// Print the best guesses out of the match rating matrix.
void MATRIX::print(const UNICHARSET &unicharset) const {
tprintf("Ratings Matrix (top 3 choices)\n");
int dim = dimension();
int band_width = bandwidth();
int row, col;
for (col = 0; col < dim; ++col) {
for (row = col; row < dim && row < col + band_width; ++row) {
BLOB_CHOICE_LIST *rating = this->get(col, row);
if (rating == NOT_CLASSIFIED) continue;
BLOB_CHOICE_IT b_it(rating);
tprintf("col=%d row=%d ", col, row);
for (b_it.mark_cycle_pt(); !b_it.cycled_list(); b_it.forward()) {
tprintf("%s rat=%g cert=%g " ,
unicharset.id_to_unichar(b_it.data()->unichar_id()),
b_it.data()->rating(), b_it.data()->certainty());
}
tprintf("\n");
}
tprintf("\n");
}
tprintf("\n");
for (col = 0; col < dim; ++col) tprintf("\t%d", col);
tprintf("\n");
for (row = 0; row < dim; ++row) {
for (col = 0; col <= row; ++col) {
if (col == 0) tprintf("%d\t", row);
if (row >= col + band_width) {
tprintf(" \t");
continue;
}
BLOB_CHOICE_LIST *rating = this->get(col, row);
if (rating != NOT_CLASSIFIED) {
BLOB_CHOICE_IT b_it(rating);
int counter = 0;
for (b_it.mark_cycle_pt(); !b_it.cycled_list(); b_it.forward()) {
tprintf("%s ",
unicharset.id_to_unichar(b_it.data()->unichar_id()));
++counter;
if (counter == 3) break;
}
tprintf("\t");
} else {
tprintf(" \t");
}
}
tprintf("\n");
}
}
// Print the best guesses out of the match rating matrix.
void MATRIX::print(const UNICHARSET &unicharset) {
tprintf("Ratings Matrix (top choices)\n");
int row, col;
for (col = 0; col < this->dimension(); ++col) tprintf("\t%d", col);
tprintf("\n");
for (row = 0; row < this->dimension(); ++row) {
for (col = 0; col <= row; ++col) {
if (col == 0) tprintf("%d\t", row);
BLOB_CHOICE_LIST *rating = this->get(col, row);
if (rating != NOT_CLASSIFIED) {
BLOB_CHOICE_IT b_it(rating);
int counter = 0;
for (b_it.mark_cycle_pt(); !b_it.cycled_list(); b_it.forward()) {
tprintf("%s ", unicharset.id_to_unichar(b_it.data()->unichar_id()));
++counter;
if (counter == 3) break;
}
tprintf("\t");
} else {
tprintf(" \t");
}
}
tprintf("\n");
}
}
/// Builds a PAGE_RES from the block_list in the way required for ApplyBoxes:
/// All fuzzy spaces are removed, and all the words are maximally chopped.
PAGE_RES* Tesseract::SetupApplyBoxes(const GenericVector<TBOX>& boxes,
BLOCK_LIST *block_list) {
PreenXHeights(block_list);
// Strip all fuzzy space markers to simplify the PAGE_RES.
BLOCK_IT b_it(block_list);
for (b_it.mark_cycle_pt(); !b_it.cycled_list(); b_it.forward()) {
BLOCK* block = b_it.data();
ROW_IT r_it(block->row_list());
for (r_it.mark_cycle_pt(); !r_it.cycled_list(); r_it.forward ()) {
ROW* row = r_it.data();
WERD_IT w_it(row->word_list());
for (w_it.mark_cycle_pt(); !w_it.cycled_list(); w_it.forward()) {
WERD* word = w_it.data();
if (word->cblob_list()->empty()) {
delete w_it.extract();
} else {
word->set_flag(W_FUZZY_SP, false);
word->set_flag(W_FUZZY_NON, false);
}
}
}
}
PAGE_RES* page_res = new PAGE_RES(false, block_list, NULL);
PAGE_RES_IT pr_it(page_res);
WERD_RES* word_res;
while ((word_res = pr_it.word()) != NULL) {
MaximallyChopWord(boxes, pr_it.block()->block,
pr_it.row()->row, word_res);
pr_it.forward();
}
return page_res;
}
// Returns true if there are any real classification results.
bool MATRIX::Classified(int col, int row, int wildcard_id) const {
if (get(col, row) == NOT_CLASSIFIED) return false;
BLOB_CHOICE_IT b_it(get(col, row));
for (b_it.mark_cycle_pt(); !b_it.cycled_list(); b_it.forward()) {
BLOB_CHOICE* choice = b_it.data();
if (choice->IsClassified())
return true;
}
return false;
}
// Generates training data for training a line recognizer, eg LSTM.
// Breaks the boxes into lines, normalizes them, converts to ImageData and
// appends them to the given training_data.
void Tesseract::TrainFromBoxes(const GenericVector<TBOX>& boxes,
const GenericVector<STRING>& texts,
BLOCK_LIST *block_list,
DocumentData* training_data) {
int box_count = boxes.size();
// Process all the text lines in this page, as defined by the boxes.
int end_box = 0;
// Don't let \t, which marks newlines in the box file, get into the line
// content, as that makes the line unusable in training.
while (end_box < texts.size() && texts[end_box] == "\t") ++end_box;
for (int start_box = end_box; start_box < box_count; start_box = end_box) {
// Find the textline of boxes starting at start and their bounding box.
TBOX line_box = boxes[start_box];
STRING line_str = texts[start_box];
for (end_box = start_box + 1; end_box < box_count && texts[end_box] != "\t";
++end_box) {
line_box += boxes[end_box];
line_str += texts[end_box];
}
// Find the most overlapping block.
BLOCK* best_block = NULL;
int best_overlap = 0;
BLOCK_IT b_it(block_list);
for (b_it.mark_cycle_pt(); !b_it.cycled_list(); b_it.forward()) {
BLOCK* block = b_it.data();
if (block->poly_block() != NULL && !block->poly_block()->IsText())
continue; // Not a text block.
TBOX block_box = block->bounding_box();
block_box.rotate(block->re_rotation());
if (block_box.major_overlap(line_box)) {
TBOX overlap_box = line_box.intersection(block_box);
if (overlap_box.area() > best_overlap) {
best_overlap = overlap_box.area();
best_block = block;
}
}
}
ImageData* imagedata = NULL;
if (best_block == NULL) {
tprintf("No block overlapping textline: %s\n", line_str.string());
} else {
imagedata = GetLineData(line_box, boxes, texts, start_box, end_box,
*best_block);
}
if (imagedata != NULL)
training_data->AddPageToDocument(imagedata);
// Don't let \t, which marks newlines in the box file, get into the line
// content, as that makes the line unusable in training.
while (end_box < texts.size() && texts[end_box] == "\t") ++end_box;
}
}
int compare_filenames(const std::string& a, const std::string& b)
{
utf8_const_iterator a_begin(a.begin()), a_end(a.end());
utf8_const_iterator b_begin(b.begin()), b_end(b.end());
utf8_const_iterator a_it(a_begin);
utf8_const_iterator b_it(b_begin);
for (; a_it != a_end && b_it != b_end; ) {
int a_chr = *a_it;
int b_chr = *b_it;
if ((a_chr >= '0') && (a_chr <= '9') && (b_chr >= '0') && (b_chr <= '9')) {
utf8_const_iterator a_it2 = a_it;
utf8_const_iterator b_it2 = b_it;
while (a_it2 != a_end && (*a_it2 >= '0') && (*a_it2 <= '9')) ++a_it2;
while (b_it2 != b_end && (*b_it2 >= '0') && (*b_it2 <= '9')) ++b_it2;
int a_num = std::strtol(std::string(a_it, a_it2).c_str(), NULL, 10);
int b_num = std::strtol(std::string(b_it, b_it2).c_str(), NULL, 10);
if (a_num != b_num)
return a_num - b_num < 0 ? -1: 1;
a_it = a_it2;
b_it = b_it2;
}
else if (is_path_separator(a_chr) && is_path_separator(b_chr)) {
++a_it;
++b_it;
}
else {
a_chr = std::tolower(a_chr);
b_chr = std::tolower(b_chr);
if (a_chr != b_chr)
return a_chr - b_chr < 0 ? -1: 1;
++a_it;
++b_it;
}
}
if (a_it == a_end && b_it == b_end)
return 0;
else if (a_it == a_end)
return -1;
else
return 1;
}
// Factory to build a TWERD from a (C_BLOB) WERD, with polygonal
// approximation along the way.
TWERD* TWERD::PolygonalCopy(WERD* src) {
TWERD* tessword = new TWERD;
tessword->latin_script = src->flag(W_SCRIPT_IS_LATIN);
C_BLOB_IT b_it(src->cblob_list());
TBLOB *tail = NULL;
for (b_it.mark_cycle_pt(); !b_it.cycled_list(); b_it.forward()) {
C_BLOB* blob = b_it.data();
TBLOB* tblob = TBLOB::PolygonalCopy(blob);
if (tail == NULL) {
tessword->blobs = tblob;
} else {
tail->next = tblob;
}
tail = tblob;
}
return tessword;
}
/// Any row xheight that is significantly different from the median is set
/// to the median.
void Tesseract::PreenXHeights(BLOCK_LIST *block_list) {
double median_xheight = MedianXHeight(block_list);
double max_deviation = kMaxXHeightDeviationFraction * median_xheight;
// Strip all fuzzy space markers to simplify the PAGE_RES.
BLOCK_IT b_it(block_list);
for (b_it.mark_cycle_pt(); !b_it.cycled_list(); b_it.forward()) {
BLOCK* block = b_it.data();
ROW_IT r_it(block->row_list());
for (r_it.mark_cycle_pt(); !r_it.cycled_list(); r_it.forward ()) {
ROW* row = r_it.data();
float diff = fabs(row->x_height() - median_xheight);
if (diff > max_deviation) {
if (applybox_debug) {
tprintf("row xheight=%g, but median xheight = %g\n",
row->x_height(), median_xheight);
}
row->set_x_height(static_cast<float>(median_xheight));
}
}
}
}
// Places a copy of blobs that are near a word (after applying rotation to the
// blob) in the most appropriate word, unless there is doubt, in which case a
// blob can end up in two words. Source blobs are not touched.
void Textord::TransferDiacriticsToWords(BLOBNBOX_LIST* diacritic_blobs,
const FCOORD& rotation,
WordGrid* word_grid) {
WordSearch ws(word_grid);
BLOBNBOX_IT b_it(diacritic_blobs);
// Apply rotation to each blob before finding the nearest words. The rotation
// allows us to only consider above/below placement and not left/right on
// vertical text, because all text is horizontal here.
for (b_it.mark_cycle_pt(); !b_it.cycled_list(); b_it.forward()) {
BLOBNBOX* blobnbox = b_it.data();
TBOX blob_box = blobnbox->bounding_box();
blob_box.rotate(rotation);
ws.StartRectSearch(blob_box);
// Above/below refer to word position relative to diacritic. Since some
// scripts eg Kannada/Telugu habitually put diacritics below words, and
// others eg Thai/Vietnamese/Latin put most diacritics above words, try
// for both if there isn't much in it.
WordWithBox* best_above_word = nullptr;
WordWithBox* best_below_word = nullptr;
int best_above_distance = 0;
int best_below_distance = 0;
for (WordWithBox* word = ws.NextRectSearch(); word != nullptr;
word = ws.NextRectSearch()) {
if (word->word()->flag(W_REP_CHAR)) continue;
TBOX word_box = word->true_bounding_box();
int x_distance = blob_box.x_gap(word_box);
int y_distance = blob_box.y_gap(word_box);
if (x_distance > 0) {
// Arbitrarily divide x-distance by 2 if there is a major y overlap,
// and the word is to the left of the diacritic. If the
// diacritic is a dropped broken character between two words, this will
// help send all the pieces to a single word, instead of splitting them
// over the 2 words.
if (word_box.major_y_overlap(blob_box) &&
blob_box.left() > word_box.right()) {
x_distance /= 2;
}
y_distance += x_distance;
}
if (word_box.y_middle() > blob_box.y_middle() &&
(best_above_word == nullptr || y_distance < best_above_distance)) {
best_above_word = word;
best_above_distance = y_distance;
}
if (word_box.y_middle() <= blob_box.y_middle() &&
(best_below_word == nullptr || y_distance < best_below_distance)) {
best_below_word = word;
best_below_distance = y_distance;
}
}
bool above_good =
best_above_word != nullptr &&
(best_below_word == nullptr ||
best_above_distance < best_below_distance + blob_box.height());
bool below_good =
best_below_word != nullptr && best_below_word != best_above_word &&
(best_above_word == nullptr ||
best_below_distance < best_above_distance + blob_box.height());
if (below_good) {
C_BLOB* copied_blob = C_BLOB::deep_copy(blobnbox->cblob());
copied_blob->rotate(rotation);
// Put the blob into the word's reject blobs list.
C_BLOB_IT blob_it(best_below_word->RejBlobs());
blob_it.add_to_end(copied_blob);
}
if (above_good) {
C_BLOB* copied_blob = C_BLOB::deep_copy(blobnbox->cblob());
copied_blob->rotate(rotation);
// Put the blob into the word's reject blobs list.
C_BLOB_IT blob_it(best_above_word->RejBlobs());
blob_it.add_to_end(copied_blob);
}
}
}
// Make the textlines and words inside each block.
void Textord::TextordPage(PageSegMode pageseg_mode, const FCOORD &reskew,
int width, int height, Pix *binary_pix,
Pix *thresholds_pix, Pix *grey_pix,
bool use_box_bottoms,
BLOCK_LIST *blocks, TO_BLOCK_LIST *to_blocks) {
page_tr_.set_x(width);
page_tr_.set_y(height);
if (to_blocks->empty()) {
// AutoPageSeg was not used, so we need to find_components first.
find_components(binary_pix, blocks, to_blocks);
TO_BLOCK_IT it(to_blocks);
for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) {
TO_BLOCK *to_block = it.data();
// Compute the edge offsets whether or not there is a grey_pix.
// We have by-passed auto page seg, so we have to run it here.
// By page segmentation mode there is no non-text to avoid running on.
to_block->ComputeEdgeOffsets(thresholds_pix, grey_pix);
}
} else if (!PSM_SPARSE(pageseg_mode)) {
// AutoPageSeg does not need to find_components as it did that already.
// Filter_blobs sets up the TO_BLOCKs the same as find_components does.
filter_blobs(page_tr_, to_blocks, true);
}
ASSERT_HOST(!to_blocks->empty());
if (pageseg_mode == PSM_SINGLE_BLOCK_VERT_TEXT) {
const FCOORD anticlockwise90(0.0f, 1.0f);
const FCOORD clockwise90(0.0f, -1.0f);
TO_BLOCK_IT it(to_blocks);
for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) {
TO_BLOCK *to_block = it.data();
BLOCK *block = to_block->block;
// Create a fake poly_block in block from its bounding box.
block->set_poly_block(new POLY_BLOCK(block->bounding_box(),
PT_VERTICAL_TEXT));
// Rotate the to_block along with its contained block and blobnbox lists.
to_block->rotate(anticlockwise90);
// Set the block's rotation values to obey the convention followed in
// layout analysis for vertical text.
block->set_re_rotation(clockwise90);
block->set_classify_rotation(clockwise90);
}
}
TO_BLOCK_IT to_block_it(to_blocks);
TO_BLOCK *to_block = to_block_it.data();
// Make the rows in the block.
float gradient = 0;
// Do it the old fashioned way.
if (PSM_LINE_FIND_ENABLED(pageseg_mode)) {
gradient = make_rows(page_tr_, to_blocks);
} else if (!PSM_SPARSE(pageseg_mode)) {
// RAW_LINE, SINGLE_LINE, SINGLE_WORD and SINGLE_CHAR all need a single row.
gradient = make_single_row(page_tr_, pageseg_mode != PSM_RAW_LINE,
to_block, to_blocks);
}
BaselineDetect baseline_detector(textord_baseline_debug,
reskew, to_blocks);
baseline_detector.ComputeStraightBaselines(use_box_bottoms);
baseline_detector.ComputeBaselineSplinesAndXheights(page_tr_, true,
textord_heavy_nr,
textord_show_final_rows,
this);
// Now make the words in the lines.
if (PSM_WORD_FIND_ENABLED(pageseg_mode)) {
// SINGLE_LINE uses the old word maker on the single line.
make_words(this, page_tr_, gradient, blocks, to_blocks);
} 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.
TO_BLOCK *to_block = to_block_it.data();
make_single_word(pageseg_mode == PSM_SINGLE_CHAR,
to_block->get_rows(), to_block->block->row_list());
}
cleanup_blocks(PSM_WORD_FIND_ENABLED(pageseg_mode), blocks);
// Remove empties.
// Compute the margins for each row in the block, to be used later for
// paragraph detection.
BLOCK_IT b_it(blocks);
for (b_it.mark_cycle_pt(); !b_it.cycled_list(); b_it.forward()) {
b_it.data()->compute_row_margins();
}
#ifndef GRAPHICS_DISABLED
close_to_win();
#endif
}
/// Consume all source blobs that strongly overlap the given box,
/// putting them into a new word, with the correct_text label.
/// Fights over which box owns which blobs are settled by
/// applying the blobs to box or next_box with the least non-overlap.
/// @return false if the box was in error, which can only be caused by
/// failing to find an overlapping blob for a box.
bool Tesseract::ResegmentWordBox(BLOCK_LIST *block_list,
const TBOX& box, const TBOX& next_box,
const char* correct_text) {
if (applybox_debug > 1) {
tprintf("\nAPPLY_BOX: in ResegmentWordBox() for %s\n", correct_text);
}
WERD* new_word = NULL;
BLOCK_IT b_it(block_list);
for (b_it.mark_cycle_pt(); !b_it.cycled_list(); b_it.forward()) {
BLOCK* block = b_it.data();
if (!box.major_overlap(block->bounding_box()))
continue;
ROW_IT r_it(block->row_list());
for (r_it.mark_cycle_pt(); !r_it.cycled_list(); r_it.forward()) {
ROW* row = r_it.data();
if (!box.major_overlap(row->bounding_box()))
continue;
WERD_IT w_it(row->word_list());
for (w_it.mark_cycle_pt(); !w_it.cycled_list(); w_it.forward()) {
WERD* word = w_it.data();
if (applybox_debug > 2) {
tprintf("Checking word:");
word->bounding_box().print();
}
if (word->text() != NULL && word->text()[0] != '\0')
continue; // Ignore words that are already done.
if (!box.major_overlap(word->bounding_box()))
continue;
C_BLOB_IT blob_it(word->cblob_list());
for (blob_it.mark_cycle_pt(); !blob_it.cycled_list();
blob_it.forward()) {
C_BLOB* blob = blob_it.data();
TBOX blob_box = blob->bounding_box();
if (!blob_box.major_overlap(box))
continue;
double current_box_miss_metric = BoxMissMetric(blob_box, box);
double next_box_miss_metric = BoxMissMetric(blob_box, next_box);
if (applybox_debug > 2) {
tprintf("Checking blob:");
blob_box.print();
tprintf("Current miss metric = %g, next = %g\n",
current_box_miss_metric, next_box_miss_metric);
}
if (current_box_miss_metric > next_box_miss_metric)
continue; // Blob is a better match for next box.
if (applybox_debug > 2) {
tprintf("Blob match: blob:");
blob_box.print();
tprintf("Matches box:");
box.print();
tprintf("With next box:");
next_box.print();
}
if (new_word == NULL) {
// Make a new word with a single blob.
new_word = word->shallow_copy();
new_word->set_text(correct_text);
w_it.add_to_end(new_word);
}
C_BLOB_IT new_blob_it(new_word->cblob_list());
new_blob_it.add_to_end(blob_it.extract());
}
}
}
}
if (new_word == NULL && applybox_debug > 0) tprintf("FAIL!\n");
return new_word != NULL;
}
void LMPainPoints::GenerateFromPath(float rating_cert_scale,
ViterbiStateEntry *vse,
WERD_RES *word_res) {
ViterbiStateEntry *curr_vse = vse;
BLOB_CHOICE *curr_b = vse->curr_b;
// The following pain point generation and priority calculation approaches
// prioritize exploring paths with low average rating of the known part of
// the path, while not relying on the ratings of the pieces to be combined.
//
// A pain point to combine the neighbors is generated for each pair of
// neighboring blobs on the path (the path is represented by vse argument
// given to GenerateFromPath()). The priority of each pain point is set to
// the average rating (per outline length) of the path, not including the
// ratings of the blobs to be combined.
// The ratings of the blobs to be combined are not used to calculate the
// priority, since it is not possible to determine from their magnitude
// whether it will be beneficial to combine the blobs. The reason is that
// chopped junk blobs (/ | - ') can have very good (low) ratings, however
// combining them will be beneficial. Blobs with high ratings might be
// over-joined pieces of characters, but also could be blobs from an unseen
// font or chopped pieces of complex characters.
while (curr_vse->parent_vse != NULL) {
ViterbiStateEntry* parent_vse = curr_vse->parent_vse;
const MATRIX_COORD& curr_cell = curr_b->matrix_cell();
const MATRIX_COORD& parent_cell = parent_vse->curr_b->matrix_cell();
MATRIX_COORD pain_coord(parent_cell.col, curr_cell.row);
if (!pain_coord.Valid(*word_res->ratings) ||
!word_res->ratings->Classified(parent_cell.col, curr_cell.row,
dict_->WildcardID())) {
// rat_subtr contains ratings sum of the two adjacent blobs to be merged.
// rat_subtr will be subtracted from the ratings sum of the path, since
// the blobs will be joined into a new blob, whose rating is yet unknown.
float rat_subtr = curr_b->rating() + parent_vse->curr_b->rating();
// ol_subtr contains the outline length of the blobs that will be joined.
float ol_subtr =
AssociateUtils::ComputeOutlineLength(rating_cert_scale, *curr_b) +
AssociateUtils::ComputeOutlineLength(rating_cert_scale,
*(parent_vse->curr_b));
// ol_dif is the outline of the path without the two blobs to be joined.
float ol_dif = vse->outline_length - ol_subtr;
// priority is set to the average rating of the path per unit of outline,
// not counting the ratings of the pieces to be joined.
float priority = ol_dif > 0 ? (vse->ratings_sum-rat_subtr)/ol_dif : 0.0;
GeneratePainPoint(pain_coord.col, pain_coord.row, LM_PPTYPE_PATH,
priority, true, max_char_wh_ratio_, word_res);
} else if (debug_level_ > 3) {
tprintf("NO pain point (Classified) for col=%d row=%d type=%s\n",
pain_coord.col, pain_coord.row,
LMPainPointsTypeName[LM_PPTYPE_PATH]);
BLOB_CHOICE_IT b_it(word_res->ratings->get(pain_coord.col,
pain_coord.row));
for (b_it.mark_cycle_pt(); !b_it.cycled_list(); b_it.forward()) {
BLOB_CHOICE* choice = b_it.data();
choice->print_full();
}
}
curr_vse = parent_vse;
curr_b = curr_vse->curr_b;
}
}