// Computes the center of mass and second moments for the old baseline and
// 2nd moment normalizations. Returns the outline length.
// The input denorm should be the normalizations that have been applied from
// the image to the current state of this TBLOB.
int TBLOB::ComputeMoments(FCOORD* center, FCOORD* second_moments) const {
// Compute 1st and 2nd moments of the original outline.
LLSQ accumulator;
TBOX box = bounding_box();
// Iterate the outlines, accumulating edges relative the box.botleft().
CollectEdges(box, NULL, &accumulator, NULL, NULL);
*center = accumulator.mean_point() + box.botleft();
// The 2nd moments are just the standard deviation of the point positions.
double x2nd = sqrt(accumulator.x_variance());
double y2nd = sqrt(accumulator.y_variance());
if (x2nd < 1.0) x2nd = 1.0;
if (y2nd < 1.0) y2nd = 1.0;
second_moments->set_x(x2nd);
second_moments->set_y(y2nd);
return accumulator.count();
}
// Helper returns the mean direction vector from the given stats. Use the
// mean direction from dirs if there is information available, otherwise, use
// the fit_vector from point_diffs.
static FCOORD MeanDirectionVector(const LLSQ& point_diffs, const LLSQ& dirs,
const FCOORD& start_pt,
const FCOORD& end_pt) {
FCOORD fit_vector;
if (dirs.count() > 0) {
// There were directions, so use them. To avoid wrap-around problems, we
// have 2 accumulators in dirs: x for normal directions and y for
// directions offset by 128. We will use the one with the least variance.
FCOORD mean_pt = dirs.mean_point();
double mean_dir = 0.0;
if (dirs.x_variance() <= dirs.y_variance()) {
mean_dir = mean_pt.x();
} else {
mean_dir = mean_pt.y() + 128;
}
fit_vector.from_direction(Modulo(IntCastRounded(mean_dir), 256));
} else {
// There were no directions, so we rely on the vector_fit to the points.
// Since the vector_fit is 180 degrees ambiguous, we align with the
// supplied feature_dir by making the scalar product non-negative.
FCOORD feature_dir(end_pt - start_pt);
fit_vector = point_diffs.vector_fit();
if (fit_vector.x() == 0.0f && fit_vector.y() == 0.0f) {
// There was only a single point. Use feature_dir directly.
fit_vector = feature_dir;
} else {
// Sometimes the least mean squares fit is wrong, due to the small sample
// of points and scaling. Use a 90 degree rotated vector if that matches
// feature_dir better.
FCOORD fit_vector2 = !fit_vector;
// The fit_vector is 180 degrees ambiguous, so resolve the ambiguity by
// insisting that the scalar product with the feature_dir should be +ve.
if (fit_vector % feature_dir < 0.0)
fit_vector = -fit_vector;
if (fit_vector2 % feature_dir < 0.0)
fit_vector2 = -fit_vector2;
// Even though fit_vector2 has a higher mean squared error, it might be
// a better fit, so use it if the dot product with feature_dir is bigger.
if (fit_vector2 % feature_dir > fit_vector % feature_dir)
fit_vector = fit_vector2;
}
}
return fit_vector;
}
// Given an initial estimate of line spacing (m_in) and the positions of each
// baseline, computes the line spacing of the block more accurately in m_out,
// and the corresponding intercept in c_out, and the number of spacings seen
// in index_delta. Returns the error of fit to the line spacing model.
// Uses a simple linear regression, but optimized the offset using the median.
double BaselineBlock::FitLineSpacingModel(
const GenericVector<double>& positions, double m_in,
double* m_out, double* c_out, int* index_delta) {
if (m_in == 0.0f || positions.size() < 2) {
*m_out = m_in;
*c_out = 0.0;
if (index_delta != NULL) *index_delta = 0;
return 0.0;
}
GenericVector<double> offsets;
// Get the offset (remainder) linespacing for each line and choose the median.
for (int i = 0; i < positions.size(); ++i)
offsets.push_back(fmod(positions[i], m_in));
// Get the median offset.
double median_offset = MedianOfCircularValues(m_in, &offsets);
// Now fit a line to quantized line number and offset.
LLSQ llsq;
int min_index = MAX_INT32;
int max_index = -MAX_INT32;
for (int i = 0; i < positions.size(); ++i) {
double y_pos = positions[i];
int row_index = IntCastRounded((y_pos - median_offset) / m_in);
UpdateRange(row_index, &min_index, &max_index);
llsq.add(row_index, y_pos);
}
// Get the refined line spacing.
*m_out = llsq.m();
// Use the median offset rather than the mean.
offsets.truncate(0);
for (int i = 0; i < positions.size(); ++i)
offsets.push_back(fmod(positions[i], *m_out));
// Get the median offset.
if (debug_level_ > 2) {
for (int i = 0; i < offsets.size(); ++i)
tprintf("%d: %g\n", i, offsets[i]);
}
*c_out = MedianOfCircularValues(*m_out, &offsets);
if (debug_level_ > 1) {
tprintf("Median offset = %g, compared to mean of %g.\n",
*c_out, llsq.c(*m_out));
}
// Index_delta is the number of hypothesized line gaps present.
if (index_delta != NULL)
*index_delta = max_index - min_index;
// Use the regression model's intercept to compute the error, as it may be
// a full line-spacing in disagreement with the median.
double rms_error = llsq.rms(*m_out, llsq.c(*m_out));
if (debug_level_ > 1) {
tprintf("Linespacing of y=%g x + %g improved to %g x + %g, rms=%g\n",
m_in, median_offset, *m_out, *c_out, rms_error);
}
return rms_error;
}
// Fits a straight baseline to the points. Returns true if it had enough
// points to be reasonably sure of the fitted baseline.
// If use_box_bottoms is false, baselines positions are formed by
// considering the outlines of the blobs.
bool BaselineRow::FitBaseline(bool use_box_bottoms) {
// Deterministic fitting is used wherever possible.
fitter_.Clear();
// Linear least squares is a backup if the DetLineFit produces a bad line.
LLSQ llsq;
BLOBNBOX_IT blob_it(blobs_);
for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) {
BLOBNBOX* blob = blob_it.data();
if (!use_box_bottoms) blob->EstimateBaselinePosition();
const TBOX& box = blob->bounding_box();
int x_middle = (box.left() + box.right()) / 2;
#ifdef kDebugYCoord
if (box.bottom() < kDebugYCoord && box.top() > kDebugYCoord) {
tprintf("Box bottom = %d, baseline pos=%d for box at:",
box.bottom(), blob->baseline_position());
box.print();
}
#endif
fitter_.Add(ICOORD(x_middle, blob->baseline_position()), box.width() / 2);
llsq.add(x_middle, blob->baseline_position());
}
// Fit the line.
ICOORD pt1, pt2;
baseline_error_ = fitter_.Fit(&pt1, &pt2);
baseline_pt1_ = pt1;
baseline_pt2_ = pt2;
if (baseline_error_ > max_baseline_error_ &&
fitter_.SufficientPointsForIndependentFit()) {
// The fit was bad but there were plenty of points, so try skipping
// the first and last few, and use the new line if it dramatically improves
// the error of fit.
double error = fitter_.Fit(kNumSkipPoints, kNumSkipPoints, &pt1, &pt2);
if (error < baseline_error_ / 2.0) {
baseline_error_ = error;
baseline_pt1_ = pt1;
baseline_pt2_ = pt2;
}
}
int debug = 0;
#ifdef kDebugYCoord
Print();
debug = bounding_box_.bottom() < kDebugYCoord &&
bounding_box_.top() > kDebugYCoord
? 3 : 2;
#endif
// Now we obtained a direction from that fit, see if we can improve the
// fit using the same direction and some other start point.
FCOORD direction(pt2 - pt1);
double target_offset = direction * pt1;
good_baseline_ = false;
FitConstrainedIfBetter(debug, direction, 0.0, target_offset);
// Wild lines can be produced because DetLineFit allows vertical lines, but
// vertical text has been rotated so angles over pi/4 should be disallowed.
// Near vertical lines can still be produced by vertically aligned components
// on very short lines.
double angle = BaselineAngle();
if (fabs(angle) > M_PI * 0.25) {
// Use the llsq fit as a backup.
baseline_pt1_ = llsq.mean_point();
baseline_pt2_ = baseline_pt1_ + FCOORD(1.0f, llsq.m());
// TODO(rays) get rid of this when m and c are no longer used.
double m = llsq.m();
double c = llsq.c(m);
baseline_error_ = llsq.rms(m, c);
good_baseline_ = false;
}
return good_baseline_;
}
// Extracts Tesseract features and appends them to the features vector.
// Startpt to lastpt, inclusive, MUST have the same src_outline member,
// which may be nullptr. The vector from lastpt to its next is included in
// the feature extraction. Hidden edges should be excluded by the caller.
// If force_poly is true, the features will be extracted from the polygonal
// approximation even if more accurate data is available.
static void ExtractFeaturesFromRun(
const EDGEPT* startpt, const EDGEPT* lastpt,
const DENORM& denorm, double feature_length, bool force_poly,
GenericVector<INT_FEATURE_STRUCT>* features) {
const EDGEPT* endpt = lastpt->next;
const C_OUTLINE* outline = startpt->src_outline;
if (outline != nullptr && !force_poly) {
// Detailed information is available. We have to normalize only from
// the root_denorm to denorm.
const DENORM* root_denorm = denorm.RootDenorm();
int total_features = 0;
// Get the features from the outline.
int step_length = outline->pathlength();
int start_index = startpt->start_step;
// pos is the integer coordinates of the binary image steps.
ICOORD pos = outline->position_at_index(start_index);
// We use an end_index that allows us to use a positive increment, but that
// may be beyond the bounds of the outline steps/ due to wrap-around, to
// so we use % step_length everywhere, except for start_index.
int end_index = lastpt->start_step + lastpt->step_count;
if (end_index <= start_index)
end_index += step_length;
LLSQ prev_points;
LLSQ prev_dirs;
FCOORD prev_normed_pos = outline->sub_pixel_pos_at_index(pos, start_index);
denorm.NormTransform(root_denorm, prev_normed_pos, &prev_normed_pos);
LLSQ points;
LLSQ dirs;
FCOORD normed_pos(0.0f, 0.0f);
int index = GatherPoints(outline, feature_length, denorm, root_denorm,
start_index, end_index, &pos, &normed_pos,
&points, &dirs);
while (index <= end_index) {
// At each iteration we nominally have 3 accumulated sets of points and
// dirs: prev_points/dirs, points/dirs, next_points/dirs and sum them
// into sum_points/dirs, but we don't necessarily get any features out,
// so if that is the case, we keep accumulating instead of rotating the
// accumulators.
LLSQ next_points;
LLSQ next_dirs;
FCOORD next_normed_pos(0.0f, 0.0f);
index = GatherPoints(outline, feature_length, denorm, root_denorm,
index, end_index, &pos, &next_normed_pos,
&next_points, &next_dirs);
LLSQ sum_points(prev_points);
// TODO(rays) find out why it is better to use just dirs and next_dirs
// in sum_dirs, instead of using prev_dirs as well.
LLSQ sum_dirs(dirs);
sum_points.add(points);
sum_points.add(next_points);
sum_dirs.add(next_dirs);
bool made_features = false;
// If we have some points, we can try making some features.
if (sum_points.count() > 0) {
// We have gone far enough from the start. Make a feature and restart.
FCOORD fit_pt = sum_points.mean_point();
FCOORD fit_vector = MeanDirectionVector(sum_points, sum_dirs,
prev_normed_pos, normed_pos);
// The segment to which we fit features is the line passing through
// fit_pt in direction of fit_vector that starts nearest to
// prev_normed_pos and ends nearest to normed_pos.
FCOORD start_pos = prev_normed_pos.nearest_pt_on_line(fit_pt,
fit_vector);
FCOORD end_pos = normed_pos.nearest_pt_on_line(fit_pt, fit_vector);
// Possible correction to match the adjacent polygon segment.
if (total_features == 0 && startpt != endpt) {
FCOORD poly_pos(startpt->pos.x, startpt->pos.y);
denorm.LocalNormTransform(poly_pos, &start_pos);
}
if (index > end_index && startpt != endpt) {
FCOORD poly_pos(endpt->pos.x, endpt->pos.y);
denorm.LocalNormTransform(poly_pos, &end_pos);
}
int num_features = ComputeFeatures(start_pos, end_pos, feature_length,
features);
if (num_features > 0) {
// We made some features so shuffle the accumulators.
prev_points = points;
prev_dirs = dirs;
prev_normed_pos = normed_pos;
points = next_points;
dirs = next_dirs;
made_features = true;
total_features += num_features;
}
// The end of the next set becomes the end next time around.
normed_pos = next_normed_pos;
}
if (!made_features) {
// We didn't make any features, so keep the prev accumulators and
// add the next ones into the current.
points.add(next_points);
dirs.add(next_dirs);
}
}
} else {
// There is no outline, so we are forced to use the polygonal approximation.
const EDGEPT* pt = startpt;
do {
FCOORD start_pos(pt->pos.x, pt->pos.y);
FCOORD end_pos(pt->next->pos.x, pt->next->pos.y);
denorm.LocalNormTransform(start_pos, &start_pos);
denorm.LocalNormTransform(end_pos, &end_pos);
ComputeFeatures(start_pos, end_pos, feature_length, features);
} while ((pt = pt->next) != endpt);
}
}
