// Compute the distance from the from_box to the to_box using curved
// projection space. Separation that involves a decrease in projection
// density (moving from the from_box to the to_box) is weighted more heavily
// than constant density, and an increase is weighted less.
// If horizontal_textline is true, then curved space is used vertically,
// as for a diacritic on the edge of a textline.
// The projection uses original image coords, so denorm is used to get
// back to the image coords from box/part space.
// How the calculation works: Think of a diacritic near a textline.
// Distance is measured from the far side of the from_box to the near side of
// the to_box. Shown is the horizontal textline case.
// |------^-----|
// | from | box |
// |------|-----|
// perpendicular |
// <------v-------->|--------------------|
// parallel | to box |
// |--------------------|
// Perpendicular distance uses "curved space" See VerticalDistance below.
// Parallel distance is linear.
// Result is perpendicular_gap + parallel_gap / kParaPerpDistRatio.
int TextlineProjection::DistanceOfBoxFromBox(const TBOX& from_box,
const TBOX& to_box,
bool horizontal_textline,
const DENORM* denorm,
bool debug) const {
// The parallel_gap is the horizontal gap between a horizontal textline and
// the box. Analogous for vertical.
int parallel_gap = 0;
// start_pt is the box end of the line to be modified for curved space.
TPOINT start_pt;
// end_pt is the partition end of the line to be modified for curved space.
TPOINT end_pt;
if (horizontal_textline) {
parallel_gap = from_box.x_gap(to_box) + from_box.width();
start_pt.x = (from_box.left() + from_box.right()) / 2;
end_pt.x = start_pt.x;
if (from_box.top() - to_box.top() >= to_box.bottom() - from_box.bottom()) {
start_pt.y = from_box.top();
end_pt.y = MIN(to_box.top(), start_pt.y);
} else {
start_pt.y = from_box.bottom();
end_pt.y = MAX(to_box.bottom(), start_pt.y);
}
} else {
parallel_gap = from_box.y_gap(to_box) + from_box.height();
if (from_box.right() - to_box.right() >= to_box.left() - from_box.left()) {
start_pt.x = from_box.right();
end_pt.x = MIN(to_box.right(), start_pt.x);
} else {
start_pt.x = from_box.left();
end_pt.x = MAX(to_box.left(), start_pt.x);
}
start_pt.y = (from_box.bottom() + from_box.top()) / 2;
end_pt.y = start_pt.y;
}
// The perpendicular gap is the max vertical distance gap out of:
// top of from_box to to_box top and bottom of from_box to to_box bottom.
// This value is then modified for curved projection space.
// Analogous for vertical.
int perpendicular_gap = 0;
// If start_pt == end_pt, then the from_box lies entirely within the to_box
// (in the perpendicular direction), so we don't need to calculate the
// perpendicular_gap.
if (start_pt.x != end_pt.x || start_pt.y != end_pt.y) {
if (denorm != NULL) {
// Denormalize the start and end.
denorm->DenormTransform(NULL, start_pt, &start_pt);
denorm->DenormTransform(NULL, end_pt, &end_pt);
}
if (abs(start_pt.y - end_pt.y) >= abs(start_pt.x - end_pt.x)) {
perpendicular_gap = VerticalDistance(debug, start_pt.x, start_pt.y,
end_pt.y);
} else {
perpendicular_gap = HorizontalDistance(debug, start_pt.x, end_pt.x,
start_pt.y);
}
}
// The parallel_gap weighs less than the perpendicular_gap.
return perpendicular_gap + parallel_gap / kParaPerpDistRatio;
}
/// Gather consecutive blobs that match the given box into the best_state
/// and corresponding correct_text.
///
/// Fights over which box owns which blobs are settled by pre-chopping and
/// 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 appropriate blob for a box.
///
/// This means that occasionally, blobs may be incorrectly segmented if the
/// chopper fails to find a suitable chop point.
bool Tesseract::ResegmentCharBox(PAGE_RES* page_res, const TBOX *prev_box,
const TBOX& box, const TBOX& next_box,
const char* correct_text) {
if (applybox_debug > 1) {
tprintf("\nAPPLY_BOX: in ResegmentCharBox() for %s\n", correct_text);
}
PAGE_RES_IT page_res_it(page_res);
WERD_RES* word_res;
for (word_res = page_res_it.word(); word_res != NULL;
word_res = page_res_it.forward()) {
if (!word_res->box_word->bounding_box().major_overlap(box))
continue;
if (applybox_debug > 1) {
tprintf("Checking word box:");
word_res->box_word->bounding_box().print();
}
int word_len = word_res->box_word->length();
for (int i = 0; i < word_len; ++i) {
TBOX char_box = TBOX();
int blob_count = 0;
for (blob_count = 0; i + blob_count < word_len; ++blob_count) {
TBOX blob_box = word_res->box_word->BlobBox(i + blob_count);
if (!blob_box.major_overlap(box))
break;
if (word_res->correct_text[i + blob_count].length() > 0)
break; // Blob is claimed already.
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)
break; // Blob is a better match for next box.
char_box += blob_box;
}
if (blob_count > 0) {
if (applybox_debug > 1) {
tprintf("Index [%d, %d) seem good.\n", i, i + blob_count);
}
if (!char_box.almost_equal(box, 3) &&
(box.x_gap(next_box) < -3 ||
(prev_box != NULL && prev_box->x_gap(box) < -3))) {
return false;
}
// We refine just the box_word, best_state and correct_text here.
// The rebuild_word is made in TidyUp.
// blob_count blobs are put together to match the box. Merge the
// box_word boxes, save the blob_count in the state and the text.
word_res->box_word->MergeBoxes(i, i + blob_count);
word_res->best_state[i] = blob_count;
word_res->correct_text[i] = correct_text;
if (applybox_debug > 2) {
tprintf("%d Blobs match: blob box:", blob_count);
word_res->box_word->BlobBox(i).print();
tprintf("Matches box:");
box.print();
tprintf("With next box:");
next_box.print();
}
// Eliminated best_state and correct_text entries for the consumed
// blobs.
for (int j = 1; j < blob_count; ++j) {
word_res->best_state.remove(i + 1);
word_res->correct_text.remove(i + 1);
}
// Assume that no box spans multiple source words, so we are done with
// this box.
if (applybox_debug > 1) {
tprintf("Best state = ");
for (int j = 0; j < word_res->best_state.size(); ++j) {
tprintf("%d ", word_res->best_state[j]);
}
tprintf("\n");
tprintf("Correct text = [[ ");
for (int j = 0; j < word_res->correct_text.size(); ++j) {
tprintf("%s ", word_res->correct_text[j].string());
}
tprintf("]]\n");
}
return true;
}
}
}
if (applybox_debug > 0) {
tprintf("FAIL!\n");
}
return false; // Failure.
}