// Rotate the grid by rotation, keeping cell contents.
// rotation must be a multiple of 90 degrees.
// NOTE: due to partial cells, cell coverage in the rotated grid will be
// inexact. This is why there is no Rotate for the generic BBGrid.
// TODO(rays) investigate fixing this inaccuracy by moving the origin after
// rotation.
void IntGrid::Rotate(const FCOORD& rotation) {
ASSERT_HOST(rotation.x() == 0.0f || rotation.y() == 0.0f);
ICOORD old_bleft(bleft());
ICOORD old_tright(tright());
int old_width = gridwidth();
int old_height = gridheight();
TBOX box(bleft(), tright());
box.rotate(rotation);
int* old_grid = grid_;
grid_ = NULL;
Init(gridsize(), box.botleft(), box.topright());
// Iterate over the old grid, copying data to the rotated position in the new.
int oldi = 0;
FCOORD x_step(rotation);
x_step *= gridsize();
for (int oldy = 0; oldy < old_height; ++oldy) {
FCOORD line_pos(old_bleft.x(), old_bleft.y() + gridsize() * oldy);
line_pos.rotate(rotation);
for (int oldx = 0; oldx < old_width; ++oldx, line_pos += x_step, ++oldi) {
int grid_x, grid_y;
GridCoords(static_cast<int>(line_pos.x() + 0.5),
static_cast<int>(line_pos.y() + 0.5),
&grid_x, &grid_y);
grid_[grid_y * gridwidth() + grid_x] = old_grid[oldi];
}
}
delete [] old_grid;
}
// Returns a full-resolution binary pix in which each cell over the given
// threshold is filled as a black square. pixDestroy after use.
// Edge cells, which have a zero 4-neighbour, are not marked.
Pix* IntGrid::ThresholdToPix(int threshold) const {
Pix* pix = pixCreate(tright().x() - bleft().x(),
tright().y() - bleft().y(), 1);
int cellsize = gridsize();
for (int y = 0; y < gridheight(); ++y) {
for (int x = 0; x < gridwidth(); ++x) {
if (GridCellValue(x, y) > threshold &&
GridCellValue(x - 1, y) > 0 && GridCellValue(x + 1, y) > 0 &&
GridCellValue(x, y - 1) > 0 && GridCellValue(x, y + 1) > 0) {
pixRasterop(pix, x * cellsize, tright().y() - ((y + 1) * cellsize),
cellsize, cellsize, PIX_SET, NULL, 0, 0);
}
}
}
return pix;
}
// Finds horizontal line objects in the given pix.
// Uses the given resolution to determine size thresholds instead of any
// that may be present in the pix.
// The output vectors are owned by the list and Frozen (cannot refit) by
// having no boxes, as there is no need to refit or merge separator lines.
void LineFinder::FindHorizontalLines(int resolution, Pix* pix,
TabVector_LIST* vectors) {
#ifdef HAVE_LIBLEPT
Pix* line_pix;
Boxa* boxes = GetHLineBoxes(resolution, pix, &line_pix);
C_BLOB_LIST line_cblobs;
int width = pixGetWidth(pix);
int height = pixGetHeight(pix);
ConvertBoxaToBlobs(height, width, &boxes, &line_cblobs);
// Make the BLOBNBOXes from the C_BLOBs.
BLOBNBOX_LIST line_bblobs;
C_BLOB_IT blob_it(&line_cblobs);
BLOBNBOX_IT bbox_it(&line_bblobs);
for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) {
C_BLOB* cblob = blob_it.data();
BLOBNBOX* bblob = new BLOBNBOX(cblob);
bbox_it.add_to_end(bblob);
}
ICOORD bleft(0, 0);
ICOORD tright(height, width);
int vertical_x, vertical_y;
FindLineVectors(bleft, tright, &line_bblobs, &vertical_x, &vertical_y,
vectors);
if (!vectors->empty()) {
// Some lines were found, so erase the unused blobs from the line image
// and then subtract the line image from the source.
bbox_it.move_to_first();
for (bbox_it.mark_cycle_pt(); !bbox_it.cycled_list(); bbox_it.forward()) {
BLOBNBOX* blob = bbox_it.data();
if (blob->left_tab_type() == TT_UNCONFIRMED) {
const TBOX& box = blob->bounding_box();
// Coords are in tess format so filp x and y and then covert
// to leptonica by height -y.
Box* pixbox = boxCreate(box.bottom(), height - box.right(),
box.height(), box.width());
pixClearInRect(line_pix, pixbox);
boxDestroy(&pixbox);
}
}
pixDilateBrick(line_pix, line_pix, 3, 1);
pixSubtract(pix, pix, line_pix);
if (textord_tabfind_show_vlines)
pixWrite("hlinesclean.png", line_pix, IFF_PNG);
ICOORD vertical;
vertical.set_with_shrink(vertical_x, vertical_y);
TabVector::MergeSimilarTabVectors(vertical, vectors, NULL);
// Iterate the vectors to flip them.
TabVector_IT h_it(vectors);
for (h_it.mark_cycle_pt(); !h_it.cycled_list(); h_it.forward()) {
h_it.data()->XYFlip();
}
}
pixDestroy(&line_pix);
#endif
}
// Tests each blob in the list to see if it is certain non-text using 2
// conditions:
// 1. blob overlaps a cell with high value in noise_density_ (previously set
// by ComputeNoiseDensity).
// OR 2. The blob overlaps more than max_blob_overlaps in *this grid. This
// condition is disabled with max_blob_overlaps == -1.
// If it does, the blob is declared non-text, and is used to mark up the
// nontext_mask. Such blobs are fully deleted, and non-noise blobs have their
// neighbours reset, as they may now point to deleted data.
// WARNING: The blobs list blobs may be in the *this grid, but they are
// not removed. If any deleted blobs might be in *this, then this must be
// Clear()ed immediately after MarkAndDeleteNonTextBlobs is called.
// If the win is not NULL, deleted blobs are drawn on it in red, and kept
// blobs are drawn on it in ok_color.
void CCNonTextDetect::MarkAndDeleteNonTextBlobs(BLOBNBOX_LIST* blobs,
int max_blob_overlaps,
ScrollView* win,
ScrollView::Color ok_color,
Pix* nontext_mask) {
int imageheight = tright().y() - bleft().x();
BLOBNBOX_IT blob_it(blobs);
BLOBNBOX_LIST dead_blobs;
BLOBNBOX_IT dead_it(&dead_blobs);
for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) {
BLOBNBOX* blob = blob_it.data();
TBOX box = blob->bounding_box();
if (!noise_density_->RectMostlyOverThreshold(box, max_noise_count_) &&
(max_blob_overlaps < 0 ||
!BlobOverlapsTooMuch(blob, max_blob_overlaps))) {
blob->ClearNeighbours();
#ifndef GRAPHICS_DISABLED
if (win != NULL)
blob->plot(win, ok_color, ok_color);
#endif // GRAPHICS_DISABLED
} else {
if (noise_density_->AnyZeroInRect(box)) {
// There is a danger that the bounding box may overlap real text, so
// we need to render the outline.
Pix* blob_pix = blob->cblob()->render_outline();
pixRasterop(nontext_mask, box.left(), imageheight - box.top(),
box.width(), box.height(), PIX_SRC | PIX_DST,
blob_pix, 0, 0);
pixDestroy(&blob_pix);
} else {
if (box.area() < gridsize() * gridsize()) {
// It is a really bad idea to make lots of small components in the
// photo mask, so try to join it to a bigger area by expanding the
// box in a way that does not touch any zero noise density cell.
box = AttemptBoxExpansion(box, *noise_density_, gridsize());
}
// All overlapped cells are non-zero, so just mark the rectangle.
pixRasterop(nontext_mask, box.left(), imageheight - box.top(),
box.width(), box.height(), PIX_SET, NULL, 0, 0);
}
#ifndef GRAPHICS_DISABLED
if (win != NULL)
blob->plot(win, ScrollView::RED, ScrollView::RED);
#endif // GRAPHICS_DISABLED
// It is safe to delete the cblob now, as it isn't used by the grid
// or BlobOverlapsTooMuch, and the BLOBNBOXes will go away with the
// dead_blobs list.
// TODO(rays) delete the delete when the BLOBNBOX destructor deletes
// the cblob.
delete blob->cblob();
dead_it.add_to_end(blob_it.extract());
}
}
}
// Finds vertical line objects in the given pix.
// Uses the given resolution to determine size thresholds instead of any
// that may be present in the pix.
// The output vertical_x and vertical_y contain a sum of the output vectors,
// thereby giving the mean vertical direction.
// The output vectors are owned by the list and Frozen (cannot refit) by
// having no boxes, as there is no need to refit or merge separator lines.
void LineFinder::FindVerticalLines(int resolution, Pix* pix,
int* vertical_x, int* vertical_y,
TabVector_LIST* vectors) {
#ifdef HAVE_LIBLEPT
Pix* line_pix;
Boxa* boxes = GetVLineBoxes(resolution, pix, &line_pix);
C_BLOB_LIST line_cblobs;
int width = pixGetWidth(pix);
int height = pixGetHeight(pix);
ConvertBoxaToBlobs(width, height, &boxes, &line_cblobs);
// Make the BLOBNBOXes from the C_BLOBs.
BLOBNBOX_LIST line_bblobs;
C_BLOB_IT blob_it(&line_cblobs);
BLOBNBOX_IT bbox_it(&line_bblobs);
for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) {
C_BLOB* cblob = blob_it.data();
BLOBNBOX* bblob = new BLOBNBOX(cblob);
bbox_it.add_to_end(bblob);
}
ICOORD bleft(0, 0);
ICOORD tright(width, height);
FindLineVectors(bleft, tright, &line_bblobs, vertical_x, vertical_y, vectors);
if (!vectors->empty()) {
// Some lines were found, so erase the unused blobs from the line image
// and then subtract the line image from the source.
bbox_it.move_to_first();
for (bbox_it.mark_cycle_pt(); !bbox_it.cycled_list(); bbox_it.forward()) {
BLOBNBOX* blob = bbox_it.data();
if (blob->left_tab_type() == TT_UNCONFIRMED) {
const TBOX& box = blob->bounding_box();
Box* pixbox = boxCreate(box.left(), height - box.top(),
box.width(), box.height());
pixClearInRect(line_pix, pixbox);
boxDestroy(&pixbox);
}
}
pixDilateBrick(line_pix, line_pix, 1, 3);
pixSubtract(pix, pix, line_pix);
if (textord_tabfind_show_vlines)
pixWrite("vlinesclean.png", line_pix, IFF_PNG);
ICOORD vertical;
vertical.set_with_shrink(*vertical_x, *vertical_y);
TabVector::MergeSimilarTabVectors(vertical, vectors, NULL);
}
pixDestroy(&line_pix);
#endif
}
// Returns a new IntGrid containing values equal to the sum of all the
// neighbouring cells. The returned grid must be deleted after use.
// For ease of implementation, edge cells are double counted, to make them
// have the same range as the non-edge cells.
IntGrid* IntGrid::NeighbourhoodSum() const {
IntGrid* sumgrid = new IntGrid(gridsize(), bleft(), tright());
for (int y = 0; y < gridheight(); ++y) {
for (int x = 0; x < gridwidth(); ++x) {
int cell_count = 0;
for (int yoffset = -1; yoffset <= 1; ++yoffset) {
for (int xoffset = -1; xoffset <= 1; ++xoffset) {
int grid_x = x + xoffset;
int grid_y = y + yoffset;
ClipGridCoords(&grid_x, &grid_y);
cell_count += GridCellValue(grid_x, grid_y);
}
}
if (GridCellValue(x, y) > 1)
sumgrid->SetGridCell(x, y, cell_count);
}
}
return sumgrid;
}
// Creates and returns a Pix with the same resolution as the original
// in which 1 (black) pixels represent likely non text (photo, line drawing)
// areas of the page, deleting from the blob_block the blobs that were
// determined to be non-text.
// The photo_map is used to bias the decision towards non-text, rather than
// supplying definite decision.
// The blob_block is the usual result of connected component analysis,
// holding the detected blobs.
// The returned Pix should be PixDestroyed after use.
Pix* CCNonTextDetect::ComputeNonTextMask(bool debug, Pix* photo_map,
TO_BLOCK* blob_block) {
// Insert the smallest blobs into the grid.
InsertBlobList(&blob_block->small_blobs);
InsertBlobList(&blob_block->noise_blobs);
// Add the medium blobs that don't have a good strokewidth neighbour.
// Those that do go into good_grid as an antidote to spreading beyond the
// real reaches of a noise region.
BlobGrid good_grid(gridsize(), bleft(), tright());
BLOBNBOX_IT blob_it(&blob_block->blobs);
for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) {
BLOBNBOX* blob = blob_it.data();
double perimeter_area_ratio = blob->cblob()->perimeter() / 4.0;
perimeter_area_ratio *= perimeter_area_ratio / blob->enclosed_area();
if (blob->GoodTextBlob() == 0 || perimeter_area_ratio < kMinGoodTextPARatio)
InsertBBox(true, true, blob);
else
good_grid.InsertBBox(true, true, blob);
}
noise_density_ = ComputeNoiseDensity(debug, photo_map, &good_grid);
good_grid.Clear(); // Not needed any more.
Pix* pix = noise_density_->ThresholdToPix(max_noise_count_);
if (debug) {
pixWrite("junknoisemask.png", pix, IFF_PNG);
}
ScrollView* win = NULL;
#ifndef GRAPHICS_DISABLED
if (debug) {
win = MakeWindow(0, 400, "Photo Mask Blobs");
}
#endif // GRAPHICS_DISABLED
// Large and medium blobs are not text if they overlap with "a lot" of small
// blobs.
MarkAndDeleteNonTextBlobs(&blob_block->large_blobs,
kMaxLargeOverlapsWithSmall,
win, ScrollView::DARK_GREEN, pix);
MarkAndDeleteNonTextBlobs(&blob_block->blobs, kMaxMediumOverlapsWithSmall,
win, ScrollView::WHITE, pix);
// Clear the grid of small blobs and insert the medium blobs.
Clear();
InsertBlobList(&blob_block->blobs);
MarkAndDeleteNonTextBlobs(&blob_block->large_blobs,
kMaxLargeOverlapsWithMedium,
win, ScrollView::DARK_GREEN, pix);
// Clear again before we start deleting the blobs in the grid.
Clear();
MarkAndDeleteNonTextBlobs(&blob_block->noise_blobs, -1,
win, ScrollView::CORAL, pix);
MarkAndDeleteNonTextBlobs(&blob_block->small_blobs, -1,
win, ScrollView::GOLDENROD, pix);
MarkAndDeleteNonTextBlobs(&blob_block->blobs, -1,
win, ScrollView::WHITE, pix);
if (debug) {
#ifndef GRAPHICS_DISABLED
win->Update();
#endif // GRAPHICS_DISABLED
pixWrite("junkccphotomask.png", pix, IFF_PNG);
#ifndef GRAPHICS_DISABLED
delete win->AwaitEvent(SVET_DESTROY);
delete win;
#endif // GRAPHICS_DISABLED
}
return pix;
}
/**
* Sets up auto page segmentation, determines the orientation, and corrects it.
* Somewhat arbitrary chunk of functionality, factored out of AutoPageSeg to
* facilitate testing.
* photo_mask_pix is a pointer to a NULL pointer that will be filled on return
* with the leptonica photo mask, which must be pixDestroyed by the caller.
* to_blocks is an empty list that will be filled with (usually a single)
* block that is used during layout analysis. This ugly API is required
* because of the possibility of a unlv zone file.
* TODO(rays) clean this up.
* See AutoPageSeg for other arguments.
* The returned ColumnFinder must be deleted after use.
*/
ColumnFinder* Tesseract::SetupPageSegAndDetectOrientation(
PageSegMode pageseg_mode, BLOCK_LIST* blocks, Tesseract* osd_tess,
OSResults* osr, TO_BLOCK_LIST* to_blocks, Pix** photo_mask_pix,
Pix** music_mask_pix) {
int vertical_x = 0;
int vertical_y = 1;
TabVector_LIST v_lines;
TabVector_LIST h_lines;
ICOORD bleft(0, 0);
ASSERT_HOST(pix_binary_ != NULL);
if (tessedit_dump_pageseg_images) {
pixa_debug_.AddPix(pix_binary_, "PageSegInput");
}
// Leptonica is used to find the rule/separator lines in the input.
LineFinder::FindAndRemoveLines(source_resolution_,
textord_tabfind_show_vlines, pix_binary_,
&vertical_x, &vertical_y, music_mask_pix,
&v_lines, &h_lines);
if (tessedit_dump_pageseg_images) {
pixa_debug_.AddPix(pix_binary_, "NoLines");
}
// Leptonica is used to find a mask of the photo regions in the input.
*photo_mask_pix = ImageFind::FindImages(pix_binary_, &pixa_debug_);
if (tessedit_dump_pageseg_images) {
pixa_debug_.AddPix(pix_binary_, "NoImages");
}
if (!PSM_COL_FIND_ENABLED(pageseg_mode)) v_lines.clear();
// The rest of the algorithm uses the usual connected components.
textord_.find_components(pix_binary_, blocks, to_blocks);
TO_BLOCK_IT to_block_it(to_blocks);
// There must be exactly one input block.
// TODO(rays) handle new textline finding with a UNLV zone file.
ASSERT_HOST(to_blocks->singleton());
TO_BLOCK* to_block = to_block_it.data();
TBOX blkbox = to_block->block->bounding_box();
ColumnFinder* finder = NULL;
int estimated_resolution = source_resolution_;
if (source_resolution_ == kMinCredibleResolution) {
// Try to estimate resolution from typical body text size.
int res = IntCastRounded(to_block->line_size * kResolutionEstimationFactor);
if (res > estimated_resolution && res < kMaxCredibleResolution) {
estimated_resolution = res;
tprintf("Estimating resolution as %d\n", estimated_resolution);
}
}
if (to_block->line_size >= 2) {
finder = new ColumnFinder(static_cast<int>(to_block->line_size),
blkbox.botleft(), blkbox.topright(),
estimated_resolution, textord_use_cjk_fp_model,
textord_tabfind_aligned_gap_fraction, &v_lines,
&h_lines, vertical_x, vertical_y);
finder->SetupAndFilterNoise(pageseg_mode, *photo_mask_pix, to_block);
if (equ_detect_) {
equ_detect_->LabelSpecialText(to_block);
}
BLOBNBOX_CLIST osd_blobs;
// osd_orientation is the number of 90 degree rotations to make the
// characters upright. (See osdetect.h for precise definition.)
// We want the text lines horizontal, (vertical text indicates vertical
// textlines) which may conflict (eg vertically written CJK).
int osd_orientation = 0;
bool vertical_text = textord_tabfind_force_vertical_text ||
pageseg_mode == PSM_SINGLE_BLOCK_VERT_TEXT;
if (!vertical_text && textord_tabfind_vertical_text &&
PSM_ORIENTATION_ENABLED(pageseg_mode)) {
vertical_text =
finder->IsVerticallyAlignedText(textord_tabfind_vertical_text_ratio,
to_block, &osd_blobs);
}
if (PSM_OSD_ENABLED(pageseg_mode) && osd_tess != NULL && osr != NULL) {
GenericVector<int> osd_scripts;
if (osd_tess != this) {
// We are running osd as part of layout analysis, so constrain the
// scripts to those allowed by *this.
AddAllScriptsConverted(unicharset, osd_tess->unicharset, &osd_scripts);
for (int s = 0; s < sub_langs_.size(); ++s) {
AddAllScriptsConverted(sub_langs_[s]->unicharset,
osd_tess->unicharset, &osd_scripts);
}
}
os_detect_blobs(&osd_scripts, &osd_blobs, osr, osd_tess);
if (pageseg_mode == PSM_OSD_ONLY) {
delete finder;
return NULL;
}
osd_orientation = osr->best_result.orientation_id;
double osd_score = osr->orientations[osd_orientation];
double osd_margin = min_orientation_margin * 2;
for (int i = 0; i < 4; ++i) {
if (i != osd_orientation &&
osd_score - osr->orientations[i] < osd_margin) {
osd_margin = osd_score - osr->orientations[i];
}
}
int best_script_id = osr->best_result.script_id;
const char* best_script_str =
osd_tess->unicharset.get_script_from_script_id(best_script_id);
bool cjk = best_script_id == osd_tess->unicharset.han_sid() ||
best_script_id == osd_tess->unicharset.hiragana_sid() ||
best_script_id == osd_tess->unicharset.katakana_sid() ||
strcmp("Japanese", best_script_str) == 0 ||
strcmp("Korean", best_script_str) == 0 ||
strcmp("Hangul", best_script_str) == 0;
if (cjk) {
finder->set_cjk_script(true);
}
if (osd_margin < min_orientation_margin) {
// The margin is weak.
if (!cjk && !vertical_text && osd_orientation == 2) {
// upside down latin text is improbable with such a weak margin.
tprintf("OSD: Weak margin (%.2f), horiz textlines, not CJK: "
"Don't rotate.\n", osd_margin);
osd_orientation = 0;
} else {
tprintf(
"OSD: Weak margin (%.2f) for %d blob text block, "
"but using orientation anyway: %d\n",
osd_margin, osd_blobs.length(), osd_orientation);
}
}
}
osd_blobs.shallow_clear();
finder->CorrectOrientation(to_block, vertical_text, osd_orientation);
}
return finder;
}
/**
* Sets up auto page segmentation, determines the orientation, and corrects it.
* Somewhat arbitrary chunk of functionality, factored out of AutoPageSeg to
* facilitate testing.
* photo_mask_pix is a pointer to a NULL pointer that will be filled on return
* with the leptonica photo mask, which must be pixDestroyed by the caller.
* to_blocks is an empty list that will be filled with (usually a single)
* block that is used during layout analysis. This ugly API is required
* because of the possibility of a unlv zone file.
* TODO(rays) clean this up.
* See AutoPageSeg for other arguments.
* The returned ColumnFinder must be deleted after use.
*/
ColumnFinder* Tesseract::SetupPageSegAndDetectOrientation(
bool single_column, bool osd, bool only_osd,
BLOCK_LIST* blocks, Tesseract* osd_tess, OSResults* osr,
TO_BLOCK_LIST* to_blocks, Pix** photo_mask_pix, Pix** music_mask_pix) {
int vertical_x = 0;
int vertical_y = 1;
TabVector_LIST v_lines;
TabVector_LIST h_lines;
ICOORD bleft(0, 0);
ASSERT_HOST(pix_binary_ != NULL);
if (tessedit_dump_pageseg_images) {
pixWrite("tessinput.png", pix_binary_, IFF_PNG);
}
// Leptonica is used to find the rule/separator lines in the input.
LineFinder::FindAndRemoveLines(source_resolution_,
textord_tabfind_show_vlines, pix_binary_,
&vertical_x, &vertical_y, music_mask_pix,
&v_lines, &h_lines);
if (tessedit_dump_pageseg_images)
pixWrite("tessnolines.png", pix_binary_, IFF_PNG);
// Leptonica is used to find a mask of the photo regions in the input.
*photo_mask_pix = ImageFind::FindImages(pix_binary_);
if (tessedit_dump_pageseg_images)
pixWrite("tessnoimages.png", pix_binary_, IFF_PNG);
if (single_column)
v_lines.clear();
// The rest of the algorithm uses the usual connected components.
textord_.find_components(pix_binary_, blocks, to_blocks);
TO_BLOCK_IT to_block_it(to_blocks);
// There must be exactly one input block.
// TODO(rays) handle new textline finding with a UNLV zone file.
ASSERT_HOST(to_blocks->singleton());
TO_BLOCK* to_block = to_block_it.data();
TBOX blkbox = to_block->block->bounding_box();
ColumnFinder* finder = NULL;
if (to_block->line_size >= 2) {
finder = new ColumnFinder(static_cast<int>(to_block->line_size),
blkbox.botleft(), blkbox.topright(),
source_resolution_,
&v_lines, &h_lines, vertical_x, vertical_y);
finder->SetupAndFilterNoise(*photo_mask_pix, to_block);
if (equ_detect_) {
equ_detect_->LabelSpecialText(to_block);
}
BLOBNBOX_CLIST osd_blobs;
// osd_orientation is the number of 90 degree rotations to make the
// characters upright. (See osdetect.h for precise definition.)
// We want the text lines horizontal, (vertical text indicates vertical
// textlines) which may conflict (eg vertically written CJK).
int osd_orientation = 0;
bool vertical_text = finder->IsVerticallyAlignedText(to_block, &osd_blobs);
if (osd && osd_tess != NULL && osr != NULL) {
os_detect_blobs(&osd_blobs, osr, osd_tess);
if (only_osd) {
delete finder;
return NULL;
}
osd_orientation = osr->best_result.orientation_id;
double osd_score = osr->orientations[osd_orientation];
double osd_margin = min_orientation_margin * 2;
for (int i = 0; i < 4; ++i) {
if (i != osd_orientation &&
osd_score - osr->orientations[i] < osd_margin) {
osd_margin = osd_score - osr->orientations[i];
}
}
if (osd_margin < min_orientation_margin) {
// The margin is weak.
int best_script_id = osr->best_result.script_id;
bool cjk = (best_script_id == osd_tess->unicharset.han_sid()) ||
(best_script_id == osd_tess->unicharset.hiragana_sid()) ||
(best_script_id == osd_tess->unicharset.katakana_sid());
if (!cjk && !vertical_text && osd_orientation == 2) {
// upside down latin text is improbable with such a weak margin.
tprintf("OSD: Weak margin (%.2f), horiz textlines, not CJK: "
"Don't rotate.\n", osd_margin);
osd_orientation = 0;
} else {
tprintf("OSD: Weak margin (%.2f) for %d blob text block, "
"but using orientation anyway: %d\n",
osd_blobs.length(), osd_margin, osd_orientation);
}
}
}
osd_blobs.shallow_clear();
finder->CorrectOrientation(to_block, vertical_text, osd_orientation);
}
return finder;
}
// Auto page segmentation. Divide the page image into blocks of uniform
// text linespacing and images.
// Width, height and resolution are derived from the input image.
// If the pix is non-NULL, then it is assumed to be the input, and it is
// copied to the image, otherwise the image is used directly.
// The output goes in the blocks list with corresponding TO_BLOCKs in the
// to_blocks list.
// If single_column is true, then no attempt is made to divide the image
// into columns, but multiple blocks are still made if the text is of
// non-uniform linespacing.
int Tesseract::AutoPageSeg(int width, int height, int resolution,
bool single_column, IMAGE* image,
BLOCK_LIST* blocks, TO_BLOCK_LIST* to_blocks) {
int vertical_x = 0;
int vertical_y = 1;
TabVector_LIST v_lines;
TabVector_LIST h_lines;
ICOORD bleft(0, 0);
Boxa* boxa = NULL;
Pixa* pixa = NULL;
// The blocks made by the ColumnFinder. Moved to blocks before return.
BLOCK_LIST found_blocks;
#ifdef HAVE_LIBLEPT
if (pix_binary_ != NULL) {
if (textord_debug_images) {
Pix* grey_pix = pixCreate(width, height, 8);
// Printable images are light grey on white, but for screen display
// they are black on dark grey so the other colors show up well.
if (textord_debug_printable) {
pixSetAll(grey_pix);
pixSetMasked(grey_pix, pix_binary_, 192);
} else {
pixSetAllArbitrary(grey_pix, 64);
pixSetMasked(grey_pix, pix_binary_, 0);
}
AlignedBlob::IncrementDebugPix();
pixWrite(AlignedBlob::textord_debug_pix().string(), grey_pix, IFF_PNG);
pixDestroy(&grey_pix);
}
if (tessedit_dump_pageseg_images)
pixWrite("tessinput.png", pix_binary_, IFF_PNG);
// Leptonica is used to find the lines and image regions in the input.
LineFinder::FindVerticalLines(resolution, pix_binary_,
&vertical_x, &vertical_y, &v_lines);
LineFinder::FindHorizontalLines(resolution, pix_binary_, &h_lines);
if (tessedit_dump_pageseg_images)
pixWrite("tessnolines.png", pix_binary_, IFF_PNG);
ImageFinder::FindImages(pix_binary_, &boxa, &pixa);
if (tessedit_dump_pageseg_images)
pixWrite("tessnoimages.png", pix_binary_, IFF_PNG);
// Copy the Pix to the IMAGE.
image->FromPix(pix_binary_);
if (single_column)
v_lines.clear();
}
#endif
TO_BLOCK_LIST land_blocks, port_blocks;
TBOX page_box;
// The rest of the algorithm uses the usual connected components.
find_components(blocks, &land_blocks, &port_blocks, &page_box);
TO_BLOCK_IT to_block_it(&port_blocks);
ASSERT_HOST(!to_block_it.empty());
for (to_block_it.mark_cycle_pt(); !to_block_it.cycled_list();
to_block_it.forward()) {
TO_BLOCK* to_block = to_block_it.data();
TBOX blkbox = to_block->block->bounding_box();
if (to_block->line_size >= 2) {
// Note: if there are multiple blocks, then v_lines, boxa, and pixa
// are empty on the next iteration, but in this case, we assume
// that there aren't any interesting line separators or images, since
// it means that we have a pre-defined unlv zone file.
ColumnFinder finder(static_cast<int>(to_block->line_size),
blkbox.botleft(), blkbox.topright(),
&v_lines, &h_lines, vertical_x, vertical_y);
if (finder.FindBlocks(height, resolution, single_column,
to_block, boxa, pixa, &found_blocks, to_blocks) < 0)
return -1;
finder.ComputeDeskewVectors(&deskew_, &reskew_);
boxa = NULL;
pixa = NULL;
}
}
#ifdef HAVE_LIBLEPT
boxaDestroy(&boxa);
pixaDestroy(&pixa);
#endif
blocks->clear();
BLOCK_IT block_it(blocks);
// Move the found blocks to the input/output blocks.
block_it.add_list_after(&found_blocks);
if (textord_debug_images) {
// The debug image is no longer needed so delete it.
unlink(AlignedBlob::textord_debug_pix().string());
}
return 0;
}
