRAPHICS_DISABLED
2037
2038 } // namespace tesseract.
2039