Tesseract
3.02
|
00001 00002 // File: TabFind.cpp 00003 // Description: Subclass of BBGrid to find vertically aligned blobs. 00004 // Author: Ray Smith 00005 // Created: Fri Mar 21 15:03:01 PST 2008 00006 // 00007 // (C) Copyright 2008, Google Inc. 00008 // Licensed under the Apache License, Version 2.0 (the "License"); 00009 // you may not use this file except in compliance with the License. 00010 // You may obtain a copy of the License at 00011 // http://www.apache.org/licenses/LICENSE-2.0 00012 // Unless required by applicable law or agreed to in writing, software 00013 // distributed under the License is distributed on an "AS IS" BASIS, 00014 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 00015 // See the License for the specific language governing permissions and 00016 // limitations under the License. 00017 // 00019 00020 #include "tabfind.h" 00021 #include "alignedblob.h" 00022 #include "blobbox.h" 00023 #include "colpartitiongrid.h" 00024 #include "detlinefit.h" 00025 #include "linefind.h" 00026 #include "ndminx.h" 00027 00028 // Include automatically generated configuration file if running autoconf. 00029 #ifdef HAVE_CONFIG_H 00030 #include "config_auto.h" 00031 #endif 00032 00033 namespace tesseract { 00034 00035 // Multiple of box size to search for initial gaps. 00036 const int kTabRadiusFactor = 5; 00037 // Min and Max multiple of height to search vertically when extrapolating. 00038 const int kMinVerticalSearch = 3; 00039 const int kMaxVerticalSearch = 12; 00040 const int kMaxRaggedSearch = 25; 00041 // Minimum number of lines in a column width to make it interesting. 00042 const int kMinLinesInColumn = 10; 00043 // Minimum width of a column to be interesting. 00044 const int kMinColumnWidth = 200; 00045 // Minimum fraction of total column lines for a column to be interesting. 00046 const double kMinFractionalLinesInColumn = 0.125; 00047 // Fraction of height used as alignment tolerance for aligned tabs. 00048 const double kAlignedFraction = 0.03125; 00049 // Minimum gutter width in absolute inch (multiplied by resolution) 00050 const double kMinGutterWidthAbsolute = 0.02; 00051 // Maximum gutter width (in absolute inch) that we care about 00052 const double kMaxGutterWidthAbsolute = 2.00; 00053 // Multiplier of gridsize for min gutter width of TT_MAYBE_RAGGED blobs. 00054 const int kRaggedGutterMultiple = 5; 00055 // Min aspect ratio of tall objects to be considered a separator line. 00056 // (These will be ignored in searching the gutter for obstructions.) 00057 const double kLineFragmentAspectRatio = 10.0; 00058 // Multiplier of new y positions in running average for skew estimation. 00059 const double kSmoothFactor = 0.25; 00060 // Min coverage for a good baseline between vectors 00061 const double kMinBaselineCoverage = 0.5; 00062 // Minimum overlap fraction when scanning text lines for column widths. 00063 const double kCharVerticalOverlapFraction = 0.375; 00064 // Maximum horizontal gap allowed when scanning for column widths 00065 const double kMaxHorizontalGap = 3.0; 00066 // Maximum upper quartile error allowed on a baseline fit as a fraction 00067 // of height. 00068 const double kMaxBaselineError = 0.4375; 00069 // Min number of points to accept after evaluation. 00070 const int kMinEvaluatedTabs = 3; 00071 // Minimum aspect ratio of a textline to make a good textline blob with a 00072 // single blob. 00073 const int kMaxTextLineBlobRatio = 5; 00074 // Minimum aspect ratio of a textline to make a good textline blob with 00075 // multiple blobs. Target ratio varies according to number of blobs. 00076 const int kMinTextLineBlobRatio = 3; 00077 // Fraction of box area covered by image to make a blob image. 00078 const double kMinImageArea = 0.5; 00079 // Upto 30 degrees is allowed for rotations of diacritic blobs. 00080 // Keep this value slightly larger than kCosSmallAngle in blobbox.cpp 00081 // so that the assert there never fails. 00082 const double kCosMaxSkewAngle = 0.866025; 00083 00084 BOOL_VAR(textord_tabfind_show_initialtabs, false, "Show tab candidates"); 00085 BOOL_VAR(textord_tabfind_show_finaltabs, false, "Show tab vectors"); 00086 double_VAR(textord_tabfind_aligned_gap_fraction, 0.75, 00087 "Fraction of height used as a minimum gap for aligned blobs."); 00088 00089 TabFind::TabFind(int gridsize, const ICOORD& bleft, const ICOORD& tright, 00090 TabVector_LIST* vlines, int vertical_x, int vertical_y, 00091 int resolution) 00092 : AlignedBlob(gridsize, bleft, tright), 00093 resolution_(resolution), 00094 image_origin_(0, tright.y() - 1) { 00095 width_cb_ = NULL; 00096 v_it_.set_to_list(&vectors_); 00097 v_it_.add_list_after(vlines); 00098 SetVerticalSkewAndParellelize(vertical_x, vertical_y); 00099 width_cb_ = NewPermanentTessCallback(this, &TabFind::CommonWidth); 00100 } 00101 00102 TabFind::~TabFind() { 00103 if (width_cb_ != NULL) 00104 delete width_cb_; 00105 } 00106 00108 00109 // Insert a list of blobs into the given grid (not necessarily this). 00110 // If take_ownership is true, then the blobs are removed from the source list. 00111 // See InsertBlob for the other arguments. 00112 // It would seem to make more sense to swap this and grid, but this way 00113 // around allows grid to not be derived from TabFind, eg a ColPartitionGrid, 00114 // while the grid that provides the tab stops(this) has to be derived from 00115 // TabFind. 00116 void TabFind::InsertBlobsToGrid(bool h_spread, bool v_spread, 00117 BLOBNBOX_LIST* blobs, 00118 BBGrid<BLOBNBOX, BLOBNBOX_CLIST, 00119 BLOBNBOX_C_IT>* grid) { 00120 BLOBNBOX_IT blob_it(blobs); 00121 int b_count = 0; 00122 int reject_count = 0; 00123 for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) { 00124 BLOBNBOX* blob = blob_it.data(); 00125 // if (InsertBlob(true, true, blob, grid)) { 00126 if (InsertBlob(h_spread, v_spread, blob, grid)) { 00127 ++b_count; 00128 } else { 00129 ++reject_count; 00130 } 00131 } 00132 if (textord_debug_tabfind) { 00133 tprintf("Inserted %d blobs into grid, %d rejected.\n", 00134 b_count, reject_count); 00135 } 00136 } 00137 00138 // Insert a single blob into the given grid (not necessarily this). 00139 // If h_spread, then all cells covered horizontally by the box are 00140 // used, otherwise, just the bottom-left. Similarly for v_spread. 00141 // A side effect is that the left and right rule edges of the blob are 00142 // set according to the tab vectors in this (not grid). 00143 bool TabFind::InsertBlob(bool h_spread, bool v_spread, BLOBNBOX* blob, 00144 BBGrid<BLOBNBOX, BLOBNBOX_CLIST, 00145 BLOBNBOX_C_IT>* grid) { 00146 TBOX box = blob->bounding_box(); 00147 blob->set_left_rule(LeftEdgeForBox(box, false, false)); 00148 blob->set_right_rule(RightEdgeForBox(box, false, false)); 00149 blob->set_left_crossing_rule(LeftEdgeForBox(box, true, false)); 00150 blob->set_right_crossing_rule(RightEdgeForBox(box, true, false)); 00151 if (blob->joined_to_prev()) 00152 return false; 00153 grid->InsertBBox(h_spread, v_spread, blob); 00154 return true; 00155 } 00156 00157 // Calls SetBlobRuleEdges for all the blobs in the given block. 00158 void TabFind::SetBlockRuleEdges(TO_BLOCK* block) { 00159 SetBlobRuleEdges(&block->blobs); 00160 SetBlobRuleEdges(&block->small_blobs); 00161 SetBlobRuleEdges(&block->noise_blobs); 00162 SetBlobRuleEdges(&block->large_blobs); 00163 } 00164 00165 // Sets the left and right rule and crossing_rules for the blobs in the given 00166 // list by fiding the next outermost tabvectors for each blob. 00167 void TabFind::SetBlobRuleEdges(BLOBNBOX_LIST* blobs) { 00168 BLOBNBOX_IT blob_it(blobs); 00169 for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) { 00170 BLOBNBOX* blob = blob_it.data(); 00171 TBOX box = blob->bounding_box(); 00172 blob->set_left_rule(LeftEdgeForBox(box, false, false)); 00173 blob->set_right_rule(RightEdgeForBox(box, false, false)); 00174 blob->set_left_crossing_rule(LeftEdgeForBox(box, true, false)); 00175 blob->set_right_crossing_rule(RightEdgeForBox(box, true, false)); 00176 } 00177 } 00178 00179 // Returns the gutter width of the given TabVector between the given y limits. 00180 // Also returns x-shift to be added to the vector to clear any intersecting 00181 // blobs. The shift is deducted from the returned gutter. 00182 // If ignore_unmergeables is true, then blobs of UnMergeableType are 00183 // ignored as if they don't exist. (Used for text on image.) 00184 // max_gutter_width is used as the maximum width worth searching for in case 00185 // there is nothing near the TabVector. 00186 int TabFind::GutterWidth(int bottom_y, int top_y, const TabVector& v, 00187 bool ignore_unmergeables, int max_gutter_width, 00188 int* required_shift) { 00189 bool right_to_left = v.IsLeftTab(); 00190 int bottom_x = v.XAtY(bottom_y); 00191 int top_x = v.XAtY(top_y); 00192 int start_x = right_to_left ? MAX(top_x, bottom_x) : MIN(top_x, bottom_x); 00193 BlobGridSearch sidesearch(this); 00194 sidesearch.StartSideSearch(start_x, bottom_y, top_y); 00195 int min_gap = max_gutter_width; 00196 *required_shift = 0; 00197 BLOBNBOX* blob = NULL; 00198 while ((blob = sidesearch.NextSideSearch(right_to_left)) != NULL) { 00199 const TBOX& box = blob->bounding_box(); 00200 if (box.bottom() >= top_y || box.top() <= bottom_y) 00201 continue; // Doesn't overlap enough. 00202 if (box.height() >= gridsize() * 2 && 00203 box.height() > box.width() * kLineFragmentAspectRatio) { 00204 // Skip likely separator line residue. 00205 continue; 00206 } 00207 if (ignore_unmergeables && BLOBNBOX::UnMergeableType(blob->region_type())) 00208 continue; // Skip non-text if required. 00209 int mid_y = (box.bottom() + box.top()) / 2; 00210 // We use the x at the mid-y so that the required_shift guarantees 00211 // to clear all the blobs on the tab-stop. If we use the min/max 00212 // of x at top/bottom of the blob, then exactness would be required, 00213 // which is not a good thing. 00214 int tab_x = v.XAtY(mid_y); 00215 int gap; 00216 if (right_to_left) { 00217 gap = tab_x - box.right(); 00218 if (gap < 0 && box.left() - tab_x < *required_shift) 00219 *required_shift = box.left() - tab_x; 00220 } else { 00221 gap = box.left() - tab_x; 00222 if (gap < 0 && box.right() - tab_x > *required_shift) 00223 *required_shift = box.right() - tab_x; 00224 } 00225 if (gap > 0 && gap < min_gap) 00226 min_gap = gap; 00227 } 00228 // Result may be negative, in which case, this is a really bad tabstop. 00229 return min_gap - abs(*required_shift); 00230 } 00231 00232 // Find the gutter width and distance to inner neighbour for the given blob. 00233 void TabFind::GutterWidthAndNeighbourGap(int tab_x, int mean_height, 00234 int max_gutter, bool left, 00235 BLOBNBOX* bbox, int* gutter_width, 00236 int* neighbour_gap ) { 00237 const TBOX& box = bbox->bounding_box(); 00238 // The gutter and internal sides of the box. 00239 int gutter_x = left ? box.left() : box.right(); 00240 int internal_x = left ? box.right() : box.left(); 00241 // On ragged edges, the gutter side of the box is away from the tabstop. 00242 int tab_gap = left ? gutter_x - tab_x : tab_x - gutter_x; 00243 *gutter_width = max_gutter; 00244 // If the box is away from the tabstop, we need to increase 00245 // the allowed gutter width. 00246 if (tab_gap > 0) 00247 *gutter_width += tab_gap; 00248 bool debug = WithinTestRegion(2, box.left(), box.bottom()); 00249 if (debug) 00250 tprintf("Looking in gutter\n"); 00251 // Find the nearest blob on the outside of the column. 00252 BLOBNBOX* gutter_bbox = AdjacentBlob(bbox, left, 00253 bbox->flow() == BTFT_TEXT_ON_IMAGE, 0.0, 00254 *gutter_width, box.top(), box.bottom()); 00255 if (gutter_bbox != NULL) { 00256 TBOX gutter_box = gutter_bbox->bounding_box(); 00257 *gutter_width = left ? tab_x - gutter_box.right() 00258 : gutter_box.left() - tab_x; 00259 } 00260 if (*gutter_width >= max_gutter) { 00261 // If there is no box because a tab was in the way, get the tab coord. 00262 TBOX gutter_box(box); 00263 if (left) { 00264 gutter_box.set_left(tab_x - max_gutter - 1); 00265 gutter_box.set_right(tab_x - max_gutter); 00266 int tab_gutter = RightEdgeForBox(gutter_box, true, false); 00267 if (tab_gutter < tab_x - 1) 00268 *gutter_width = tab_x - tab_gutter; 00269 } else { 00270 gutter_box.set_left(tab_x + max_gutter); 00271 gutter_box.set_right(tab_x + max_gutter + 1); 00272 int tab_gutter = LeftEdgeForBox(gutter_box, true, false); 00273 if (tab_gutter > tab_x + 1) 00274 *gutter_width = tab_gutter - tab_x; 00275 } 00276 } 00277 if (*gutter_width > max_gutter) 00278 *gutter_width = max_gutter; 00279 // Now look for a neighbour on the inside. 00280 if (debug) 00281 tprintf("Looking for neighbour\n"); 00282 BLOBNBOX* neighbour = AdjacentBlob(bbox, !left, 00283 bbox->flow() == BTFT_TEXT_ON_IMAGE, 0.0, 00284 *gutter_width, box.top(), box.bottom()); 00285 int neighbour_edge = left ? RightEdgeForBox(box, true, false) 00286 : LeftEdgeForBox(box, true, false); 00287 if (neighbour != NULL) { 00288 TBOX n_box = neighbour->bounding_box(); 00289 if (debug) { 00290 tprintf("Found neighbour:"); 00291 n_box.print(); 00292 } 00293 if (left && n_box.left() < neighbour_edge) 00294 neighbour_edge = n_box.left(); 00295 else if (!left && n_box.right() > neighbour_edge) 00296 neighbour_edge = n_box.right(); 00297 } 00298 *neighbour_gap = left ? neighbour_edge - internal_x 00299 : internal_x - neighbour_edge; 00300 } 00301 00302 // Return the x-coord that corresponds to the right edge for the given 00303 // box. If there is a rule line to the right that vertically overlaps it, 00304 // then return the x-coord of the rule line, otherwise return the right 00305 // edge of the page. For details see RightTabForBox below. 00306 int TabFind::RightEdgeForBox(const TBOX& box, bool crossing, bool extended) { 00307 TabVector* v = RightTabForBox(box, crossing, extended); 00308 return v == NULL ? tright_.x() : v->XAtY((box.top() + box.bottom()) / 2); 00309 } 00310 // As RightEdgeForBox, but finds the left Edge instead. 00311 int TabFind::LeftEdgeForBox(const TBOX& box, bool crossing, bool extended) { 00312 TabVector* v = LeftTabForBox(box, crossing, extended); 00313 return v == NULL ? bleft_.x() : v->XAtY((box.top() + box.bottom()) / 2); 00314 } 00315 00316 // This comment documents how this function works. 00317 // For its purpose and arguments, see the comment in tabfind.h. 00318 // TabVectors are stored sorted by perpendicular distance of middle from 00319 // the global mean vertical vector. Since the individual vectors can have 00320 // differing directions, their XAtY for a given y is not necessarily in the 00321 // right order. Therefore the search has to be run with a margin. 00322 // The middle of a vector that passes through (x,y) cannot be higher than 00323 // halfway from y to the top, or lower than halfway from y to the bottom 00324 // of the coordinate range; therefore, the search margin is the range of 00325 // sort keys between these halfway points. Any vector with a sort key greater 00326 // than the upper margin must be to the right of x at y, and likewise any 00327 // vector with a sort key less than the lower margin must pass to the left 00328 // of x at y. 00329 TabVector* TabFind::RightTabForBox(const TBOX& box, bool crossing, 00330 bool extended) { 00331 if (v_it_.empty()) 00332 return NULL; 00333 int top_y = box.top(); 00334 int bottom_y = box.bottom(); 00335 int mid_y = (top_y + bottom_y) / 2; 00336 int right = crossing ? (box.left() + box.right()) / 2 : box.right(); 00337 int min_key, max_key; 00338 SetupTabSearch(right, mid_y, &min_key, &max_key); 00339 // Position the iterator at the first TabVector with sort_key >= min_key. 00340 while (!v_it_.at_first() && v_it_.data()->sort_key() >= min_key) 00341 v_it_.backward(); 00342 while (!v_it_.at_last() && v_it_.data()->sort_key() < min_key) 00343 v_it_.forward(); 00344 // Find the leftmost tab vector that overlaps and has XAtY(mid_y) >= right. 00345 TabVector* best_v = NULL; 00346 int best_x = -1; 00347 int key_limit = -1; 00348 do { 00349 TabVector* v = v_it_.data(); 00350 int x = v->XAtY(mid_y); 00351 if (x >= right && 00352 (v->VOverlap(top_y, bottom_y) > 0 || 00353 (extended && v->ExtendedOverlap(top_y, bottom_y) > 0))) { 00354 if (best_v == NULL || x < best_x) { 00355 best_v = v; 00356 best_x = x; 00357 // We can guarantee that no better vector can be found if the 00358 // sort key exceeds that of the best by max_key - min_key. 00359 key_limit = v->sort_key() + max_key - min_key; 00360 } 00361 } 00362 // Break when the search is done to avoid wrapping the iterator and 00363 // thereby potentially slowing the next search. 00364 if (v_it_.at_last() || 00365 (best_v != NULL && v->sort_key() > key_limit)) 00366 break; // Prevent restarting list for next call. 00367 v_it_.forward(); 00368 } while (!v_it_.at_first()); 00369 return best_v; 00370 } 00371 00372 // As RightTabForBox, but finds the left TabVector instead. 00373 TabVector* TabFind::LeftTabForBox(const TBOX& box, bool crossing, 00374 bool extended) { 00375 if (v_it_.empty()) 00376 return NULL; 00377 int top_y = box.top(); 00378 int bottom_y = box.bottom(); 00379 int mid_y = (top_y + bottom_y) / 2; 00380 int left = crossing ? (box.left() + box.right()) / 2 : box.left(); 00381 int min_key, max_key; 00382 SetupTabSearch(left, mid_y, &min_key, &max_key); 00383 // Position the iterator at the last TabVector with sort_key <= max_key. 00384 while (!v_it_.at_last() && v_it_.data()->sort_key() <= max_key) 00385 v_it_.forward(); 00386 while (!v_it_.at_first() && v_it_.data()->sort_key() > max_key) { 00387 v_it_.backward(); 00388 } 00389 // Find the rightmost tab vector that overlaps and has XAtY(mid_y) <= left. 00390 TabVector* best_v = NULL; 00391 int best_x = -1; 00392 int key_limit = -1; 00393 do { 00394 TabVector* v = v_it_.data(); 00395 int x = v->XAtY(mid_y); 00396 if (x <= left && 00397 (v->VOverlap(top_y, bottom_y) > 0 || 00398 (extended && v->ExtendedOverlap(top_y, bottom_y) > 0))) { 00399 if (best_v == NULL || x > best_x) { 00400 best_v = v; 00401 best_x = x; 00402 // We can guarantee that no better vector can be found if the 00403 // sort key is less than that of the best by max_key - min_key. 00404 key_limit = v->sort_key() - (max_key - min_key); 00405 } 00406 } 00407 // Break when the search is done to avoid wrapping the iterator and 00408 // thereby potentially slowing the next search. 00409 if (v_it_.at_first() || 00410 (best_v != NULL && v->sort_key() < key_limit)) 00411 break; // Prevent restarting list for next call. 00412 v_it_.backward(); 00413 } while (!v_it_.at_last()); 00414 return best_v; 00415 } 00416 00417 // Return true if the given width is close to one of the common 00418 // widths in column_widths_. 00419 bool TabFind::CommonWidth(int width) { 00420 width /= kColumnWidthFactor; 00421 ICOORDELT_IT it(&column_widths_); 00422 for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) { 00423 ICOORDELT* w = it.data(); 00424 if (NearlyEqual<int>(width, w->x(), 1)) 00425 return true; 00426 } 00427 return false; 00428 } 00429 00430 // Return true if the sizes are more than a 00431 // factor of 2 different. 00432 bool TabFind::DifferentSizes(int size1, int size2) { 00433 return size1 > size2 * 2 || size2 > size1 * 2; 00434 } 00435 00436 // Return true if the sizes are more than a 00437 // factor of 5 different. 00438 bool TabFind::VeryDifferentSizes(int size1, int size2) { 00439 return size1 > size2 * 5 || size2 > size1 * 5; 00440 } 00441 00443 00444 // Top-level function to find TabVectors in an input page block. 00445 // Returns false if the detected skew angle is impossible. 00446 // Applies the detected skew angle to deskew the tabs, blobs and part_grid. 00447 bool TabFind::FindTabVectors(TabVector_LIST* hlines, 00448 BLOBNBOX_LIST* image_blobs, TO_BLOCK* block, 00449 int min_gutter_width, 00450 ColPartitionGrid* part_grid, 00451 FCOORD* deskew, FCOORD* reskew) { 00452 ScrollView* tab_win = FindInitialTabVectors(image_blobs, min_gutter_width, 00453 block); 00454 ComputeColumnWidths(tab_win, part_grid); 00455 TabVector::MergeSimilarTabVectors(vertical_skew_, &vectors_, this); 00456 SortVectors(); 00457 CleanupTabs(); 00458 if (!Deskew(hlines, image_blobs, block, deskew, reskew)) 00459 return false; // Skew angle is too large. 00460 part_grid->Deskew(*deskew); 00461 ApplyTabConstraints(); 00462 #ifndef GRAPHICS_DISABLED 00463 if (textord_tabfind_show_finaltabs) { 00464 tab_win = MakeWindow(640, 50, "FinalTabs"); 00465 if (textord_debug_images) { 00466 tab_win->Image(AlignedBlob::textord_debug_pix().string(), 00467 image_origin_.x(), image_origin_.y()); 00468 } else { 00469 DisplayBoxes(tab_win); 00470 DisplayTabs("FinalTabs", tab_win); 00471 } 00472 tab_win = DisplayTabVectors(tab_win); 00473 } 00474 #endif // GRAPHICS_DISABLED 00475 return true; 00476 } 00477 00478 // Top-level function to not find TabVectors in an input page block, 00479 // but setup for single column mode. 00480 void TabFind::DontFindTabVectors(BLOBNBOX_LIST* image_blobs, TO_BLOCK* block, 00481 FCOORD* deskew, FCOORD* reskew) { 00482 InsertBlobsToGrid(false, false, image_blobs, this); 00483 InsertBlobsToGrid(true, false, &block->blobs, this); 00484 deskew->set_x(1.0f); 00485 deskew->set_y(0.0f); 00486 reskew->set_x(1.0f); 00487 reskew->set_y(0.0f); 00488 } 00489 00490 // Cleans up the lists of blobs in the block ready for use by TabFind. 00491 // Large blobs that look like text are moved to the main blobs list. 00492 // Main blobs that are superseded by the image blobs are deleted. 00493 void TabFind::TidyBlobs(TO_BLOCK* block) { 00494 BLOBNBOX_IT large_it = &block->large_blobs; 00495 BLOBNBOX_IT blob_it = &block->blobs; 00496 int b_count = 0; 00497 for (large_it.mark_cycle_pt(); !large_it.cycled_list(); large_it.forward()) { 00498 BLOBNBOX* large_blob = large_it.data(); 00499 if (large_blob->owner() != NULL) { 00500 blob_it.add_to_end(large_it.extract()); 00501 ++b_count; 00502 } 00503 } 00504 if (textord_debug_tabfind) { 00505 tprintf("Moved %d large blobs to normal list\n", 00506 b_count); 00507 #ifndef GRAPHICS_DISABLED 00508 ScrollView* rej_win = MakeWindow(500, 300, "Image blobs"); 00509 block->plot_graded_blobs(rej_win); 00510 block->plot_noise_blobs(rej_win); 00511 rej_win->Update(); 00512 #endif // GRAPHICS_DISABLED 00513 } 00514 block->DeleteUnownedNoise(); 00515 } 00516 00517 // Helper function to setup search limits for *TabForBox. 00518 void TabFind::SetupTabSearch(int x, int y, int* min_key, int* max_key) { 00519 int key1 = TabVector::SortKey(vertical_skew_, x, (y + tright_.y()) / 2); 00520 int key2 = TabVector::SortKey(vertical_skew_, x, (y + bleft_.y()) / 2); 00521 *min_key = MIN(key1, key2); 00522 *max_key = MAX(key1, key2); 00523 } 00524 00525 ScrollView* TabFind::DisplayTabVectors(ScrollView* tab_win) { 00526 #ifndef GRAPHICS_DISABLED 00527 // For every vector, display it. 00528 TabVector_IT it(&vectors_); 00529 for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) { 00530 TabVector* vector = it.data(); 00531 vector->Display(tab_win); 00532 } 00533 tab_win->Update(); 00534 #endif 00535 return tab_win; 00536 } 00537 00538 // PRIVATE CODE. 00539 // 00540 // First part of FindTabVectors, which may be used twice if the text 00541 // is mostly of vertical alignment. 00542 ScrollView* TabFind::FindInitialTabVectors(BLOBNBOX_LIST* image_blobs, 00543 int min_gutter_width, 00544 TO_BLOCK* block) { 00545 if (textord_tabfind_show_initialtabs) { 00546 ScrollView* line_win = MakeWindow(0, 0, "VerticalLines"); 00547 line_win = DisplayTabVectors(line_win); 00548 } 00549 // Prepare the grid. 00550 if (image_blobs != NULL) 00551 InsertBlobsToGrid(true, false, image_blobs, this); 00552 InsertBlobsToGrid(true, false, &block->blobs, this); 00553 ScrollView* initial_win = FindTabBoxes(min_gutter_width); 00554 FindAllTabVectors(min_gutter_width); 00555 00556 TabVector::MergeSimilarTabVectors(vertical_skew_, &vectors_, this); 00557 SortVectors(); 00558 EvaluateTabs(); 00559 if (textord_tabfind_show_initialtabs && initial_win != NULL) 00560 initial_win = DisplayTabVectors(initial_win); 00561 MarkVerticalText(); 00562 return initial_win; 00563 } 00564 00565 // Helper displays all the boxes in the given vector on the given window. 00566 static void DisplayBoxVector(const GenericVector<BLOBNBOX*> boxes, 00567 ScrollView* win) { 00568 #ifndef GRAPHICS_DISABLED 00569 for (int i = 0; i < boxes.size(); ++i) { 00570 TBOX box = boxes[i]->bounding_box(); 00571 int left_x = box.left(); 00572 int right_x = box.right(); 00573 int top_y = box.top(); 00574 int bottom_y = box.bottom(); 00575 ScrollView::Color box_color = boxes[i]->BoxColor(); 00576 win->Pen(box_color); 00577 win->Rectangle(left_x, bottom_y, right_x, top_y); 00578 } 00579 win->Update(); 00580 #endif // GRAPHICS_DISABLED 00581 } 00582 00583 // For each box in the grid, decide whether it is a candidate tab-stop, 00584 // and if so add it to the left/right tab boxes. 00585 ScrollView* TabFind::FindTabBoxes(int min_gutter_width) { 00586 left_tab_boxes_.clear(); 00587 right_tab_boxes_.clear(); 00588 // For every bbox in the grid, determine whether it uses a tab on an edge. 00589 GridSearch<BLOBNBOX, BLOBNBOX_CLIST, BLOBNBOX_C_IT> gsearch(this); 00590 gsearch.StartFullSearch(); 00591 BLOBNBOX* bbox; 00592 while ((bbox = gsearch.NextFullSearch()) != NULL) { 00593 if (TestBoxForTabs(bbox, min_gutter_width)) { 00594 // If it is any kind of tab, insert it into the vectors. 00595 if (bbox->left_tab_type() != TT_NONE) 00596 left_tab_boxes_.push_back(bbox); 00597 if (bbox->right_tab_type() != TT_NONE) 00598 right_tab_boxes_.push_back(bbox); 00599 } 00600 } 00601 // Sort left tabs by left and right by right to see the outermost one first 00602 // on a ragged tab. 00603 left_tab_boxes_.sort(SortByBoxLeft<BLOBNBOX>); 00604 right_tab_boxes_.sort(SortRightToLeft<BLOBNBOX>); 00605 ScrollView* tab_win = NULL; 00606 #ifndef GRAPHICS_DISABLED 00607 if (textord_tabfind_show_initialtabs) { 00608 tab_win = MakeWindow(0, 100, "InitialTabs"); 00609 tab_win->Pen(ScrollView::BLUE); 00610 tab_win->Brush(ScrollView::NONE); 00611 // Display the left and right tab boxes. 00612 DisplayBoxVector(left_tab_boxes_, tab_win); 00613 DisplayBoxVector(right_tab_boxes_, tab_win); 00614 tab_win = DisplayTabs("Tabs", tab_win); 00615 } 00616 #endif // GRAPHICS_DISABLED 00617 return tab_win; 00618 } 00619 00620 bool TabFind::TestBoxForTabs(BLOBNBOX* bbox, int min_gutter_width) { 00621 GridSearch<BLOBNBOX, BLOBNBOX_CLIST, BLOBNBOX_C_IT> radsearch(this); 00622 TBOX box = bbox->bounding_box(); 00623 // If there are separator lines, get the column edges. 00624 int left_column_edge = bbox->left_rule(); 00625 int right_column_edge = bbox->right_rule(); 00626 // The edges of the bounding box of the blob being processed. 00627 int left_x = box.left(); 00628 int right_x = box.right(); 00629 int top_y = box.top(); 00630 int bottom_y = box.bottom(); 00631 int height = box.height(); 00632 bool debug = WithinTestRegion(3, left_x, top_y); 00633 if (debug) { 00634 tprintf("Column edges for blob at (%d,%d)->(%d,%d) are [%d, %d]\n", 00635 left_x, top_y, right_x, bottom_y, 00636 left_column_edge, right_column_edge); 00637 } 00638 // Compute a search radius based on a multiple of the height. 00639 int radius = (height * kTabRadiusFactor + gridsize_ - 1) / gridsize_; 00640 radsearch.StartRadSearch((left_x + right_x)/2, (top_y + bottom_y)/2, radius); 00641 // In Vertical Page mode, once we have an estimate of the vertical line 00642 // spacing, the minimum amount of gutter space before a possible tab is 00643 // increased under the assumption that column partition is always larger 00644 // than line spacing. 00645 int min_spacing = 00646 static_cast<int>(height * textord_tabfind_aligned_gap_fraction); 00647 if (min_gutter_width > min_spacing) 00648 min_spacing = min_gutter_width; 00649 int min_ragged_gutter = kRaggedGutterMultiple * gridsize(); 00650 if (min_gutter_width > min_ragged_gutter) 00651 min_ragged_gutter = min_gutter_width; 00652 int target_right = left_x - min_spacing; 00653 int target_left = right_x + min_spacing; 00654 // We will be evaluating whether the left edge could be a left tab, and 00655 // whether the right edge could be a right tab. 00656 // A box can be a tab if its bool is_(left/right)_tab remains true, meaning 00657 // that no blobs have been found in the gutter during the radial search. 00658 // A box can also be a tab if there are objects in the gutter only above 00659 // or only below, and there are aligned objects on the opposite side, but 00660 // not too many unaligned objects. The maybe_(left/right)_tab_up counts 00661 // aligned objects above and negatively counts unaligned objects above, 00662 // and is set to -MAX_INT32 if a gutter object is found above. 00663 // The other 3 maybe ints work similarly for the other sides. 00664 // These conditions are very strict, to minimize false positives, and really 00665 // only aligned tabs and outermost ragged tab blobs will qualify, so we 00666 // also have maybe_ragged_left/right with less stringent rules. 00667 // A blob that is maybe_ragged_left/right will be further qualified later, 00668 // using the min_ragged_gutter. 00669 bool is_left_tab = true; 00670 bool is_right_tab = true; 00671 bool maybe_ragged_left = true; 00672 bool maybe_ragged_right = true; 00673 int maybe_left_tab_up = 0; 00674 int maybe_right_tab_up = 0; 00675 int maybe_left_tab_down = 0; 00676 int maybe_right_tab_down = 0; 00677 if (bbox->leader_on_left()) { 00678 is_left_tab = false; 00679 maybe_ragged_left = false; 00680 maybe_left_tab_up = -MAX_INT32; 00681 maybe_left_tab_down = -MAX_INT32; 00682 } 00683 if (bbox->leader_on_right()) { 00684 is_right_tab = false; 00685 maybe_ragged_right = false; 00686 maybe_right_tab_up = -MAX_INT32; 00687 maybe_right_tab_down = -MAX_INT32; 00688 } 00689 int alignment_tolerance = static_cast<int>(resolution_ * kAlignedFraction); 00690 BLOBNBOX* neighbour = NULL; 00691 while ((neighbour = radsearch.NextRadSearch()) != NULL) { 00692 if (neighbour == bbox) 00693 continue; 00694 TBOX nbox = neighbour->bounding_box(); 00695 int n_left = nbox.left(); 00696 int n_right = nbox.right(); 00697 if (debug) 00698 tprintf("Neighbour at (%d,%d)->(%d,%d)\n", 00699 n_left, nbox.bottom(), n_right, nbox.top()); 00700 // If the neighbouring blob is the wrong side of a separator line, then it 00701 // "doesn't exist" as far as we are concerned. 00702 if (n_right > right_column_edge || n_left < left_column_edge || 00703 left_x < neighbour->left_rule() || right_x > neighbour->right_rule()) 00704 continue; // Separator line in the way. 00705 int n_mid_x = (n_left + n_right) / 2; 00706 int n_mid_y = (nbox.top() + nbox.bottom()) / 2; 00707 if (n_mid_x <= left_x && n_right >= target_right) { 00708 if (debug) 00709 tprintf("Not a left tab\n"); 00710 is_left_tab = false; 00711 if (n_mid_y < top_y) 00712 maybe_left_tab_down = -MAX_INT32; 00713 if (n_mid_y > bottom_y) 00714 maybe_left_tab_up = -MAX_INT32; 00715 } else if (NearlyEqual(left_x, n_left, alignment_tolerance)) { 00716 if (debug) 00717 tprintf("Maybe a left tab\n"); 00718 if (n_mid_y > top_y && maybe_left_tab_up > -MAX_INT32) 00719 ++maybe_left_tab_up; 00720 if (n_mid_y < bottom_y && maybe_left_tab_down > -MAX_INT32) 00721 ++maybe_left_tab_down; 00722 } else if (n_left < left_x && n_right >= left_x) { 00723 // Overlaps but not aligned so negative points on a maybe. 00724 if (debug) 00725 tprintf("Maybe Not a left tab\n"); 00726 if (n_mid_y > top_y && maybe_left_tab_up > -MAX_INT32) 00727 --maybe_left_tab_up; 00728 if (n_mid_y < bottom_y && maybe_left_tab_down > -MAX_INT32) 00729 --maybe_left_tab_down; 00730 } 00731 if (n_left < left_x && nbox.y_overlap(box) && n_right >= target_right) { 00732 maybe_ragged_left = false; 00733 if (debug) 00734 tprintf("Not a ragged left\n"); 00735 } 00736 if (n_mid_x >= right_x && n_left <= target_left) { 00737 if (debug) 00738 tprintf("Not a right tab\n"); 00739 is_right_tab = false; 00740 if (n_mid_y < top_y) 00741 maybe_right_tab_down = -MAX_INT32; 00742 if (n_mid_y > bottom_y) 00743 maybe_right_tab_up = -MAX_INT32; 00744 } else if (NearlyEqual(right_x, n_right, alignment_tolerance)) { 00745 if (debug) 00746 tprintf("Maybe a right tab\n"); 00747 if (n_mid_y > top_y && maybe_right_tab_up > -MAX_INT32) 00748 ++maybe_right_tab_up; 00749 if (n_mid_y < bottom_y && maybe_right_tab_down > -MAX_INT32) 00750 ++maybe_right_tab_down; 00751 } else if (n_right > right_x && n_left <= right_x) { 00752 // Overlaps but not aligned so negative points on a maybe. 00753 if (debug) 00754 tprintf("Maybe Not a right tab\n"); 00755 if (n_mid_y > top_y && maybe_right_tab_up > -MAX_INT32) 00756 --maybe_right_tab_up; 00757 if (n_mid_y < bottom_y && maybe_right_tab_down > -MAX_INT32) 00758 --maybe_right_tab_down; 00759 } 00760 if (n_right > right_x && nbox.y_overlap(box) && n_left <= target_left) { 00761 maybe_ragged_right = false; 00762 if (debug) 00763 tprintf("Not a ragged right\n"); 00764 } 00765 if (maybe_left_tab_down == -MAX_INT32 && maybe_left_tab_up == -MAX_INT32 && 00766 maybe_right_tab_down == -MAX_INT32 && maybe_right_tab_up == -MAX_INT32) 00767 break; 00768 } 00769 if (is_left_tab || maybe_left_tab_up > 1 || maybe_left_tab_down > 1) { 00770 bbox->set_left_tab_type(TT_MAYBE_ALIGNED); 00771 } else if (maybe_ragged_left && ConfirmRaggedLeft(bbox, min_ragged_gutter)) { 00772 bbox->set_left_tab_type(TT_MAYBE_RAGGED); 00773 } else { 00774 bbox->set_left_tab_type(TT_NONE); 00775 } 00776 if (is_right_tab || maybe_right_tab_up > 1 || maybe_right_tab_down > 1) { 00777 bbox->set_right_tab_type(TT_MAYBE_ALIGNED); 00778 } else if (maybe_ragged_right && 00779 ConfirmRaggedRight(bbox, min_ragged_gutter)) { 00780 bbox->set_right_tab_type(TT_MAYBE_RAGGED); 00781 } else { 00782 bbox->set_right_tab_type(TT_NONE); 00783 } 00784 if (debug) { 00785 tprintf("Left result = %s, Right result=%s\n", 00786 bbox->left_tab_type() == TT_MAYBE_ALIGNED ? "Aligned" : 00787 (bbox->left_tab_type() == TT_MAYBE_RAGGED ? "Ragged" : "None"), 00788 bbox->right_tab_type() == TT_MAYBE_ALIGNED ? "Aligned" : 00789 (bbox->right_tab_type() == TT_MAYBE_RAGGED ? "Ragged" : "None")); 00790 } 00791 return bbox->left_tab_type() != TT_NONE || bbox->right_tab_type() != TT_NONE; 00792 } 00793 00794 // Returns true if there is nothing in the rectangle of width min_gutter to 00795 // the left of bbox. 00796 bool TabFind::ConfirmRaggedLeft(BLOBNBOX* bbox, int min_gutter) { 00797 TBOX search_box(bbox->bounding_box()); 00798 search_box.set_right(search_box.left()); 00799 search_box.set_left(search_box.left() - min_gutter); 00800 return NothingYOverlapsInBox(search_box, bbox->bounding_box()); 00801 } 00802 00803 // Returns true if there is nothing in the rectangle of width min_gutter to 00804 // the right of bbox. 00805 bool TabFind::ConfirmRaggedRight(BLOBNBOX* bbox, int min_gutter) { 00806 TBOX search_box(bbox->bounding_box()); 00807 search_box.set_left(search_box.right()); 00808 search_box.set_right(search_box.right() + min_gutter); 00809 return NothingYOverlapsInBox(search_box, bbox->bounding_box()); 00810 } 00811 00812 // Returns true if there is nothing in the given search_box that vertically 00813 // overlaps target_box other than target_box itself. 00814 bool TabFind::NothingYOverlapsInBox(const TBOX& search_box, 00815 const TBOX& target_box) { 00816 BlobGridSearch rsearch(this); 00817 rsearch.StartRectSearch(search_box); 00818 BLOBNBOX* blob; 00819 while ((blob = rsearch.NextRectSearch()) != NULL) { 00820 const TBOX& box = blob->bounding_box(); 00821 if (box.y_overlap(target_box) && !(box == target_box)) 00822 return false; 00823 } 00824 return true; 00825 } 00826 00827 void TabFind::FindAllTabVectors(int min_gutter_width) { 00828 // A list of vectors that will be created in estimating the skew. 00829 TabVector_LIST dummy_vectors; 00830 // An estimate of the vertical direction, revised as more lines are added. 00831 int vertical_x = 0; 00832 int vertical_y = 1; 00833 // Find an estimate of the vertical direction by finding some tab vectors. 00834 // Slowly up the search size until we get some vectors. 00835 for (int search_size = kMinVerticalSearch; search_size < kMaxVerticalSearch; 00836 search_size += kMinVerticalSearch) { 00837 int vector_count = FindTabVectors(search_size, TA_LEFT_ALIGNED, 00838 min_gutter_width, 00839 &dummy_vectors, 00840 &vertical_x, &vertical_y); 00841 vector_count += FindTabVectors(search_size, TA_RIGHT_ALIGNED, 00842 min_gutter_width, 00843 &dummy_vectors, 00844 &vertical_x, &vertical_y); 00845 if (vector_count > 0) 00846 break; 00847 } 00848 // Get rid of the test vectors and reset the types of the tabs. 00849 dummy_vectors.clear(); 00850 for (int i = 0; i < left_tab_boxes_.size(); ++i) { 00851 BLOBNBOX* bbox = left_tab_boxes_[i]; 00852 if (bbox->left_tab_type() == TT_CONFIRMED) 00853 bbox->set_left_tab_type(TT_MAYBE_ALIGNED); 00854 } 00855 for (int i = 0; i < right_tab_boxes_.size(); ++i) { 00856 BLOBNBOX* bbox = right_tab_boxes_[i]; 00857 if (bbox->right_tab_type() == TT_CONFIRMED) 00858 bbox->set_right_tab_type(TT_MAYBE_ALIGNED); 00859 } 00860 if (textord_debug_tabfind) { 00861 tprintf("Beginning real tab search with vertical = %d,%d...\n", 00862 vertical_x, vertical_y); 00863 } 00864 // Now do the real thing ,but keep the vectors in the dummy_vectors list 00865 // until they are all done, so we don't get the tab vectors confused with 00866 // the rule line vectors. 00867 FindTabVectors(kMaxVerticalSearch, TA_LEFT_ALIGNED, min_gutter_width, 00868 &dummy_vectors, &vertical_x, &vertical_y); 00869 FindTabVectors(kMaxVerticalSearch, TA_RIGHT_ALIGNED, min_gutter_width, 00870 &dummy_vectors, &vertical_x, &vertical_y); 00871 FindTabVectors(kMaxRaggedSearch, TA_LEFT_RAGGED, min_gutter_width, 00872 &dummy_vectors, &vertical_x, &vertical_y); 00873 FindTabVectors(kMaxRaggedSearch, TA_RIGHT_RAGGED, min_gutter_width, 00874 &dummy_vectors, &vertical_x, &vertical_y); 00875 // Now add the vectors to the vectors_ list. 00876 TabVector_IT v_it(&vectors_); 00877 v_it.add_list_after(&dummy_vectors); 00878 // Now use the summed (mean) vertical vector as the direction for everything. 00879 SetVerticalSkewAndParellelize(vertical_x, vertical_y); 00880 } 00881 00882 // Helper for FindAllTabVectors finds the vectors of a particular type. 00883 int TabFind::FindTabVectors(int search_size_multiple, TabAlignment alignment, 00884 int min_gutter_width, TabVector_LIST* vectors, 00885 int* vertical_x, int* vertical_y) { 00886 TabVector_IT vector_it(vectors); 00887 int vector_count = 0; 00888 // Search the right or left tab boxes, looking for tab vectors. 00889 bool right = alignment == TA_RIGHT_ALIGNED || alignment == TA_RIGHT_RAGGED; 00890 const GenericVector<BLOBNBOX*>& boxes = right ? right_tab_boxes_ 00891 : left_tab_boxes_; 00892 for (int i = 0; i < boxes.size(); ++i) { 00893 BLOBNBOX* bbox = boxes[i]; 00894 if ((!right && bbox->left_tab_type() == TT_MAYBE_ALIGNED) || 00895 (right && bbox->right_tab_type() == TT_MAYBE_ALIGNED)) { 00896 TabVector* vector = FindTabVector(search_size_multiple, min_gutter_width, 00897 alignment, 00898 bbox, vertical_x, vertical_y); 00899 if (vector != NULL) { 00900 ++vector_count; 00901 vector_it.add_to_end(vector); 00902 } 00903 } 00904 } 00905 return vector_count; 00906 } 00907 00908 // Finds a vector corresponding to a tabstop running through the 00909 // given box of the given alignment type. 00910 // search_size_multiple is a multiple of height used to control 00911 // the size of the search. 00912 // vertical_x and y are updated with an estimate of the real 00913 // vertical direction. (skew finding.) 00914 // Returns NULL if no decent tabstop can be found. 00915 TabVector* TabFind::FindTabVector(int search_size_multiple, 00916 int min_gutter_width, 00917 TabAlignment alignment, 00918 BLOBNBOX* bbox, 00919 int* vertical_x, int* vertical_y) { 00920 int height = MAX(bbox->bounding_box().height(), gridsize()); 00921 AlignedBlobParams align_params(*vertical_x, *vertical_y, 00922 height, 00923 search_size_multiple, min_gutter_width, 00924 resolution_, alignment); 00925 // FindVerticalAlignment is in the parent (AlignedBlob) class. 00926 return FindVerticalAlignment(align_params, bbox, vertical_x, vertical_y); 00927 } 00928 00929 // Set the vertical_skew_ member from the given vector and refit 00930 // all vectors parallel to the skew vector. 00931 void TabFind::SetVerticalSkewAndParellelize(int vertical_x, int vertical_y) { 00932 // Fit the vertical vector into an ICOORD, which is 16 bit. 00933 vertical_skew_.set_with_shrink(vertical_x, vertical_y); 00934 if (textord_debug_tabfind) 00935 tprintf("Vertical skew vector=(%d,%d)\n", 00936 vertical_skew_.x(), vertical_skew_.y()); 00937 v_it_.set_to_list(&vectors_); 00938 for (v_it_.mark_cycle_pt(); !v_it_.cycled_list(); v_it_.forward()) { 00939 TabVector* v = v_it_.data(); 00940 v->Fit(vertical_skew_, true); 00941 } 00942 // Now sort the vectors as their direction has potentially changed. 00943 SortVectors(); 00944 } 00945 00946 // Sort all the current vectors using the given vertical direction vector. 00947 void TabFind::SortVectors() { 00948 vectors_.sort(TabVector::SortVectorsByKey); 00949 v_it_.set_to_list(&vectors_); 00950 } 00951 00952 // Evaluate all the current tab vectors. 00953 void TabFind::EvaluateTabs() { 00954 TabVector_IT rule_it(&vectors_); 00955 for (rule_it.mark_cycle_pt(); !rule_it.cycled_list(); rule_it.forward()) { 00956 TabVector* tab = rule_it.data(); 00957 if (!tab->IsSeparator()) { 00958 tab->Evaluate(vertical_skew_, this); 00959 if (tab->BoxCount() < kMinEvaluatedTabs) { 00960 if (textord_debug_tabfind > 2) 00961 tab->Print("Too few boxes"); 00962 delete rule_it.extract(); 00963 v_it_.set_to_list(&vectors_); 00964 } else if (WithinTestRegion(3, tab->startpt().x(), tab->startpt().y())) { 00965 tab->Print("Evaluated tab"); 00966 } 00967 } 00968 } 00969 } 00970 00971 // Trace textlines from one side to the other of each tab vector, saving 00972 // the most frequent column widths found in a list so that a given width 00973 // can be tested for being a common width with a simple callback function. 00974 void TabFind::ComputeColumnWidths(ScrollView* tab_win, 00975 ColPartitionGrid* part_grid) { 00976 #ifndef GRAPHICS_DISABLED 00977 if (tab_win != NULL) 00978 tab_win->Pen(ScrollView::WHITE); 00979 // Accumulate column sections into a STATS 00980 int col_widths_size = (tright_.x() - bleft_.x()) / kColumnWidthFactor; 00981 STATS col_widths(0, col_widths_size + 1); 00982 ApplyPartitionsToColumnWidths(part_grid, &col_widths); 00983 if (tab_win != NULL) { 00984 tab_win->Update(); 00985 } 00986 if (textord_debug_tabfind > 1) 00987 col_widths.print(); 00988 // Now make a list of column widths. 00989 #endif // GRAPHICS_DISABLED 00990 } 00991 00992 // Find column width and pair-up tab vectors with existing ColPartitions. 00993 void TabFind::ApplyPartitionsToColumnWidths(ColPartitionGrid* part_grid, 00994 STATS* col_widths) { 00995 // For every ColPartition in the part_grid, add partners to the tabvectors 00996 // and accumulate the column widths. 00997 ColPartitionGridSearch gsearch(part_grid); 00998 gsearch.StartFullSearch(); 00999 ColPartition* part; 01000 while ((part = gsearch.NextFullSearch()) != NULL) { 01001 BLOBNBOX_C_IT blob_it(part->boxes()); 01002 if (blob_it.empty()) 01003 continue; 01004 BLOBNBOX* left_blob = blob_it.data(); 01005 blob_it.move_to_last(); 01006 BLOBNBOX* right_blob = blob_it.data(); 01007 TabVector* left_vector = LeftTabForBox(left_blob->bounding_box(), 01008 true, false); 01009 if (left_vector == NULL || left_vector->IsRightTab()) 01010 continue; 01011 TabVector* right_vector = RightTabForBox(right_blob->bounding_box(), 01012 true, false); 01013 if (right_vector == NULL || right_vector->IsLeftTab()) 01014 continue; 01015 01016 AddPartnerVector(left_blob, right_blob, left_vector, right_vector); 01017 int line_left = left_vector->XAtY(left_blob->bounding_box().bottom()); 01018 int line_right = right_vector->XAtY(right_blob->bounding_box().bottom()); 01019 // Add to STATS of measurements if the width is significant. 01020 int width = line_right - line_left; 01021 if (width >= kMinColumnWidth) 01022 col_widths->add(width / kColumnWidthFactor, 1); 01023 } 01024 } 01025 01026 // Helper makes the list of common column widths in column_widths_ from the 01027 // input col_widths. Destroys the content of col_widths by repeatedly 01028 // finding the mode and erasing the peak. 01029 void TabFind::MakeColumnWidths(int col_widths_size, STATS* col_widths) { 01030 ICOORDELT_IT w_it(&column_widths_); 01031 int total_col_count = col_widths->get_total(); 01032 while (col_widths->get_total() > 0) { 01033 int width = col_widths->mode(); 01034 int col_count = col_widths->pile_count(width); 01035 col_widths->add(width, -col_count); 01036 // Get the entire peak. 01037 for (int left = width - 1; left > 0 && 01038 col_widths->pile_count(left) > 0; 01039 --left) { 01040 int new_count = col_widths->pile_count(left); 01041 col_count += new_count; 01042 col_widths->add(left, -new_count); 01043 } 01044 for (int right = width + 1; right < col_widths_size && 01045 col_widths->pile_count(right) > 0; 01046 ++right) { 01047 int new_count = col_widths->pile_count(right); 01048 col_count += new_count; 01049 col_widths->add(right, -new_count); 01050 } 01051 if (col_count > kMinLinesInColumn && 01052 col_count > kMinFractionalLinesInColumn * total_col_count) { 01053 ICOORDELT* w = new ICOORDELT(width, col_count); 01054 w_it.add_after_then_move(w); 01055 if (textord_debug_tabfind) 01056 tprintf("Column of width %d has %d = %.2f%% lines\n", 01057 width * kColumnWidthFactor, col_count, 01058 100.0 * col_count / total_col_count); 01059 } 01060 } 01061 } 01062 01063 // Mark blobs as being in a vertical text line where that is the case. 01064 // Returns true if the majority of the image is vertical text lines. 01065 void TabFind::MarkVerticalText() { 01066 if (textord_debug_tabfind) 01067 tprintf("Checking for vertical lines\n"); 01068 BlobGridSearch gsearch(this); 01069 gsearch.StartFullSearch(); 01070 BLOBNBOX* blob = NULL; 01071 while ((blob = gsearch.NextFullSearch()) != NULL) { 01072 if (blob->region_type() < BRT_UNKNOWN) 01073 continue; 01074 if (blob->UniquelyVertical()) { 01075 blob->set_region_type(BRT_VERT_TEXT); 01076 } 01077 } 01078 } 01079 01080 int TabFind::FindMedianGutterWidth(TabVector_LIST *lines) { 01081 TabVector_IT it(lines); 01082 int prev_right = -1; 01083 int max_gap = static_cast<int>(kMaxGutterWidthAbsolute * resolution_); 01084 STATS gaps(0, max_gap); 01085 STATS heights(0, max_gap); 01086 for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) { 01087 TabVector* v = it.data(); 01088 TabVector* partner = v->GetSinglePartner(); 01089 if (!v->IsLeftTab() || v->IsSeparator() || !partner) continue; 01090 heights.add(partner->startpt().x() - v->startpt().x(), 1); 01091 if (prev_right > 0 && v->startpt().x() > prev_right) { 01092 gaps.add(v->startpt().x() - prev_right, 1); 01093 } 01094 prev_right = partner->startpt().x(); 01095 } 01096 if (textord_debug_tabfind) 01097 tprintf("TabGutter total %d median_gap %.2f median_hgt %.2f\n", 01098 gaps.get_total(), gaps.median(), heights.median()); 01099 if (gaps.get_total() < kMinLinesInColumn) return 0; 01100 return static_cast<int>(gaps.median()); 01101 } 01102 01103 // Find the next adjacent (looking to the left or right) blob on this text 01104 // line, with the constraint that it must vertically significantly overlap 01105 // the [top_y, bottom_y] range. 01106 // If ignore_images is true, then blobs with aligned_text() < 0 are treated 01107 // as if they do not exist. 01108 BLOBNBOX* TabFind::AdjacentBlob(const BLOBNBOX* bbox, 01109 bool look_left, bool ignore_images, 01110 double min_overlap_fraction, 01111 int gap_limit, int top_y, int bottom_y) { 01112 GridSearch<BLOBNBOX, BLOBNBOX_CLIST, BLOBNBOX_C_IT> sidesearch(this); 01113 const TBOX& box = bbox->bounding_box(); 01114 int left = box.left(); 01115 int right = box.right(); 01116 int mid_x = (left + right) / 2; 01117 sidesearch.StartSideSearch(mid_x, bottom_y, top_y); 01118 int best_gap = 0; 01119 bool debug = WithinTestRegion(3, left, bottom_y); 01120 BLOBNBOX* result = NULL; 01121 BLOBNBOX* neighbour = NULL; 01122 while ((neighbour = sidesearch.NextSideSearch(look_left)) != NULL) { 01123 if (debug) { 01124 tprintf("Adjacent blob: considering box:"); 01125 neighbour->bounding_box().print(); 01126 } 01127 if (neighbour == bbox || 01128 (ignore_images && neighbour->region_type() < BRT_UNKNOWN)) 01129 continue; 01130 const TBOX& nbox = neighbour->bounding_box(); 01131 int n_top_y = nbox.top(); 01132 int n_bottom_y = nbox.bottom(); 01133 int v_overlap = MIN(n_top_y, top_y) - MAX(n_bottom_y, bottom_y); 01134 int height = top_y - bottom_y; 01135 int n_height = n_top_y - n_bottom_y; 01136 if (v_overlap > min_overlap_fraction * MIN(height, n_height) && 01137 (min_overlap_fraction == 0.0 || !DifferentSizes(height, n_height))) { 01138 int n_left = nbox.left(); 01139 int n_right = nbox.right(); 01140 int h_gap = MAX(n_left, left) - MIN(n_right, right); 01141 int n_mid_x = (n_left + n_right) / 2; 01142 if (look_left == (n_mid_x < mid_x) && n_mid_x != mid_x) { 01143 if (h_gap > gap_limit) { 01144 // Hit a big gap before next tab so don't return anything. 01145 if (debug) 01146 tprintf("Giving up due to big gap = %d vs %d\n", 01147 h_gap, gap_limit); 01148 return result; 01149 } 01150 if (h_gap > 0 && (look_left ? neighbour->right_tab_type() 01151 : neighbour->left_tab_type()) >= TT_CONFIRMED) { 01152 // Hit a tab facing the wrong way. Stop in case we are crossing 01153 // the column boundary. 01154 if (debug) 01155 tprintf("Collision with like tab of type %d at %d,%d\n", 01156 look_left ? neighbour->right_tab_type() 01157 : neighbour->left_tab_type(), 01158 n_left, nbox.bottom()); 01159 return result; 01160 } 01161 // This is a good fit to the line. Continue with this 01162 // neighbour as the bbox if the best gap. 01163 if (result == NULL || h_gap < best_gap) { 01164 if (debug) 01165 tprintf("Good result\n"); 01166 result = neighbour; 01167 best_gap = h_gap; 01168 } else { 01169 // The new one is worse, so we probably already have the best result. 01170 return result; 01171 } 01172 } else if (debug) { 01173 tprintf("Wrong way\n"); 01174 } 01175 } else if (debug) { 01176 tprintf("Insufficient overlap\n"); 01177 } 01178 } 01179 if (WithinTestRegion(3, left, box.top())) 01180 tprintf("Giving up due to end of search\n"); 01181 return result; // Hit the edge and found nothing. 01182 } 01183 01184 // Add a bi-directional partner relationship between the left 01185 // and the right. If one (or both) of the vectors is a separator, 01186 // extend a nearby extendable vector or create a new one of the 01187 // correct type, using the given left or right blob as a guide. 01188 void TabFind::AddPartnerVector(BLOBNBOX* left_blob, BLOBNBOX* right_blob, 01189 TabVector* left, TabVector* right) { 01190 const TBOX& left_box = left_blob->bounding_box(); 01191 const TBOX& right_box = right_blob->bounding_box(); 01192 if (left->IsSeparator()) { 01193 // Try to find a nearby left edge to extend. 01194 TabVector* v = LeftTabForBox(left_box, true, true); 01195 if (v != NULL && v != left && v->IsLeftTab() && 01196 v->XAtY(left_box.top()) > left->XAtY(left_box.top())) { 01197 left = v; // Found a good replacement. 01198 left->ExtendToBox(left_blob); 01199 } else { 01200 // Fake a vector. 01201 left = new TabVector(*left, TA_LEFT_RAGGED, vertical_skew_, left_blob); 01202 vectors_.add_sorted(TabVector::SortVectorsByKey, left); 01203 v_it_.move_to_first(); 01204 } 01205 } 01206 if (right->IsSeparator()) { 01207 // Try to find a nearby left edge to extend. 01208 if (WithinTestRegion(3, right_box.right(), right_box.bottom())) { 01209 tprintf("Box edge (%d,%d-%d)", 01210 right_box.right(), right_box.bottom(), right_box.top()); 01211 right->Print(" looking for improvement for"); 01212 } 01213 TabVector* v = RightTabForBox(right_box, true, true); 01214 if (v != NULL && v != right && v->IsRightTab() && 01215 v->XAtY(right_box.top()) < right->XAtY(right_box.top())) { 01216 right = v; // Found a good replacement. 01217 right->ExtendToBox(right_blob); 01218 if (WithinTestRegion(3, right_box.right(), right_box.bottom())) { 01219 right->Print("Extended vector"); 01220 } 01221 } else { 01222 // Fake a vector. 01223 right = new TabVector(*right, TA_RIGHT_RAGGED, vertical_skew_, 01224 right_blob); 01225 vectors_.add_sorted(TabVector::SortVectorsByKey, right); 01226 v_it_.move_to_first(); 01227 if (WithinTestRegion(3, right_box.right(), right_box.bottom())) { 01228 right->Print("Created new vector"); 01229 } 01230 } 01231 } 01232 left->AddPartner(right); 01233 right->AddPartner(left); 01234 } 01235 01236 // Remove separators and unused tabs from the main vectors_ list 01237 // to the dead_vectors_ list. 01238 void TabFind::CleanupTabs() { 01239 // TODO(rays) Before getting rid of separators and unused vectors, it 01240 // would be useful to try moving ragged vectors outwards to see if this 01241 // allows useful extension. Could be combined with checking ends of partners. 01242 TabVector_IT it(&vectors_); 01243 TabVector_IT dead_it(&dead_vectors_); 01244 for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) { 01245 TabVector* v = it.data(); 01246 if (v->IsSeparator() || v->Partnerless()) { 01247 dead_it.add_after_then_move(it.extract()); 01248 v_it_.set_to_list(&vectors_); 01249 } else { 01250 v->FitAndEvaluateIfNeeded(vertical_skew_, this); 01251 } 01252 } 01253 } 01254 01255 // Apply the given rotation to the given list of blobs. 01256 void TabFind::RotateBlobList(const FCOORD& rotation, BLOBNBOX_LIST* blobs) { 01257 BLOBNBOX_IT it(blobs); 01258 for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) { 01259 it.data()->rotate_box(rotation); 01260 } 01261 } 01262 01263 // Recreate the grid with deskewed BLOBNBOXes. 01264 // Returns false if the detected skew angle is impossible. 01265 bool TabFind::Deskew(TabVector_LIST* hlines, BLOBNBOX_LIST* image_blobs, 01266 TO_BLOCK* block, FCOORD* deskew, FCOORD* reskew) { 01267 ComputeDeskewVectors(deskew, reskew); 01268 if (deskew->x() < kCosMaxSkewAngle) 01269 return false; 01270 RotateBlobList(*deskew, image_blobs); 01271 RotateBlobList(*deskew, &block->blobs); 01272 RotateBlobList(*deskew, &block->small_blobs); 01273 RotateBlobList(*deskew, &block->noise_blobs); 01274 if (textord_debug_images) { 01275 // Rotate the debug pix and arrange for it to be drawn at the correct 01276 // pixel offset. 01277 Pix* pix_grey = pixRead(AlignedBlob::textord_debug_pix().string()); 01278 int width = pixGetWidth(pix_grey); 01279 int height = pixGetHeight(pix_grey); 01280 float angle = atan2(deskew->y(), deskew->x()); 01281 // Positive angle is clockwise to pixRotate. 01282 Pix* pix_rot = pixRotate(pix_grey, -angle, L_ROTATE_AREA_MAP, 01283 L_BRING_IN_WHITE, width, height); 01284 // The image must be translated by the rotation of its center, since it 01285 // has just been rotated about its center. 01286 ICOORD center_offset(width / 2, height / 2); 01287 ICOORD new_center_offset(center_offset); 01288 new_center_offset.rotate(*deskew); 01289 image_origin_ += new_center_offset - center_offset; 01290 // The image grew as it was rotated, so offset the (top/left) origin 01291 // by half the change in size. y is opposite to x because it is drawn 01292 // at ist top/left, not bottom/left. 01293 ICOORD corner_offset((width - pixGetWidth(pix_rot)) / 2, 01294 (pixGetHeight(pix_rot) - height) / 2); 01295 image_origin_ += corner_offset; 01296 pixWrite(AlignedBlob::textord_debug_pix().string(), pix_rot, IFF_PNG); 01297 pixDestroy(&pix_grey); 01298 pixDestroy(&pix_rot); 01299 } 01300 01301 // Rotate the horizontal vectors. The vertical vectors don't need 01302 // rotating as they can just be refitted. 01303 TabVector_IT h_it(hlines); 01304 for (h_it.mark_cycle_pt(); !h_it.cycled_list(); h_it.forward()) { 01305 TabVector* h = h_it.data(); 01306 h->Rotate(*deskew); 01307 } 01308 TabVector_IT d_it(&dead_vectors_); 01309 for (d_it.mark_cycle_pt(); !d_it.cycled_list(); d_it.forward()) { 01310 TabVector* d = d_it.data(); 01311 d->Rotate(*deskew); 01312 } 01313 SetVerticalSkewAndParellelize(0, 1); 01314 // Rebuild the grid to the new size. 01315 TBOX grid_box(bleft_, tright_); 01316 grid_box.rotate_large(*deskew); 01317 Init(gridsize(), grid_box.botleft(), grid_box.topright()); 01318 InsertBlobsToGrid(false, false, image_blobs, this); 01319 InsertBlobsToGrid(true, false, &block->blobs, this); 01320 return true; 01321 } 01322 01323 // Flip the vertical and horizontal lines and rotate the grid ready 01324 // for working on the rotated image. 01325 // This also makes parameter adjustments for FindInitialTabVectors(). 01326 void TabFind::ResetForVerticalText(const FCOORD& rotate, const FCOORD& rerotate, 01327 TabVector_LIST* horizontal_lines, 01328 int* min_gutter_width) { 01329 // Rotate the horizontal and vertical vectors and swap them over. 01330 // Only the separators are kept and rotated; other tabs are used 01331 // to estimate the gutter width then thrown away. 01332 TabVector_LIST ex_verticals; 01333 TabVector_IT ex_v_it(&ex_verticals); 01334 TabVector_LIST vlines; 01335 TabVector_IT v_it(&vlines); 01336 while (!v_it_.empty()) { 01337 TabVector* v = v_it_.extract(); 01338 if (v->IsSeparator()) { 01339 v->Rotate(rotate); 01340 ex_v_it.add_after_then_move(v); 01341 } else { 01342 v_it.add_after_then_move(v); 01343 } 01344 v_it_.forward(); 01345 } 01346 01347 // Adjust the min gutter width for better tabbox selection 01348 // in 2nd call to FindInitialTabVectors(). 01349 int median_gutter = FindMedianGutterWidth(&vlines); 01350 if (median_gutter > *min_gutter_width) 01351 *min_gutter_width = median_gutter; 01352 01353 TabVector_IT h_it(horizontal_lines); 01354 for (h_it.mark_cycle_pt(); !h_it.cycled_list(); h_it.forward()) { 01355 TabVector* h = h_it.data(); 01356 h->Rotate(rotate); 01357 } 01358 v_it_.add_list_after(horizontal_lines); 01359 v_it_.move_to_first(); 01360 h_it.set_to_list(horizontal_lines); 01361 h_it.add_list_after(&ex_verticals); 01362 01363 // Rebuild the grid to the new size. 01364 TBOX grid_box(bleft(), tright()); 01365 grid_box.rotate_large(rotate); 01366 Init(gridsize(), grid_box.botleft(), grid_box.topright()); 01367 } 01368 01369 // Clear the grid and get rid of the tab vectors, but not separators, 01370 // ready to start again. 01371 void TabFind::Reset() { 01372 v_it_.move_to_first(); 01373 for (v_it_.mark_cycle_pt(); !v_it_.cycled_list(); v_it_.forward()) { 01374 if (!v_it_.data()->IsSeparator()) 01375 delete v_it_.extract(); 01376 } 01377 Clear(); 01378 } 01379 01380 // Reflect the separator tab vectors and the grids in the y-axis. 01381 // Can only be called after Reset! 01382 void TabFind::ReflectInYAxis() { 01383 TabVector_LIST temp_list; 01384 TabVector_IT temp_it(&temp_list); 01385 v_it_.move_to_first(); 01386 // The TabVector list only contains vertical lines, but they need to be 01387 // reflected and the list needs to be reversed, so they are still in 01388 // sort_key order. 01389 while (!v_it_.empty()) { 01390 TabVector* v = v_it_.extract(); 01391 v_it_.forward(); 01392 v->ReflectInYAxis(); 01393 temp_it.add_before_then_move(v); 01394 } 01395 v_it_.add_list_after(&temp_list); 01396 v_it_.move_to_first(); 01397 // Reset this grid with reflected bounding boxes. 01398 TBOX grid_box(bleft(), tright()); 01399 int tmp = grid_box.left(); 01400 grid_box.set_left(-grid_box.right()); 01401 grid_box.set_right(-tmp); 01402 Init(gridsize(), grid_box.botleft(), grid_box.topright()); 01403 } 01404 01405 // Compute the rotation required to deskew, and its inverse rotation. 01406 void TabFind::ComputeDeskewVectors(FCOORD* deskew, FCOORD* reskew) { 01407 double length = vertical_skew_ % vertical_skew_; 01408 length = sqrt(length); 01409 deskew->set_x(static_cast<float>(vertical_skew_.y() / length)); 01410 deskew->set_y(static_cast<float>(vertical_skew_.x() / length)); 01411 reskew->set_x(deskew->x()); 01412 reskew->set_y(-deskew->y()); 01413 } 01414 01415 // Compute and apply constraints to the end positions of TabVectors so 01416 // that where possible partners end at the same y coordinate. 01417 void TabFind::ApplyTabConstraints() { 01418 TabVector_IT it(&vectors_); 01419 for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) { 01420 TabVector* v = it.data(); 01421 v->SetupConstraints(); 01422 } 01423 for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) { 01424 TabVector* v = it.data(); 01425 // With the first and last partner, we want a common bottom and top, 01426 // respectively, and for each change of partner, we want a common 01427 // top of first with bottom of next. 01428 v->SetupPartnerConstraints(); 01429 } 01430 // TODO(rays) The back-to-back pairs should really be done like the 01431 // front-to-front pairs, but there is no convenient way of producing the 01432 // list of partners like there is with the front-to-front. 01433 for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) { 01434 TabVector* v = it.data(); 01435 if (!v->IsRightTab()) 01436 continue; 01437 // For each back-to-back pair of vectors, try for common top and bottom. 01438 TabVector_IT partner_it(it); 01439 for (partner_it.forward(); !partner_it.at_first(); partner_it.forward()) { 01440 TabVector* partner = partner_it.data(); 01441 if (!partner->IsLeftTab() || !v->VOverlap(*partner)) 01442 continue; 01443 v->SetupPartnerConstraints(partner); 01444 } 01445 } 01446 // Now actually apply the constraints to get common start/end points. 01447 for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) { 01448 TabVector* v = it.data(); 01449 if (!v->IsSeparator()) 01450 v->ApplyConstraints(); 01451 } 01452 // TODO(rays) Where constraint application fails, it would be good to try 01453 // checking the ends to see if they really should be moved. 01454 } 01455 01456 } // namespace tesseract.