// Make a block using lines parallel to the given vector that fit between
// the min and max coordinates specified by the ColPartitions.
// Construct a block from the given list of partitions.
void WorkingPartSet::MakeBlocks(const ICOORD& bleft, const ICOORD& tright,
int resolution, ColPartition_LIST* used_parts) {
part_it_.move_to_first();
while (!part_it_.empty()) {
// Gather a list of ColPartitions in block_parts that will be split
// by linespacing into smaller blocks.
ColPartition_LIST block_parts;
ColPartition_IT block_it(&block_parts);
ColPartition* next_part = NULL;
bool text_block = false;
do {
ColPartition* part = part_it_.extract();
if (part->blob_type() == BRT_UNKNOWN || part->blob_type() == BRT_TEXT)
text_block = true;
part->set_working_set(NULL);
part_it_.forward();
block_it.add_after_then_move(part);
next_part = part->SingletonPartner(false);
if (part_it_.empty() || next_part != part_it_.data()) {
// Sequences of partitions can get split by titles.
next_part = NULL;
}
// Merge adjacent blocks that are of the same type and let the
// linespacing determine the real boundaries.
if (next_part == NULL && !part_it_.empty()) {
ColPartition* next_block_part = part_it_.data();
const TBOX& part_box = part->bounding_box();
const TBOX& next_box = next_block_part->bounding_box();
// In addition to the same type, the next box must not be above the
// current box, nor (if image) too far below.
PolyBlockType type = part->type(), next_type = next_block_part->type();
if (ColPartition::TypesSimilar(type, next_type) &&
next_box.bottom() <= part_box.top() &&
(text_block ||
part_box.bottom() - next_box.top() < part_box.height()))
next_part = next_block_part;
}
} while (!part_it_.empty() && next_part != NULL);
if (!text_block) {
TO_BLOCK* to_block = ColPartition::MakeBlock(bleft, tright,
&block_parts, used_parts);
if (to_block != NULL) {
TO_BLOCK_IT to_block_it(&to_blocks_);
to_block_it.add_to_end(to_block);
BLOCK_IT block_it(&completed_blocks_);
block_it.add_to_end(to_block->block);
}
} else {
// Further sub-divide text blocks where linespacing changes.
ColPartition::LineSpacingBlocks(bleft, tright, resolution, &block_parts,
used_parts,
&completed_blocks_, &to_blocks_);
}
}
part_it_.set_to_list(&part_set_);
latest_part_ = NULL;
ASSERT_HOST(completed_blocks_.length() == to_blocks_.length());
}
void PrintSegmentationStats(BLOCK_LIST* block_list) {
int num_blocks = 0;
int num_rows = 0;
int num_words = 0;
int num_blobs = 0;
BLOCK_IT block_it(block_list);
for (block_it.mark_cycle_pt(); !block_it.cycled_list(); block_it.forward()) {
BLOCK* block = block_it.data();
++num_blocks;
ROW_IT row_it(block->row_list());
for (row_it.mark_cycle_pt(); !row_it.cycled_list(); row_it.forward()) {
++num_rows;
ROW* row = row_it.data();
// Iterate over all werds in the row.
WERD_IT werd_it(row->word_list());
for (werd_it.mark_cycle_pt(); !werd_it.cycled_list(); werd_it.forward()) {
WERD* werd = werd_it.data();
++num_words;
num_blobs += werd->cblob_list()->length();
}
}
}
tprintf("Block list stats:\nBlocks = %d\nRows = %d\nWords = %d\nBlobs = %d\n",
num_blocks, num_rows, num_words, num_blobs);
}
// Insert the given blocks at the front of the completed_blocks_ list so
// they can be kept in the correct reading order.
void WorkingPartSet::InsertCompletedBlocks(BLOCK_LIST* blocks,
TO_BLOCK_LIST* to_blocks) {
BLOCK_IT block_it(&completed_blocks_);
block_it.add_list_before(blocks);
TO_BLOCK_IT to_block_it(&to_blocks_);
to_block_it.add_list_before(to_blocks);
}
// Helper computes median xheight in the image.
static double MedianXHeight(BLOCK_LIST *block_list) {
BLOCK_IT block_it(block_list);
STATS xheights(0, block_it.data()->bounding_box().height());
for (block_it.mark_cycle_pt();
!block_it.cycled_list(); block_it.forward()) {
ROW_IT row_it(block_it.data()->row_list());
for (row_it.mark_cycle_pt(); !row_it.cycled_list(); row_it.forward()) {
xheights.add(IntCastRounded(row_it.data()->x_height()), 1);
}
}
return xheights.median();
}
/**
* Auto page segmentation. Divide the page image into blocks of uniform
* text linespacing and images.
*
* Resolution (in ppi) is derived from the input image.
*
* 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.
*
* If osd (orientation and script detection) is true then that is performed
* as well. If only_osd is true, then only orientation and script detection is
* performed. If osd is desired, (osd or only_osd) then osr_tess must be
* another Tesseract that was initialized especially for osd, and the results
* will be output into osr (orientation and script result).
*/
int Tesseract::AutoPageSeg(PageSegMode pageseg_mode,
BLOCK_LIST* blocks, TO_BLOCK_LIST* to_blocks,
Tesseract* osd_tess, OSResults* osr) {
if (textord_debug_images) {
WriteDebugBackgroundImage(textord_debug_printable, pix_binary_);
}
Pix* photomask_pix = NULL;
Pix* musicmask_pix = NULL;
// The blocks made by the ColumnFinder. Moved to blocks before return.
BLOCK_LIST found_blocks;
TO_BLOCK_LIST temp_blocks;
bool single_column = !PSM_COL_FIND_ENABLED(pageseg_mode);
bool osd_enabled = PSM_OSD_ENABLED(pageseg_mode);
bool osd_only = pageseg_mode == PSM_OSD_ONLY;
ColumnFinder* finder = SetupPageSegAndDetectOrientation(
single_column, osd_enabled, osd_only, blocks, osd_tess, osr,
&temp_blocks, &photomask_pix, &musicmask_pix);
int result = 0;
if (finder != NULL) {
TO_BLOCK_IT to_block_it(&temp_blocks);
TO_BLOCK* to_block = to_block_it.data();
if (musicmask_pix != NULL) {
// TODO(rays) pass the musicmask_pix into FindBlocks and mark music
// blocks separately. For now combine with photomask_pix.
pixOr(photomask_pix, photomask_pix, musicmask_pix);
}
if (equ_detect_) {
finder->SetEquationDetect(equ_detect_);
}
result = finder->FindBlocks(pageseg_mode, scaled_color_, scaled_factor_,
to_block, photomask_pix,
pix_thresholds_, pix_grey_,
&found_blocks, to_blocks);
if (result >= 0)
finder->GetDeskewVectors(&deskew_, &reskew_);
delete finder;
}
pixDestroy(&photomask_pix);
pixDestroy(&musicmask_pix);
if (result < 0) return result;
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 result;
}
// Make blocks out of any partitions in this WorkingPartSet, and append
// them to the end of the blocks list. bleft, tright and resolution give
// the bounds and resolution of the source image, so that blocks can be
// made to fit in the bounds.
// All ColPartitions go in the used_parts list, as they need to be kept
// around, but are no longer needed.
void WorkingPartSet::ExtractCompletedBlocks(const ICOORD& bleft,
const ICOORD& tright,
int resolution,
ColPartition_LIST* used_parts,
BLOCK_LIST* blocks,
TO_BLOCK_LIST* to_blocks) {
MakeBlocks(bleft, tright, resolution, used_parts);
BLOCK_IT block_it(blocks);
block_it.move_to_last();
block_it.add_list_after(&completed_blocks_);
TO_BLOCK_IT to_block_it(to_blocks);
to_block_it.move_to_last();
to_block_it.add_list_after(&to_blocks_);
}
void apply_box_training(BLOCK_LIST *block_list) {
BLOCK_IT block_it(block_list);
ROW_IT row_it;
ROW *row;
WERD_IT word_it;
WERD *word;
WERD *bln_word;
WERD copy_outword; // copy to denorm
PBLOB_IT blob_it;
DENORM denorm;
INT16 count = 0;
char ch[2];
ch[1] = '\0';
tprintf ("Generating training data\n");
for (block_it.mark_cycle_pt ();
!block_it.cycled_list (); block_it.forward ()) {
row_it.set_to_list (block_it.data ()->row_list ());
for (row_it.mark_cycle_pt (); !row_it.cycled_list (); row_it.forward ()) {
row = row_it.data ();
word_it.set_to_list (row->word_list ());
for (word_it.mark_cycle_pt ();
!word_it.cycled_list (); word_it.forward ()) {
word = word_it.data ();
if ((strlen (word->text ()) == 1) &&
(word->gblob_list ()->length () == 1)) {
/* Here is a word with a single char label and a single blob so train on it */
bln_word =
make_bln_copy (word, row, row->x_height (), &denorm);
blob_it.set_to_list (bln_word->blob_list ());
ch[0] = *word->text ();
tess_training_tester (blob_it.data (),
//single blob
&denorm, TRUE, //correct
ch, //correct ASCII char
1, //ASCII length
NULL);
copy_outword = *(bln_word);
copy_outword.baseline_denormalise (&denorm);
blob_it.set_to_list (copy_outword.blob_list ());
ch[0] = *word->text ();
delete bln_word;
count++;
}
}
}
}
tprintf ("Generated training data for %d blobs\n", count);
}
static void clear_any_old_text(BLOCK_LIST *block_list) {
BLOCK_IT block_it(block_list);
for (block_it.mark_cycle_pt();
!block_it.cycled_list(); block_it.forward()) {
ROW_IT row_it(block_it.data()->row_list());
for (row_it.mark_cycle_pt(); !row_it.cycled_list(); row_it.forward()) {
WERD_IT word_it(row_it.data()->word_list());
for (word_it.mark_cycle_pt();
!word_it.cycled_list(); word_it.forward()) {
word_it.data()->set_text("");
}
}
}
}
void clear_any_old_text( //remove correct text
BLOCK_LIST *block_list //real blocks
) {
BLOCK_IT block_it(block_list);
ROW_IT row_it;
WERD_IT word_it;
for (block_it.mark_cycle_pt ();
!block_it.cycled_list (); block_it.forward ()) {
row_it.set_to_list (block_it.data ()->row_list ());
for (row_it.mark_cycle_pt (); !row_it.cycled_list (); row_it.forward ()) {
word_it.set_to_list (row_it.data ()->word_list ());
for (word_it.mark_cycle_pt ();
!word_it.cycled_list (); word_it.forward ()) {
word_it.data ()->set_text ("");
}
}
}
}
/**
* Auto page segmentation. Divide the page image into blocks of uniform
* text linespacing and images.
*
* Resolution (in ppi) is derived from the input image.
*
* The output goes in the blocks list with corresponding TO_BLOCKs in the
* to_blocks list.
*
* If !PSM_COL_FIND_ENABLED(pageseg_mode), 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.
*
* If diacritic_blobs is non-null, then diacritics/noise blobs, that would
* confuse layout anaylsis by causing textline overlap, are placed there,
* with the expectation that they will be reassigned to words later and
* noise/diacriticness determined via classification.
*
* If osd (orientation and script detection) is true then that is performed
* as well. If only_osd is true, then only orientation and script detection is
* performed. If osd is desired, (osd or only_osd) then osr_tess must be
* another Tesseract that was initialized especially for osd, and the results
* will be output into osr (orientation and script result).
*/
int Tesseract::AutoPageSeg(PageSegMode pageseg_mode, BLOCK_LIST* blocks,
TO_BLOCK_LIST* to_blocks,
BLOBNBOX_LIST* diacritic_blobs, Tesseract* osd_tess,
OSResults* osr) {
Pix* photomask_pix = NULL;
Pix* musicmask_pix = NULL;
// The blocks made by the ColumnFinder. Moved to blocks before return.
BLOCK_LIST found_blocks;
TO_BLOCK_LIST temp_blocks;
ColumnFinder* finder = SetupPageSegAndDetectOrientation(
pageseg_mode, blocks, osd_tess, osr, &temp_blocks, &photomask_pix,
&musicmask_pix);
int result = 0;
if (finder != NULL) {
TO_BLOCK_IT to_block_it(&temp_blocks);
TO_BLOCK* to_block = to_block_it.data();
if (musicmask_pix != NULL) {
// TODO(rays) pass the musicmask_pix into FindBlocks and mark music
// blocks separately. For now combine with photomask_pix.
pixOr(photomask_pix, photomask_pix, musicmask_pix);
}
if (equ_detect_) {
finder->SetEquationDetect(equ_detect_);
}
result = finder->FindBlocks(pageseg_mode, scaled_color_, scaled_factor_,
to_block, photomask_pix, pix_thresholds_,
pix_grey_, &pixa_debug_, &found_blocks,
diacritic_blobs, to_blocks);
if (result >= 0)
finder->GetDeskewVectors(&deskew_, &reskew_);
delete finder;
}
pixDestroy(&photomask_pix);
pixDestroy(&musicmask_pix);
if (result < 0) return result;
blocks->clear();
BLOCK_IT block_it(blocks);
// Move the found blocks to the input/output blocks.
block_it.add_list_after(&found_blocks);
return result;
}
// This method resolves the cc bbox to a particular row and returns the row's
// xheight.
int ShiroRekhaSplitter::GetXheightForCC(Box* cc_bbox) {
if (!segmentation_block_list_) {
return global_xheight_;
}
// Compute the box coordinates in Tesseract's coordinate system.
TBOX bbox(cc_bbox->x,
pixGetHeight(orig_pix_) - cc_bbox->y - cc_bbox->h - 1,
cc_bbox->x + cc_bbox->w,
pixGetHeight(orig_pix_) - cc_bbox->y - 1);
// Iterate over all blocks.
BLOCK_IT block_it(segmentation_block_list_);
for (block_it.mark_cycle_pt(); !block_it.cycled_list(); block_it.forward()) {
BLOCK* block = block_it.data();
// Iterate over all rows in the block.
ROW_IT row_it(block->row_list());
for (row_it.mark_cycle_pt(); !row_it.cycled_list(); row_it.forward()) {
ROW* row = row_it.data();
if (!row->bounding_box().major_overlap(bbox)) {
continue;
}
// Row could be skewed, warped, etc. Use the position of the box to
// determine the baseline position of the row for that x-coordinate.
// Create a square TBOX whose baseline's mid-point lies at this point
// and side is row's xheight. Take the overlap of this box with the input
// box and check if it is a 'major overlap'. If so, this box lies in this
// row. In that case, return the xheight for this row.
float box_middle = 0.5 * (bbox.left() + bbox.right());
int baseline = static_cast<int>(row->base_line(box_middle) + 0.5);
TBOX test_box(box_middle - row->x_height() / 2,
baseline,
box_middle + row->x_height() / 2,
static_cast<int>(baseline + row->x_height()));
// Compute overlap. If it is is a major overlap, this is the right row.
if (bbox.major_overlap(test_box)) {
return row->x_height();
}
}
}
// No row found for this bbox.
return kUnspecifiedXheight;
}
void RefreshWordBlobsFromNewBlobs(BLOCK_LIST* block_list,
C_BLOB_LIST* new_blobs,
C_BLOB_LIST* not_found_blobs) {
// Now iterate over all the blobs in the segmentation_block_list_, and just
// replace the corresponding c-blobs inside the werds.
BLOCK_IT block_it(block_list);
for (block_it.mark_cycle_pt(); !block_it.cycled_list(); block_it.forward()) {
BLOCK* block = block_it.data();
if (block->poly_block() != NULL && !block->poly_block()->IsText())
continue; // Don't touch non-text blocks.
// Iterate over all rows in the block.
ROW_IT row_it(block->row_list());
for (row_it.mark_cycle_pt(); !row_it.cycled_list(); row_it.forward()) {
ROW* row = row_it.data();
// Iterate over all werds in the row.
WERD_IT werd_it(row->word_list());
WERD_LIST new_words;
WERD_IT new_words_it(&new_words);
for (werd_it.mark_cycle_pt(); !werd_it.cycled_list(); werd_it.forward()) {
WERD* werd = werd_it.extract();
WERD* new_werd = werd->ConstructWerdWithNewBlobs(new_blobs,
not_found_blobs);
if (new_werd) {
// Insert this new werd into the actual row's werd-list. Remove the
// existing one.
new_words_it.add_after_then_move(new_werd);
delete werd;
} else {
// Reinsert the older word back, for lack of better options.
// This is critical since dropping the words messes up segmentation:
// eg. 1st word in the row might otherwise have W_FUZZY_NON turned on.
new_words_it.add_after_then_move(werd);
}
}
// Get rid of the old word list & replace it with the new one.
row->word_list()->clear();
werd_it.move_to_first();
werd_it.add_list_after(&new_words);
}
}
}
void ExtractBlobsFromSegmentation(BLOCK_LIST* blocks,
C_BLOB_LIST* output_blob_list) {
C_BLOB_IT return_list_it(output_blob_list);
BLOCK_IT block_it(blocks);
for (block_it.mark_cycle_pt(); !block_it.cycled_list(); block_it.forward()) {
BLOCK* block = block_it.data();
ROW_IT row_it(block->row_list());
for (row_it.mark_cycle_pt(); !row_it.cycled_list(); row_it.forward()) {
ROW* row = row_it.data();
// Iterate over all werds in the row.
WERD_IT werd_it(row->word_list());
for (werd_it.mark_cycle_pt(); !werd_it.cycled_list(); werd_it.forward()) {
WERD* werd = werd_it.data();
return_list_it.move_to_last();
return_list_it.add_list_after(werd->cblob_list());
return_list_it.move_to_last();
return_list_it.add_list_after(werd->rej_cblob_list());
}
}
}
}
void Textord::find_components(Pix* pix, BLOCK_LIST *blocks,
TO_BLOCK_LIST *to_blocks) {
int width = pixGetWidth(pix);
int height = pixGetHeight(pix);
if (width > INT16_MAX || height > INT16_MAX) {
tprintf("Input image too large! (%d, %d)\n", width, height);
return; // Can't handle it.
}
set_global_loc_code(LOC_EDGE_PROG);
BLOCK_IT block_it(blocks); // iterator
for (block_it.mark_cycle_pt(); !block_it.cycled_list();
block_it.forward()) {
BLOCK* block = block_it.data();
if (block->pdblk.poly_block() == nullptr || block->pdblk.poly_block()->IsText()) {
extract_edges(pix, block);
}
}
assign_blobs_to_blocks2(pix, blocks, to_blocks);
ICOORD page_tr(width, height);
filter_blobs(page_tr, to_blocks, !textord_test_landscape);
}
/*************************************************************************
* tidy_up()
* - report >1 block
* - sort the words in each row.
* - report any rows with no labelled words.
* - report any remaining unlabelled words
* - report total labelled words
*
*************************************************************************/
void tidy_up( //
BLOCK_LIST *block_list, //real blocks
INT16 &ok_char_count,
INT16 &ok_row_count,
INT16 &unlabelled_words,
INT16 *tgt_char_counts,
INT16 &rebalance_count,
char &min_char,
INT16 &min_samples,
INT16 &final_labelled_blob_count) {
BLOCK_IT block_it(block_list);
ROW_IT row_it;
ROW *row;
WERD_IT word_it;
WERD *word;
WERD *duplicate_word;
INT16 block_idx = 0;
INT16 row_idx;
INT16 all_row_idx = 0;
BOOL8 row_ok;
BOOL8 rebalance_needed = FALSE;
//No. of unique labelled samples
INT16 labelled_char_counts[128];
INT16 i;
char ch;
char prev_ch = '\0';
BOOL8 at_dupe_of_prev_word;
ROW *prev_row = NULL;
INT16 left;
INT16 prev_left = -1;
for (i = 0; i < 128; i++)
labelled_char_counts[i] = 0;
ok_char_count = 0;
ok_row_count = 0;
unlabelled_words = 0;
if ((applybox_debug > 4) && (block_it.length () != 1))
tprintf ("APPLY_BOXES: More than one block??\n");
for (block_it.mark_cycle_pt ();
!block_it.cycled_list (); block_it.forward ()) {
block_idx++;
row_idx = 0;
row_ok = FALSE;
row_it.set_to_list (block_it.data ()->row_list ());
for (row_it.mark_cycle_pt (); !row_it.cycled_list (); row_it.forward ()) {
row_idx++;
all_row_idx++;
row = row_it.data ();
word_it.set_to_list (row->word_list ());
word_it.sort (word_comparator);
for (word_it.mark_cycle_pt ();
!word_it.cycled_list (); word_it.forward ()) {
word = word_it.data ();
if (strlen (word->text ()) == 0) {
unlabelled_words++;
if (applybox_debug > 4) {
tprintf
("APPLY_BOXES: Unlabelled word blk:%d row:%d allrows:%d\n",
block_idx, row_idx, all_row_idx);
}
}
else {
if (word->gblob_list ()->length () != 1)
tprintf
("APPLY_BOXES: FATALITY - MULTIBLOB Labelled word blk:%d row:%d allrows:%d\n",
block_idx, row_idx, all_row_idx);
ok_char_count++;
labelled_char_counts[*word->text ()]++;
row_ok = TRUE;
}
}
if ((applybox_debug > 4) && (!row_ok)) {
tprintf
("APPLY_BOXES: Row with no labelled words blk:%d row:%d allrows:%d\n",
block_idx, row_idx, all_row_idx);
}
else
ok_row_count++;
}
}
min_samples = 9999;
for (i = 0; i < 128; i++) {
if (tgt_char_counts[i] > labelled_char_counts[i]) {
if (labelled_char_counts[i] <= 1) {
tprintf
("APPLY_BOXES: FATALITY - %d labelled samples of \"%c\" - target is %d\n",
labelled_char_counts[i], (char) i, tgt_char_counts[i]);
}
else {
rebalance_needed = TRUE;
if (applybox_debug > 0)
tprintf
("APPLY_BOXES: REBALANCE REQD \"%c\" - target of %d from %d labelled samples\n",
(char) i, tgt_char_counts[i], labelled_char_counts[i]);
}
}
if ((min_samples > labelled_char_counts[i]) && (tgt_char_counts[i] > 0)) {
min_samples = labelled_char_counts[i];
min_char = (char) i;
}
}
while (applybox_rebalance && rebalance_needed) {
block_it.set_to_list (block_list);
for (block_it.mark_cycle_pt ();
!block_it.cycled_list (); block_it.forward ()) {
row_it.set_to_list (block_it.data ()->row_list ());
for (row_it.mark_cycle_pt ();
!row_it.cycled_list (); row_it.forward ()) {
row = row_it.data ();
word_it.set_to_list (row->word_list ());
for (word_it.mark_cycle_pt ();
!word_it.cycled_list (); word_it.forward ()) {
word = word_it.data ();
left = word->bounding_box ().left ();
ch = *word->text ();
at_dupe_of_prev_word = ((row == prev_row) &&
(left = prev_left) &&
(ch == prev_ch));
if ((ch != '\0') &&
(labelled_char_counts[ch] > 1) &&
(tgt_char_counts[ch] > labelled_char_counts[ch]) &&
(!at_dupe_of_prev_word)) {
/* Duplicate the word to rebalance the labelled samples */
if (applybox_debug > 9) {
tprintf ("Duping \"%c\" from ", ch);
word->bounding_box ().print ();
}
duplicate_word = new WERD;
*duplicate_word = *word;
word_it.add_after_then_move (duplicate_word);
rebalance_count++;
labelled_char_counts[ch]++;
}
prev_row = row;
prev_left = left;
prev_ch = ch;
}
}
}
rebalance_needed = FALSE;
for (i = 0; i < 128; i++) {
if ((tgt_char_counts[i] > labelled_char_counts[i]) &&
(labelled_char_counts[i] > 1)) {
rebalance_needed = TRUE;
break;
}
}
}
/* Now final check - count labelled blobs */
final_labelled_blob_count = 0;
block_it.set_to_list (block_list);
for (block_it.mark_cycle_pt ();
!block_it.cycled_list (); block_it.forward ()) {
row_it.set_to_list (block_it.data ()->row_list ());
for (row_it.mark_cycle_pt (); !row_it.cycled_list (); row_it.forward ()) {
row = row_it.data ();
word_it.set_to_list (row->word_list ());
word_it.sort (word_comparator);
for (word_it.mark_cycle_pt ();
!word_it.cycled_list (); word_it.forward ()) {
word = word_it.data ();
if ((strlen (word->text ()) == 1) &&
(word->gblob_list ()->length () == 1))
final_labelled_blob_count++;
}
}
}
}
ROW *find_row_of_box( //
BLOCK_LIST *block_list, //real blocks
BOX box, //from boxfile
INT16 &block_id,
INT16 &row_id_to_process) {
BLOCK_IT block_it(block_list);
BLOCK *block;
ROW_IT row_it;
ROW *row;
ROW *row_to_process = NULL;
INT16 row_id;
WERD_IT word_it;
WERD *word;
BOOL8 polyg;
PBLOB_IT blob_it;
PBLOB *blob;
OUTLINE_IT outline_it;
OUTLINE *outline;
/*
Find row to process - error if box REALLY overlaps more than one row. (I.e
it overlaps blobs in the row - not just overlaps the bounding box of the
whole row.)
*/
block_id = 0;
for (block_it.mark_cycle_pt ();
!block_it.cycled_list (); block_it.forward ()) {
block_id++;
row_id = 0;
block = block_it.data ();
if (block->bounding_box ().overlap (box)) {
row_it.set_to_list (block->row_list ());
for (row_it.mark_cycle_pt ();
!row_it.cycled_list (); row_it.forward ()) {
row_id++;
row = row_it.data ();
if (row->bounding_box ().overlap (box)) {
word_it.set_to_list (row->word_list ());
for (word_it.mark_cycle_pt ();
!word_it.cycled_list (); word_it.forward ()) {
word = word_it.data ();
polyg = word->flag (W_POLYGON);
if (word->bounding_box ().overlap (box)) {
blob_it.set_to_list (word->gblob_list ());
for (blob_it.mark_cycle_pt ();
!blob_it.cycled_list (); blob_it.forward ()) {
blob = blob_it.data ();
if (gblob_bounding_box (blob, polyg).
overlap (box)) {
outline_it.
set_to_list (gblob_out_list
(blob, polyg));
for (outline_it.mark_cycle_pt ();
!outline_it.cycled_list ();
outline_it.forward ()) {
outline = outline_it.data ();
if (goutline_bounding_box
(outline, polyg).major_overlap (box)) {
if ((row_to_process == NULL) ||
(row_to_process == row)) {
row_to_process = row;
row_id_to_process = row_id;
}
else
/* RETURN ERROR Box overlaps blobs in more than one row */
return NULL;
}
}
}
}
}
}
}
}
}
}
return row_to_process;
}
// 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;
}
// Segment the page according to the current value of tessedit_pageseg_mode.
// If the pix_binary_ member is not NULL, it is used as the source image,
// and copied to image, otherwise it just uses image as the input.
// On return the blocks list owns all the constructed page layout.
int Tesseract::SegmentPage(const STRING* input_file,
IMAGE* image, BLOCK_LIST* blocks) {
int width = image->get_xsize();
int height = image->get_ysize();
int resolution = image->get_res();
#ifdef HAVE_LIBLEPT
if (pix_binary_ != NULL) {
width = pixGetWidth(pix_binary_);
height = pixGetHeight(pix_binary_);
resolution = pixGetXRes(pix_binary_);
}
#endif
// Zero resolution messes up the algorithms, so make sure it is credible.
if (resolution < kMinCredibleResolution)
resolution = kDefaultResolution;
// Get page segmentation mode.
PageSegMode pageseg_mode = static_cast<PageSegMode>(
static_cast<int>(tessedit_pageseg_mode));
// If a UNLV zone file can be found, use that instead of segmentation.
if (pageseg_mode != tesseract::PSM_AUTO &&
input_file != NULL && input_file->length() > 0) {
STRING name = *input_file;
const char* lastdot = strrchr(name.string(), '.');
if (lastdot != NULL)
name[lastdot - name.string()] = '\0';
read_unlv_file(name, width, height, blocks);
}
bool single_column = pageseg_mode > PSM_AUTO;
if (blocks->empty()) {
// No UNLV file present. Work according to the PageSegMode.
// First make a single block covering the whole image.
BLOCK_IT block_it(blocks);
BLOCK* block = new BLOCK("", TRUE, 0, 0, 0, 0, width, height);
block_it.add_to_end(block);
} else {
// UNLV file present. Use PSM_SINGLE_COLUMN.
pageseg_mode = PSM_SINGLE_COLUMN;
}
TO_BLOCK_LIST land_blocks, port_blocks;
TBOX page_box;
if (pageseg_mode <= PSM_SINGLE_COLUMN) {
if (AutoPageSeg(width, height, resolution, single_column,
image, blocks, &port_blocks) < 0) {
return -1;
}
// To create blobs from the image region bounds uncomment this line:
// port_blocks.clear(); // Uncomment to go back to the old mode.
} else {
#if HAVE_LIBLEPT
image->FromPix(pix_binary_);
#endif
deskew_ = FCOORD(1.0f, 0.0f);
reskew_ = FCOORD(1.0f, 0.0f);
}
if (blocks->empty()) {
tprintf("Empty page\n");
return 0; // AutoPageSeg found an empty page.
}
if (port_blocks.empty()) {
// AutoPageSeg was not used, so we need to find_components first.
find_components(blocks, &land_blocks, &port_blocks, &page_box);
} else {
// AutoPageSeg does not need to find_components as it did that already.
page_box.set_left(0);
page_box.set_bottom(0);
page_box.set_right(width);
page_box.set_top(height);
// Filter_blobs sets up the TO_BLOCKs the same as find_components does.
filter_blobs(page_box.topright(), &port_blocks, true);
}
TO_BLOCK_IT to_block_it(&port_blocks);
ASSERT_HOST(!port_blocks.empty());
TO_BLOCK* to_block = to_block_it.data();
if (pageseg_mode <= PSM_SINGLE_BLOCK ||
to_block->line_size < 2) {
// For now, AUTO, SINGLE_COLUMN and SINGLE_BLOCK all map to the old
// textord. The difference is the number of blocks and how the are made.
textord_page(page_box.topright(), blocks, &land_blocks, &port_blocks,
this);
} else {
// SINGLE_LINE, SINGLE_WORD and SINGLE_CHAR all need a single row.
float gradient = make_single_row(page_box.topright(),
to_block, &port_blocks, this);
if (pageseg_mode == PSM_SINGLE_LINE) {
// SINGLE_LINE uses the old word maker on the single line.
make_words(page_box.topright(), gradient, blocks,
&land_blocks, &port_blocks, this);
} else {
// SINGLE_WORD and SINGLE_CHAR cram all the blobs into a
// single word, and in SINGLE_CHAR mode, all the outlines
// go in a single blob.
make_single_word(pageseg_mode == PSM_SINGLE_CHAR,
to_block->get_rows(), to_block->block->row_list());
}
}
return 0;
}
void apply_box_testing(BLOCK_LIST *block_list) {
BLOCK_IT block_it(block_list);
ROW_IT row_it;
ROW *row;
INT16 row_count = 0;
WERD_IT word_it;
WERD *word;
WERD *bln_word;
INT16 word_count = 0;
PBLOB_IT blob_it;
DENORM denorm;
INT16 count = 0;
char ch[2];
WERD *outword; //bln best choice
//segmentation
WERD_CHOICE *best_choice; //tess output
WERD_CHOICE *raw_choice; //top choice permuter
//detailed results
BLOB_CHOICE_LIST_CLIST blob_choices;
INT16 char_count = 0;
INT16 correct_count = 0;
INT16 err_count = 0;
INT16 rej_count = 0;
#ifndef SECURE_NAMES
WERDSTATS wordstats; //As from newdiff
#endif
char tess_rej_str[3];
char tess_long_str[3];
ch[1] = '\0';
strcpy (tess_rej_str, "|A");
strcpy (tess_long_str, "|B");
for (block_it.mark_cycle_pt ();
!block_it.cycled_list (); block_it.forward ()) {
row_it.set_to_list (block_it.data ()->row_list ());
for (row_it.mark_cycle_pt (); !row_it.cycled_list (); row_it.forward ()) {
row = row_it.data ();
row_count++;
word_count = 0;
word_it.set_to_list (row->word_list ());
for (word_it.mark_cycle_pt ();
!word_it.cycled_list (); word_it.forward ()) {
word = word_it.data ();
word_count++;
if ((strlen (word->text ()) == 1) &&
!STRING (applybox_test_exclusions).contains (*word->text ())
&& (word->gblob_list ()->length () == 1)) {
/* Here is a word with a single char label and a single blob so test it */
bln_word =
make_bln_copy (word, row, row->x_height (), &denorm);
blob_it.set_to_list (bln_word->blob_list ());
ch[0] = *word->text ();
char_count++;
best_choice = tess_segment_pass1 (bln_word,
&denorm,
tess_default_matcher,
raw_choice,
&blob_choices, outword);
/*
Test for TESS screw up on word. Recog_word has already ensured that the
choice list, outword blob lists and best_choice string are the same
length. A TESS screw up is indicated by a blank filled or 0 length string.
*/
if ((best_choice->string ().length () == 0) ||
(strspn (best_choice->string ().string (), " ") ==
best_choice->string ().length ())) {
rej_count++;
tprintf ("%d:%d: \"%s\" -> TESS FAILED\n",
row_count, word_count, ch);
#ifndef SECURE_NAMES
wordstats.word (tess_rej_str, 2, ch, 1);
#endif
}
else {
if ((best_choice->string ().length () !=
outword->blob_list ()->length ()) ||
(best_choice->string ().length () !=
blob_choices.length ())) {
tprintf
("ASSERT FAIL String:\"%s\"; Strlen=%d; #Blobs=%d; #Choices=%d\n",
best_choice->string ().string (),
best_choice->string ().length (),
outword->blob_list ()->length (),
blob_choices.length ());
}
ASSERT_HOST (best_choice->string ().length () ==
outword->blob_list ()->length ());
ASSERT_HOST (best_choice->string ().length () ==
blob_choices.length ());
fix_quotes ((char *) best_choice->string ().string (),
//turn to double
outword, &blob_choices);
if (strcmp (best_choice->string ().string (), ch) != 0) {
err_count++;
tprintf ("%d:%d: \"%s\" -> \"%s\"\n",
row_count, word_count, ch,
best_choice->string ().string ());
}
else
correct_count++;
#ifndef SECURE_NAMES
if (best_choice->string ().length () > 2)
wordstats.word (tess_long_str, 2, ch, 1);
else
wordstats.word ((char *) best_choice->string ().
string (),
best_choice->string ().length (), ch,
1);
#endif
}
delete bln_word;
delete outword;
delete best_choice;
delete raw_choice;
blob_choices.deep_clear ();
count++;
}
}
}
}
#ifndef SECURE_NAMES
wordstats.print (1, 100.0);
wordstats.conf_matrix ();
tprintf ("Tested %d chars: %d correct; %d rejected by tess; %d errs\n",
char_count, correct_count, rej_count, err_count);
#endif
}
/**
* Segment the page according to the current value of tessedit_pageseg_mode.
* pix_binary_ is used as the source image and should not be NULL.
* On return the blocks list owns all the constructed page layout.
*/
int Tesseract::SegmentPage(const STRING* input_file, BLOCK_LIST* blocks,
Tesseract* osd_tess, OSResults* osr) {
ASSERT_HOST(pix_binary_ != NULL);
int width = pixGetWidth(pix_binary_);
int height = pixGetHeight(pix_binary_);
// Get page segmentation mode.
PageSegMode pageseg_mode = static_cast<PageSegMode>(
static_cast<int>(tessedit_pageseg_mode));
// If a UNLV zone file can be found, use that instead of segmentation.
if (!PSM_COL_FIND_ENABLED(pageseg_mode) &&
input_file != NULL && input_file->length() > 0) {
STRING name = *input_file;
const char* lastdot = strrchr(name.string(), '.');
if (lastdot != NULL)
name[lastdot - name.string()] = '\0';
read_unlv_file(name, width, height, blocks);
}
if (blocks->empty()) {
// No UNLV file present. Work according to the PageSegMode.
// First make a single block covering the whole image.
BLOCK_IT block_it(blocks);
BLOCK* block = new BLOCK("", TRUE, 0, 0, 0, 0, width, height);
block->set_right_to_left(right_to_left());
block_it.add_to_end(block);
} else {
// UNLV file present. Use PSM_SINGLE_BLOCK.
pageseg_mode = PSM_SINGLE_BLOCK;
}
int auto_page_seg_ret_val = 0;
TO_BLOCK_LIST to_blocks;
if (PSM_OSD_ENABLED(pageseg_mode) || PSM_BLOCK_FIND_ENABLED(pageseg_mode) ||
PSM_SPARSE(pageseg_mode)) {
auto_page_seg_ret_val =
AutoPageSeg(pageseg_mode, blocks, &to_blocks, osd_tess, osr);
if (pageseg_mode == PSM_OSD_ONLY)
return auto_page_seg_ret_val;
// To create blobs from the image region bounds uncomment this line:
// to_blocks.clear(); // Uncomment to go back to the old mode.
} else {
deskew_ = FCOORD(1.0f, 0.0f);
reskew_ = FCOORD(1.0f, 0.0f);
if (pageseg_mode == PSM_CIRCLE_WORD) {
Pix* pixcleaned = RemoveEnclosingCircle(pix_binary_);
if (pixcleaned != NULL) {
pixDestroy(&pix_binary_);
pix_binary_ = pixcleaned;
}
}
}
if (auto_page_seg_ret_val < 0) {
return -1;
}
if (blocks->empty()) {
if (textord_debug_tabfind)
tprintf("Empty page\n");
return 0; // AutoPageSeg found an empty page.
}
textord_.TextordPage(pageseg_mode, width, height, pix_binary_,
blocks, &to_blocks);
return auto_page_seg_ret_val;
}
/**
* Segment the page according to the current value of tessedit_pageseg_mode.
* pix_binary_ is used as the source image and should not be NULL.
* On return the blocks list owns all the constructed page layout.
*/
int Tesseract::SegmentPage(const STRING* input_file, BLOCK_LIST* blocks,
Tesseract* osd_tess, OSResults* osr) {
ASSERT_HOST(pix_binary_ != NULL);
int width = pixGetWidth(pix_binary_);
int height = pixGetHeight(pix_binary_);
// Get page segmentation mode.
PageSegMode pageseg_mode = static_cast<PageSegMode>(
static_cast<int>(tessedit_pageseg_mode));
// If a UNLV zone file can be found, use that instead of segmentation.
if (!PSM_COL_FIND_ENABLED(pageseg_mode) &&
input_file != NULL && input_file->length() > 0) {
STRING name = *input_file;
const char* lastdot = strrchr(name.string(), '.');
if (lastdot != NULL)
name[lastdot - name.string()] = '\0';
read_unlv_file(name, width, height, blocks);
}
if (blocks->empty()) {
// No UNLV file present. Work according to the PageSegMode.
// First make a single block covering the whole image.
BLOCK_IT block_it(blocks);
BLOCK* block = new BLOCK("", TRUE, 0, 0, 0, 0, width, height);
block->set_right_to_left(right_to_left());
block_it.add_to_end(block);
} else {
// UNLV file present. Use PSM_SINGLE_BLOCK.
pageseg_mode = PSM_SINGLE_BLOCK;
}
// The diacritic_blobs holds noise blobs that may be diacritics. They
// are separated out on areas of the image that seem noisy and short-circuit
// the layout process, going straight from the initial partition creation
// right through to after word segmentation, where they are added to the
// rej_cblobs list of the most appropriate word. From there classification
// will determine whether they are used.
BLOBNBOX_LIST diacritic_blobs;
int auto_page_seg_ret_val = 0;
TO_BLOCK_LIST to_blocks;
if (PSM_OSD_ENABLED(pageseg_mode) || PSM_BLOCK_FIND_ENABLED(pageseg_mode) ||
PSM_SPARSE(pageseg_mode)) {
auto_page_seg_ret_val = AutoPageSeg(
pageseg_mode, blocks, &to_blocks,
enable_noise_removal ? &diacritic_blobs : NULL, osd_tess, osr);
if (pageseg_mode == PSM_OSD_ONLY)
return auto_page_seg_ret_val;
// To create blobs from the image region bounds uncomment this line:
// to_blocks.clear(); // Uncomment to go back to the old mode.
} else {
deskew_ = FCOORD(1.0f, 0.0f);
reskew_ = FCOORD(1.0f, 0.0f);
if (pageseg_mode == PSM_CIRCLE_WORD) {
Pix* pixcleaned = RemoveEnclosingCircle(pix_binary_);
if (pixcleaned != NULL) {
pixDestroy(&pix_binary_);
pix_binary_ = pixcleaned;
}
}
}
if (auto_page_seg_ret_val < 0) {
return -1;
}
if (blocks->empty()) {
if (textord_debug_tabfind)
tprintf("Empty page\n");
return 0; // AutoPageSeg found an empty page.
}
bool splitting =
pageseg_devanagari_split_strategy != ShiroRekhaSplitter::NO_SPLIT;
bool cjk_mode = textord_use_cjk_fp_model;
textord_.TextordPage(pageseg_mode, reskew_, width, height, pix_binary_,
pix_thresholds_, pix_grey_, splitting || cjk_mode,
&diacritic_blobs, blocks, &to_blocks);
return auto_page_seg_ret_val;
}
void FullPageBlock(int width, int height, BLOCK_LIST *blocks) {
BLOCK_IT block_it(blocks);
BLOCK* block = new BLOCK("", TRUE, 0, 0, 0, 0, width, height);
block_it.add_to_end(block);
}
