// Returns the direction of the fitted line as a unit vector, using the
// least mean squared perpendicular distance. The line runs through the
// mean_point, i.e. a point p on the line is given by:
// p = mean_point() + lambda * vector_fit() for some real number lambda.
// Note that the result (0<=x<=1, -1<=y<=-1) is directionally ambiguous
// and may be negated without changing its meaning.
FCOORD LLSQ::vector_fit() const {
double x_var = x_variance();
double y_var = y_variance();
double covar = covariance();
FCOORD result;
if (x_var >= y_var) {
if (x_var == 0.0)
return FCOORD(0.0f, 0.0f);
result.set_x(x_var / sqrt(x_var * x_var + covar * covar));
result.set_y(sqrt(1.0 - result.x() * result.x()));
} else {
result.set_y(y_var / sqrt(y_var * y_var + covar * covar));
result.set_x(sqrt(1.0 - result.y() * result.y()));
}
if (covar < 0.0)
result.set_y(-result.y());
return result;
}
void POLY_BLOCK::rotate(FCOORD rotation) {
FCOORD pos; //current pos;
ICOORDELT *pt; //current point
ICOORDELT_IT pts = &vertices; //iterator
do {
pt = pts.data ();
pos.set_x (pt->x ());
pos.set_y (pt->y ());
pos.rotate (rotation);
pt->set_x ((inT16) (floor (pos.x () + 0.5)));
pt->set_y ((inT16) (floor (pos.y () + 0.5)));
pts.forward ();
}
while (!pts.at_first ());
compute_bb();
}
// Groups blocks by rotation, then, for each group, makes a WordGrid and calls
// TransferDiacriticsToWords to copy the diacritic blobs to the most
// appropriate words in the group of blocks. Source blobs are not touched.
void Textord::TransferDiacriticsToBlockGroups(BLOBNBOX_LIST* diacritic_blobs,
BLOCK_LIST* blocks) {
// Angle difference larger than this is too much to consider equal.
// They should only be in multiples of M_PI/2 anyway.
const double kMaxAngleDiff = 0.01; // About 0.6 degrees.
PointerVector<BlockGroup> groups;
BLOCK_IT bk_it(blocks);
for (bk_it.mark_cycle_pt(); !bk_it.cycled_list(); bk_it.forward()) {
BLOCK* block = bk_it.data();
if (block->pdblk.poly_block() != nullptr && !block->pdblk.poly_block()->IsText()) {
continue;
}
// Linear search of the groups to find a matching rotation.
float block_angle = block->re_rotation().angle();
int best_g = 0;
float best_angle_diff = MAX_FLOAT32;
for (int g = 0; g < groups.size(); ++g) {
double angle_diff = fabs(block_angle - groups[g]->angle);
if (angle_diff > M_PI) angle_diff = fabs(angle_diff - 2.0 * M_PI);
if (angle_diff < best_angle_diff) {
best_angle_diff = angle_diff;
best_g = g;
}
}
if (best_angle_diff > kMaxAngleDiff) {
groups.push_back(new BlockGroup(block));
} else {
groups[best_g]->blocks.push_back(block);
groups[best_g]->bounding_box += block->pdblk.bounding_box();
float x_height = block->x_height();
if (x_height < groups[best_g]->min_xheight)
groups[best_g]->min_xheight = x_height;
}
}
// Now process each group of blocks.
PointerVector<WordWithBox> word_ptrs;
for (int g = 0; g < groups.size(); ++g) {
const BlockGroup* group = groups[g];
if (group->bounding_box.null_box()) continue;
WordGrid word_grid(group->min_xheight, group->bounding_box.botleft(),
group->bounding_box.topright());
for (int b = 0; b < group->blocks.size(); ++b) {
ROW_IT row_it(group->blocks[b]->row_list());
for (row_it.mark_cycle_pt(); !row_it.cycled_list(); row_it.forward()) {
ROW* row = row_it.data();
// Put the words of the row into the grid.
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();
WordWithBox* box_word = new WordWithBox(word);
word_grid.InsertBBox(true, true, box_word);
// Save the pointer where it will be auto-deleted.
word_ptrs.push_back(box_word);
}
}
}
FCOORD rotation = group->rotation;
// Make it a forward rotation that will transform blob coords to block.
rotation.set_y(-rotation.y());
TransferDiacriticsToWords(diacritic_blobs, rotation, &word_grid);
}
}