float blob_noise_score(PBLOB *blob) {
OUTLINE_IT outline_it;
TBOX box; //BB of outline
inT16 outline_count = 0;
inT16 max_dimension;
inT16 largest_outline_dimension = 0;
outline_it.set_to_list (blob->out_list ());
for (outline_it.mark_cycle_pt ();
!outline_it.cycled_list (); outline_it.forward ()) {
outline_count++;
box = outline_it.data ()->bounding_box ();
if (box.height () > box.width ())
max_dimension = box.height ();
else
max_dimension = box.width ();
if (largest_outline_dimension < max_dimension)
largest_outline_dimension = max_dimension;
}
if (fixsp_noise_score_fixing) {
if (outline_count > 5)
//penalise LOTS of blobs
largest_outline_dimension *= 2;
box = blob->bounding_box ();
if ((box.bottom () > bln_baseline_offset * 4) ||
(box.top () < bln_baseline_offset / 2))
//Lax blob is if high or low
largest_outline_dimension /= 2;
}
return largest_outline_dimension;
}
float Tesseract::blob_noise_score(TBLOB *blob) {
TBOX box; // BB of outline
inT16 outline_count = 0;
inT16 max_dimension;
inT16 largest_outline_dimension = 0;
for (TESSLINE* ol = blob->outlines; ol != NULL; ol= ol->next) {
outline_count++;
box = ol->bounding_box();
if (box.height() > box.width()) {
max_dimension = box.height();
} else {
max_dimension = box.width();
}
if (largest_outline_dimension < max_dimension)
largest_outline_dimension = max_dimension;
}
if (outline_count > 5) {
// penalise LOTS of blobs
largest_outline_dimension *= 2;
}
box = blob->bounding_box();
if (box.bottom() > kBlnBaselineOffset * 4 ||
box.top() < kBlnBaselineOffset / 2) {
// Lax blob is if high or low
largest_outline_dimension /= 2;
}
return largest_outline_dimension;
}
// Tests each blob in the list to see if it is certain non-text using 2
// conditions:
// 1. blob overlaps a cell with high value in noise_density_ (previously set
// by ComputeNoiseDensity).
// OR 2. The blob overlaps more than max_blob_overlaps in *this grid. This
// condition is disabled with max_blob_overlaps == -1.
// If it does, the blob is declared non-text, and is used to mark up the
// nontext_mask. Such blobs are fully deleted, and non-noise blobs have their
// neighbours reset, as they may now point to deleted data.
// WARNING: The blobs list blobs may be in the *this grid, but they are
// not removed. If any deleted blobs might be in *this, then this must be
// Clear()ed immediately after MarkAndDeleteNonTextBlobs is called.
// If the win is not NULL, deleted blobs are drawn on it in red, and kept
// blobs are drawn on it in ok_color.
void CCNonTextDetect::MarkAndDeleteNonTextBlobs(BLOBNBOX_LIST* blobs,
int max_blob_overlaps,
ScrollView* win,
ScrollView::Color ok_color,
Pix* nontext_mask) {
int imageheight = tright().y() - bleft().x();
BLOBNBOX_IT blob_it(blobs);
BLOBNBOX_LIST dead_blobs;
BLOBNBOX_IT dead_it(&dead_blobs);
for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) {
BLOBNBOX* blob = blob_it.data();
TBOX box = blob->bounding_box();
if (!noise_density_->RectMostlyOverThreshold(box, max_noise_count_) &&
(max_blob_overlaps < 0 ||
!BlobOverlapsTooMuch(blob, max_blob_overlaps))) {
blob->ClearNeighbours();
#ifndef GRAPHICS_DISABLED
if (win != NULL)
blob->plot(win, ok_color, ok_color);
#endif // GRAPHICS_DISABLED
} else {
if (noise_density_->AnyZeroInRect(box)) {
// There is a danger that the bounding box may overlap real text, so
// we need to render the outline.
Pix* blob_pix = blob->cblob()->render_outline();
pixRasterop(nontext_mask, box.left(), imageheight - box.top(),
box.width(), box.height(), PIX_SRC | PIX_DST,
blob_pix, 0, 0);
pixDestroy(&blob_pix);
} else {
if (box.area() < gridsize() * gridsize()) {
// It is a really bad idea to make lots of small components in the
// photo mask, so try to join it to a bigger area by expanding the
// box in a way that does not touch any zero noise density cell.
box = AttemptBoxExpansion(box, *noise_density_, gridsize());
}
// All overlapped cells are non-zero, so just mark the rectangle.
pixRasterop(nontext_mask, box.left(), imageheight - box.top(),
box.width(), box.height(), PIX_SET, NULL, 0, 0);
}
#ifndef GRAPHICS_DISABLED
if (win != NULL)
blob->plot(win, ScrollView::RED, ScrollView::RED);
#endif // GRAPHICS_DISABLED
// It is safe to delete the cblob now, as it isn't used by the grid
// or BlobOverlapsTooMuch, and the BLOBNBOXes will go away with the
// dead_blobs list.
// TODO(rays) delete the delete when the BLOBNBOX destructor deletes
// the cblob.
delete blob->cblob();
dead_it.add_to_end(blob_it.extract());
}
}
}
PBLOB::PBLOB( //constructor
C_BLOB *cblob, //compact blob
float xheight //height of line
) {
TBOX bbox; //bounding box
if (!cblob->out_list ()->empty ()) {
//get bounding box
bbox = cblob->bounding_box ();
if (bbox.height () > xheight)
xheight = bbox.height (); //max of line and blob
//copy it
approximate_outline_list (cblob->out_list (), &outlines, xheight);
}
}
// 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;
}
// Sets up the DENORM to execute a non-linear transformation based on
// preserving an even distribution of stroke edges. The transformation
// operates only within the given box.
// x_coords is a collection of the x-coords of vertical edges for each
// y-coord starting at box.bottom().
// y_coords is a collection of the y-coords of horizontal edges for each
// x-coord starting at box.left().
// Eg x_coords[0] is a collection of the x-coords of edges at y=bottom.
// Eg x_coords[1] is a collection of the x-coords of edges at y=bottom + 1.
// The second-level vectors must all be sorted in ascending order.
// See comments on the helper functions above for more details.
void DENORM::SetupNonLinear(
const DENORM* predecessor, const TBOX& box, float target_width,
float target_height, float final_xshift, float final_yshift,
const GenericVector<GenericVector<int> >& x_coords,
const GenericVector<GenericVector<int> >& y_coords) {
Clear();
predecessor_ = predecessor;
// x_map_ and y_map_ store a mapping from input x and y coordinate to output
// x and y coordinate, based on scaling to the supplied target_width and
// target_height.
x_map_ = new GenericVector<float>;
y_map_ = new GenericVector<float>;
// Set a 2-d image array to the run lengths at each pixel.
int width = box.width();
int height = box.height();
GENERIC_2D_ARRAY<int> minruns(width, height, 0);
ComputeRunlengthImage(box, x_coords, y_coords, &minruns);
// Edge density is the sum of the inverses of the run lengths. Compute
// edge density projection profiles.
ComputeEdgeDensityProfiles(box, minruns, x_map_, y_map_);
// Convert the edge density profiles to the coordinates by multiplying by
// the desired size and accumulating.
(*x_map_)[width] = target_width;
for (int x = width - 1; x >= 0; --x) {
(*x_map_)[x] = (*x_map_)[x + 1] - (*x_map_)[x] * target_width;
}
(*y_map_)[height] = target_height;
for (int y = height - 1; y >= 0; --y) {
(*y_map_)[y] = (*y_map_)[y + 1] - (*y_map_)[y] * target_height;
}
x_origin_ = box.left();
y_origin_ = box.bottom();
final_xshift_ = final_xshift;
final_yshift_ = final_yshift;
}
// Converts the run-length image (see above to the edge density profiles used
// for scaling, thus:
// ______________
// |7 1_1_1_1_1 7| = 5.28
// |1|5 5 1 5 5|1| = 3.8
// |1|2 2|1|2 2|1| = 5
// |1|2 2|1|2 2|1| = 5
// |1|2 2|1|2 2|1| = 5
// |1|2 2|1|2 2|1| = 5
// |1|5_5_1_5_5|1| = 3.8
// |7_1_1_1_1_1_7| = 5.28
// 6 4 4 8 4 4 6
// . . . . . . .
// 2 4 4 0 4 4 2
// 8 8
// Each profile is the sum of the reciprocals of the pixels in the image in
// the appropriate row or column, and these are then normalized to sum to 1.
// On output hx, hy contain an extra element, which will eventually be used
// to guarantee that the top/right edge of the box (and anything beyond) always
// gets mapped to the maximum target coordinate.
static void ComputeEdgeDensityProfiles(const TBOX& box,
const GENERIC_2D_ARRAY<int>& minruns,
GenericVector<float>* hx,
GenericVector<float>* hy) {
int width = box.width();
int height = box.height();
hx->init_to_size(width + 1, 0.0);
hy->init_to_size(height + 1, 0.0);
double total = 0.0;
for (int iy = 0; iy < height; ++iy) {
for (int ix = 0; ix < width; ++ix) {
int run = minruns(ix, iy);
if (run == 0) run = 1;
float density = 1.0f / run;
(*hx)[ix] += density;
(*hy)[iy] += density;
}
total += (*hy)[iy];
}
// Normalize each profile to sum to 1.
if (total > 0.0) {
for (int ix = 0; ix < width; ++ix) {
(*hx)[ix] /= total;
}
for (int iy = 0; iy < height; ++iy) {
(*hy)[iy] /= total;
}
}
// There is an extra element in each array, so initialize to 1.
(*hx)[width] = 1.0f;
(*hy)[height] = 1.0f;
}
// Adds edges to the given vectors.
// For all the edge steps in all the outlines, or polygonal approximation
// where there are no edge steps, collects the steps into x_coords/y_coords.
// x_coords is a collection of the x-coords of vertical edges for each
// y-coord starting at box.bottom().
// y_coords is a collection of the y-coords of horizontal edges for each
// x-coord starting at box.left().
// Eg x_coords[0] is a collection of the x-coords of edges at y=bottom.
// Eg x_coords[1] is a collection of the x-coords of edges at y=bottom + 1.
void TBLOB::GetEdgeCoords(const TBOX& box,
GenericVector<GenericVector<int> >* x_coords,
GenericVector<GenericVector<int> >* y_coords) const {
GenericVector<int> empty;
x_coords->init_to_size(box.height(), empty);
y_coords->init_to_size(box.width(), empty);
CollectEdges(box, nullptr, nullptr, x_coords, y_coords);
// Sort the output vectors.
for (int i = 0; i < x_coords->size(); ++i) (*x_coords)[i].sort();
for (int i = 0; i < y_coords->size(); ++i) (*y_coords)[i].sort();
}
/**********************************************************************
* char_box_to_tbox
*
* Create a TBOX from a character bounding box. If nonzero, the
* x_offset accounts for any additional padding of the word box that
* should be taken into account.
*
**********************************************************************/
TBOX char_box_to_tbox(Box* char_box, TBOX word_box, int x_offset) {
l_int32 left;
l_int32 top;
l_int32 width;
l_int32 height;
l_int32 right;
l_int32 bottom;
boxGetGeometry(char_box, &left, &top, &width, &height);
left += word_box.left() - x_offset;
right = left + width;
top = word_box.bottom() + word_box.height() - top;
bottom = top - height;
return TBOX(left, bottom, right, top);
}
// Helper for SetupNonLinear computes an image of shortest run-lengths from
// the x/y edges provided.
// Based on "A nonlinear normalization method for handprinted Kanji character
// recognition -- line density equalization" by Hiromitsu Yamada et al.
// Eg below is an O in a 1-pixel margin-ed bounding box and the corresponding
// ______________ input x_coords and y_coords.
// | _________ | <empty>
// | | _ | | 1, 6
// | | | | | | 1, 3, 4, 6
// | | | | | | 1, 3, 4, 6
// | | | | | | 1, 3, 4, 6
// | | |_| | | 1, 3, 4, 6
// | |_________| | 1, 6
// |_____________| <empty>
// E 1 1 1 1 1 E
// m 7 7 2 7 7 m
// p 6 p
// t 7 t
// y y
// The output image contains the min of the x and y run-length (distance
// between edges) at each coordinate in the image thus:
// ______________
// |7 1_1_1_1_1 7|
// |1|5 5 1 5 5|1|
// |1|2 2|1|2 2|1|
// |1|2 2|1|2 2|1|
// |1|2 2|1|2 2|1|
// |1|2 2|1|2 2|1|
// |1|5_5_1_5_5|1|
// |7_1_1_1_1_1_7|
// Note that the input coords are all integer, so all partial pixels are dealt
// with elsewhere. Although it is nice for outlines to be properly connected
// and continuous, there is no requirement that they be as such, so they could
// have been derived from a flaky source, such as greyscale.
// This function works only within the provided box, and it is assumed that the
// input x_coords and y_coords have already been translated to have the bottom-
// left of box as the origin. Although an output, the minruns should have been
// pre-initialized to be the same size as box. Each element will contain the
// minimum of x and y run-length as shown above.
static void ComputeRunlengthImage(
const TBOX& box,
const GenericVector<GenericVector<int> >& x_coords,
const GenericVector<GenericVector<int> >& y_coords,
GENERIC_2D_ARRAY<int>* minruns) {
int width = box.width();
int height = box.height();
ASSERT_HOST(minruns->dim1() == width);
ASSERT_HOST(minruns->dim2() == height);
// Set a 2-d image array to the run lengths at each pixel.
for (int ix = 0; ix < width; ++ix) {
int y = 0;
for (int i = 0; i < y_coords[ix].size(); ++i) {
int y_edge = ClipToRange(y_coords[ix][i], 0, height);
int gap = y_edge - y;
// Every pixel between the last and current edge get set to the gap.
while (y < y_edge) {
(*minruns)(ix, y) = gap;
++y;
}
}
// Pretend there is a bounding box of edges all around the image.
int gap = height - y;
while (y < height) {
(*minruns)(ix, y) = gap;
++y;
}
}
// Now set the image pixels the the MIN of the x and y runlengths.
for (int iy = 0; iy < height; ++iy) {
int x = 0;
for (int i = 0; i < x_coords[iy].size(); ++i) {
int x_edge = ClipToRange(x_coords[iy][i], 0, width);
int gap = x_edge - x;
while (x < x_edge) {
if (gap < (*minruns)(x, iy))
(*minruns)(x, iy) = gap;
++x;
}
}
int gap = width - x;
while (x < width) {
if (gap < (*minruns)(x, iy))
(*minruns)(x, iy) = gap;
++x;
}
}
}
// Given an input pix, and a box, the sides of the box are shrunk inwards until
// they bound any black pixels found within the original box.
// The function converts between tesseract coords and the pix coords assuming
// that this pix is full resolution equal in size to the original image.
// Returns an empty box if there are no black pixels in the source box.
static TBOX BoundsWithinBox(Pix* pix, const TBOX& box) {
int im_height = pixGetHeight(pix);
Box* input_box = boxCreate(box.left(), im_height - box.top(),
box.width(), box.height());
Box* output_box = NULL;
pixClipBoxToForeground(pix, input_box, NULL, &output_box);
TBOX result_box;
if (output_box != NULL) {
l_int32 x, y, width, height;
boxGetGeometry(output_box, &x, &y, &width, &height);
result_box.set_left(x);
result_box.set_right(x + width);
result_box.set_top(im_height - y);
result_box.set_bottom(result_box.top() - height);
boxDestroy(&output_box);
}
boxDestroy(&input_box);
return result_box;
}
BOOL8 suspect_fullstop(WERD_RES *word, inT16 i) {
float aspect_ratio;
PBLOB_LIST *blobs = word->outword->blob_list ();
PBLOB_IT blob_it(blobs);
inT16 j;
TBOX box;
inT16 width;
inT16 height;
for (j = 0; j < i; j++)
blob_it.forward ();
box = blob_it.data ()->bounding_box ();
width = box.width ();
height = box.height ();
aspect_ratio = ((width > height) ? ((float) width) / height :
((float) height) / width);
return (aspect_ratio > tessed_fullstop_aspect_ratio);
}
/**********************************************************************
* make_rotated_tess_blob
*
* Make a single Tess style blob, applying the given rotation and
* renormalizing.
**********************************************************************/
TBLOB *make_rotated_tess_blob(const DENORM* denorm, PBLOB *blob,
BOOL8 flatten) {
if (denorm != NULL && denorm->block() != NULL &&
denorm->block()->classify_rotation().y() != 0.0) {
TBOX box = blob->bounding_box();
int src_width = box.width();
int src_height = box.height();
src_width = static_cast<int>(src_width / denorm->scale() + 0.5);
src_height = static_cast<int>(src_height / denorm->scale() + 0.5);
int x_middle = (box.left() + box.right()) / 2;
int y_middle = (box.top() + box.bottom()) / 2;
PBLOB* rotated_blob = PBLOB::deep_copy(blob);
rotated_blob->move(FCOORD(-x_middle, -y_middle));
rotated_blob->rotate(denorm->block()->classify_rotation());
ICOORD median_size = denorm->block()->median_size();
int tolerance = median_size.x() / 8;
// TODO(dsl/rays) find a better normalization solution. In the mean time
// make it work for CJK by normalizing for Cap height in the same way
// as is applied in compute_block_xheight when the row is presumed to
// be ALLCAPS, i.e. the x-height is the fixed fraction
// blob height * textord_merge_x / (textord_merge_x + textord_merge_asc)
if (NearlyEqual(src_width, static_cast<int>(median_size.x()), tolerance) &&
NearlyEqual(src_height, static_cast<int>(median_size.y()), tolerance)) {
float target_height = bln_x_height * (textord_merge_x + textord_merge_asc)
/ textord_merge_x;
rotated_blob->scale(target_height / box.width());
rotated_blob->move(FCOORD(0.0f,
bln_baseline_offset -
rotated_blob->bounding_box().bottom()));
}
TBLOB* result = make_tess_blob(rotated_blob, flatten);
delete rotated_blob;
return result;
} else {
return make_tess_blob(blob, flatten);
}
}
// Fixes the block so it obeys all the rules:
// Must have at least one ROW.
// Must have at least one WERD.
// WERDs contain a fake blob.
void Textord::cleanup_nontext_block(BLOCK* block) {
// Non-text blocks must contain at least one row.
ROW_IT row_it(block->row_list());
if (row_it.empty()) {
TBOX box = block->bounding_box();
float height = box.height();
inT32 xstarts[2] = {box.left(), box.right()};
double coeffs[3] = {0.0, 0.0, static_cast<double>(box.bottom())};
ROW* row = new ROW(1, xstarts, coeffs, height / 2.0f, height / 4.0f,
height / 4.0f, 0, 1);
row_it.add_after_then_move(row);
}
// Each row must contain at least one word.
for (row_it.mark_cycle_pt(); !row_it.cycled_list(); row_it.forward()) {
ROW* row = row_it.data();
WERD_IT w_it(row->word_list());
if (w_it.empty()) {
// Make a fake blob to put in the word.
TBOX box = block->row_list()->singleton() ? block->bounding_box()
: row->bounding_box();
C_BLOB* blob = C_BLOB::FakeBlob(box);
C_BLOB_LIST blobs;
C_BLOB_IT blob_it(&blobs);
blob_it.add_after_then_move(blob);
WERD* word = new WERD(&blobs, 0, NULL);
w_it.add_after_then_move(word);
}
// Each word must contain a fake blob.
for (w_it.mark_cycle_pt(); !w_it.cycled_list(); w_it.forward()) {
WERD* word = w_it.data();
// Just assert that this is true, as it would be useful to find
// out why it isn't.
ASSERT_HOST(!word->cblob_list()->empty());
}
row->recalc_bounding_box();
}
}
// Returns true if the blob is small enough to be a large speckle.
bool Classify::LargeSpeckle(const TBLOB &blob) {
double speckle_size = kBlnXHeight * speckle_large_max_size;
TBOX bbox = blob.bounding_box();
return bbox.width() < speckle_size && bbox.height() < speckle_size;
}
/**
* This routine returns TRUE if both the width of height
* of Blob are less than the MaxLargeSpeckleSize.
*
* Globals:
* - #speckle_large_max_size largest allowed speckle
*
* Exceptions: none
* History: Mon Mar 11 10:06:49 1991, DSJ, Created.
*
* @param blob blob to test against speckle criteria
*
* @return TRUE if blob is speckle, FALSE otherwise.
*/
BOOL8 LargeSpeckle(TBLOB *blob) {
double speckle_size = BASELINE_SCALE * speckle_large_max_size;
TBOX bbox = blob->bounding_box();
return (bbox.width() < speckle_size && bbox.height() < speckle_size);
} /* LargeSpeckle */
void char_clip_word( //
WERD *word, //word to be processed
IMAGE &bin_image, //whole image
PIXROW_LIST *&pixrow_list, //pixrows built
IMAGELINE *&imlines, //lines cut from image
TBOX &pix_box //box defining imlines
) {
TBOX word_box = word->bounding_box ();
PBLOB_LIST *blob_list;
PBLOB_IT blob_it;
PIXROW_IT pixrow_it;
inT16 pix_offset; //Y pos of pixrow[0]
inT16 row_height; //No of pix rows
inT16 imlines_x_offset;
PIXROW *prev;
PIXROW *next;
PIXROW *current;
BOOL8 changed; //still improving
BOOL8 just_changed; //still improving
inT16 iteration_count = 0;
inT16 foreground_colour;
if (word->flag (W_INVERSE))
foreground_colour = 1;
else
foreground_colour = 0;
/* Define region for max pixrow expansion */
pix_box = word_box;
pix_box.move_bottom_edge (-pix_word_margin);
pix_box.move_top_edge (pix_word_margin);
pix_box.move_left_edge (-pix_word_margin);
pix_box.move_right_edge (pix_word_margin);
pix_box -= TBOX (ICOORD (0, 0 + BUG_OFFSET),
ICOORD (bin_image.get_xsize (),
bin_image.get_ysize () - BUG_OFFSET));
/* Generate pixrows list */
pix_offset = pix_box.bottom ();
row_height = pix_box.height ();
blob_list = word->blob_list ();
blob_it.set_to_list (blob_list);
pixrow_list = new PIXROW_LIST;
pixrow_it.set_to_list (pixrow_list);
for (blob_it.mark_cycle_pt (); !blob_it.cycled_list (); blob_it.forward ()) {
PIXROW *row = new PIXROW (pix_offset, row_height, blob_it.data ());
ASSERT_HOST (!row->
bad_box (bin_image.get_xsize (), bin_image.get_ysize ()));
pixrow_it.add_after_then_move (row);
}
imlines = generate_imlines (bin_image, pix_box);
/* Contract pixrows - shrink min and max back to black pixels */
imlines_x_offset = pix_box.left ();
pixrow_it.move_to_first ();
for (pixrow_it.mark_cycle_pt ();
!pixrow_it.cycled_list (); pixrow_it.forward ()) {
ASSERT_HOST (!pixrow_it.data ()->
bad_box (bin_image.get_xsize (), bin_image.get_ysize ()));
pixrow_it.data ()->contract (imlines, imlines_x_offset,
foreground_colour);
ASSERT_HOST (!pixrow_it.data ()->
bad_box (bin_image.get_xsize (), bin_image.get_ysize ()));
}
/* Expand pixrows iteratively 1 pixel at a time */
do {
changed = FALSE;
pixrow_it.move_to_first ();
prev = NULL;
current = NULL;
next = pixrow_it.data ();
for (pixrow_it.mark_cycle_pt ();
!pixrow_it.cycled_list (); pixrow_it.forward ()) {
prev = current;
current = next;
if (pixrow_it.at_last ())
next = NULL;
else
next = pixrow_it.data_relative (1);
just_changed = current->extend (imlines, pix_box, prev, next,
foreground_colour);
ASSERT_HOST (!current->
bad_box (bin_image.get_xsize (),
bin_image.get_ysize ()));
changed = changed || just_changed;
}
iteration_count++;
}
while (changed);
}
// Search vertically for a blob that is aligned with the input bbox.
// The search parameters are determined by AlignedBlobParams.
// top_to_bottom tells whether to search down or up.
// The return value is nullptr if nothing was found in the search box
// or if a blob was found in the gutter. On a nullptr return, end_y
// is set to the edge of the search box or the leading edge of the
// gutter blob if one was found.
BLOBNBOX* AlignedBlob::FindAlignedBlob(const AlignedBlobParams& p,
bool top_to_bottom, BLOBNBOX* bbox,
int x_start, int* end_y) {
TBOX box = bbox->bounding_box();
// If there are separator lines, get the column edges.
int left_column_edge = bbox->left_rule();
int right_column_edge = bbox->right_rule();
// start_y is used to guarantee that forward progress is made and the
// search does not go into an infinite loop. New blobs must extend the
// line beyond start_y.
int start_y = top_to_bottom ? box.bottom() : box.top();
if (WithinTestRegion(2, x_start, start_y)) {
tprintf("Column edges for blob at (%d,%d)->(%d,%d) are [%d, %d]\n",
box.left(), box.top(), box.right(), box.bottom(),
left_column_edge, right_column_edge);
}
// Compute skew tolerance.
int skew_tolerance = p.max_v_gap / kMaxSkewFactor;
// Calculate xmin and xmax of the search box so that it contains
// all possibly relevant boxes up to p.max_v_gap above or below accoording
// to top_to_bottom.
// Start with a notion of vertical with the current estimate.
int x2 = (p.max_v_gap * p.vertical.x() + p.vertical.y()/2) / p.vertical.y();
if (top_to_bottom) {
x2 = x_start - x2;
*end_y = start_y - p.max_v_gap;
} else {
x2 = x_start + x2;
*end_y = start_y + p.max_v_gap;
}
// Expand the box by an additional skew tolerance
int xmin = std::min(x_start, x2) - skew_tolerance;
int xmax = std::max(x_start, x2) + skew_tolerance;
// Now add direction-specific tolerances.
if (p.right_tab) {
xmax += p.min_gutter;
xmin -= p.l_align_tolerance;
} else {
xmax += p.r_align_tolerance;
xmin -= p.min_gutter;
}
// Setup a vertical search for an aligned blob.
GridSearch<BLOBNBOX, BLOBNBOX_CLIST, BLOBNBOX_C_IT> vsearch(this);
if (WithinTestRegion(2, x_start, start_y))
tprintf("Starting %s %s search at %d-%d,%d, search_size=%d, gutter=%d\n",
p.ragged ? "Ragged" : "Aligned", p.right_tab ? "Right" : "Left",
xmin, xmax, start_y, p.max_v_gap, p.min_gutter);
vsearch.StartVerticalSearch(xmin, xmax, start_y);
// result stores the best real return value.
BLOBNBOX* result = nullptr;
// The backup_result is not a tab candidate and can be used if no
// real tab candidate result is found.
BLOBNBOX* backup_result = nullptr;
// neighbour is the blob that is currently being investigated.
BLOBNBOX* neighbour = nullptr;
while ((neighbour = vsearch.NextVerticalSearch(top_to_bottom)) != nullptr) {
if (neighbour == bbox)
continue;
TBOX nbox = neighbour->bounding_box();
int n_y = (nbox.top() + nbox.bottom()) / 2;
if ((!top_to_bottom && n_y > start_y + p.max_v_gap) ||
(top_to_bottom && n_y < start_y - p.max_v_gap)) {
if (WithinTestRegion(2, x_start, start_y))
tprintf("Neighbour too far at (%d,%d)->(%d,%d)\n",
nbox.left(), nbox.bottom(), nbox.right(), nbox.top());
break; // Gone far enough.
}
// It is CRITICAL to ensure that forward progress is made, (strictly
// in/decreasing n_y) or the caller could loop infinitely, while
// waiting for a sequence of blobs in a line to end.
// NextVerticalSearch alone does not guarantee this, as there may be
// more than one blob in a grid cell. See comment in AlignTabs.
if ((n_y < start_y) != top_to_bottom || nbox.y_overlap(box))
continue; // Only look in the required direction.
if (result != nullptr && result->bounding_box().y_gap(nbox) > gridsize())
return result; // This result is clear.
if (backup_result != nullptr && p.ragged && result == nullptr &&
backup_result->bounding_box().y_gap(nbox) > gridsize())
return backup_result; // This result is clear.
// If the neighbouring blob is the wrong side of a separator line, then it
// "doesn't exist" as far as we are concerned.
int x_at_n_y = x_start + (n_y - start_y) * p.vertical.x() / p.vertical.y();
if (x_at_n_y < neighbour->left_crossing_rule() ||
x_at_n_y > neighbour->right_crossing_rule())
continue; // Separator line in the way.
int n_left = nbox.left();
int n_right = nbox.right();
int n_x = p.right_tab ? n_right : n_left;
if (WithinTestRegion(2, x_start, start_y))
tprintf("neighbour at (%d,%d)->(%d,%d), n_x=%d, n_y=%d, xatn=%d\n",
nbox.left(), nbox.bottom(), nbox.right(), nbox.top(),
n_x, n_y, x_at_n_y);
if (p.right_tab &&
n_left < x_at_n_y + p.min_gutter &&
n_right > x_at_n_y + p.r_align_tolerance &&
(p.ragged || n_left < x_at_n_y + p.gutter_fraction * nbox.height())) {
// In the gutter so end of line.
if (bbox->right_tab_type() >= TT_MAYBE_ALIGNED)
bbox->set_right_tab_type(TT_DELETED);
*end_y = top_to_bottom ? nbox.top() : nbox.bottom();
if (WithinTestRegion(2, x_start, start_y))
tprintf("gutter\n");
return nullptr;
}
if (!p.right_tab &&
n_left < x_at_n_y - p.l_align_tolerance &&
n_right > x_at_n_y - p.min_gutter &&
(p.ragged || n_right > x_at_n_y - p.gutter_fraction * nbox.height())) {
// In the gutter so end of line.
if (bbox->left_tab_type() >= TT_MAYBE_ALIGNED)
bbox->set_left_tab_type(TT_DELETED);
*end_y = top_to_bottom ? nbox.top() : nbox.bottom();
if (WithinTestRegion(2, x_start, start_y))
tprintf("gutter\n");
return nullptr;
}
if ((p.right_tab && neighbour->leader_on_right()) ||
(!p.right_tab && neighbour->leader_on_left()))
continue; // Neighbours of leaders are not allowed to be used.
if (n_x <= x_at_n_y + p.r_align_tolerance &&
n_x >= x_at_n_y - p.l_align_tolerance) {
// Aligned so keep it. If it is a marked tab save it as result,
// otherwise keep it as backup_result to return in case of later failure.
if (WithinTestRegion(2, x_start, start_y))
tprintf("aligned, seeking%d, l=%d, r=%d\n",
p.right_tab, neighbour->left_tab_type(),
neighbour->right_tab_type());
TabType n_type = p.right_tab ? neighbour->right_tab_type()
: neighbour->left_tab_type();
if (n_type != TT_NONE && (p.ragged || n_type != TT_MAYBE_RAGGED)) {
if (result == nullptr) {
result = neighbour;
} else {
// Keep the closest neighbour by Euclidean distance.
// This prevents it from picking a tab blob in another column.
const TBOX& old_box = result->bounding_box();
int x_diff = p.right_tab ? old_box.right() : old_box.left();
x_diff -= x_at_n_y;
int y_diff = (old_box.top() + old_box.bottom()) / 2 - start_y;
int old_dist = x_diff * x_diff + y_diff * y_diff;
x_diff = n_x - x_at_n_y;
y_diff = n_y - start_y;
int new_dist = x_diff * x_diff + y_diff * y_diff;
if (new_dist < old_dist)
result = neighbour;
}
} else if (backup_result == nullptr) {
if (WithinTestRegion(2, x_start, start_y))
tprintf("Backup\n");
backup_result = neighbour;
} else {
TBOX backup_box = backup_result->bounding_box();
if ((p.right_tab && backup_box.right() < nbox.right()) ||
(!p.right_tab && backup_box.left() > nbox.left())) {
if (WithinTestRegion(2, x_start, start_y))
tprintf("Better backup\n");
backup_result = neighbour;
}
}
}
}
return result != nullptr ? result : backup_result;
}
// Returns a Pix rendering of the blob. pixDestroy after use.
Pix* C_BLOB::render() {
TBOX box = bounding_box();
Pix* pix = pixCreate(box.width(), box.height(), 1);
render_outline_list(&outlines, box.left(), box.top(), pix);
return pix;
}
// Collects edges into the given bounding box, LLSQ accumulator and/or x_coords,
// y_coords vectors.
// For a description of x_coords/y_coords, see GetEdgeCoords above.
// Startpt to lastpt, inclusive, MUST have the same src_outline member,
// which may be NULL. The vector from lastpt to its next is included in
// the accumulation. Hidden edges should be excluded by the caller.
// The input denorm should be the normalizations that have been applied from
// the image to the current state of the TBLOB from which startpt, lastpt come.
// box is the bounding box of the blob from which the EDGEPTs are taken and
// indices into x_coords, y_coords are offset by box.botleft().
static void CollectEdgesOfRun(const EDGEPT* startpt, const EDGEPT* lastpt,
const DENORM& denorm, const TBOX& box,
TBOX* bounding_box,
LLSQ* accumulator,
GenericVector<GenericVector<int> > *x_coords,
GenericVector<GenericVector<int> > *y_coords) {
const C_OUTLINE* outline = startpt->src_outline;
int x_limit = box.width() - 1;
int y_limit = box.height() - 1;
if (outline != NULL) {
// Use higher-resolution edge points stored on the outline.
// The outline coordinates may not match the binary image because of the
// rotation for vertical text lines, but the root_denorm IS the matching
// start of the DENORM chain.
const DENORM* root_denorm = denorm.RootDenorm();
int step_length = outline->pathlength();
int start_index = startpt->start_step;
// Note that if this run straddles the wrap-around point of the outline,
// that lastpt->start_step may have a lower index than startpt->start_step,
// and we want to use an end_index that allows us to use a positive
// increment, so we add step_length if necessary, but that may be beyond the
// bounds of the outline steps/ due to wrap-around, 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;
// pos is the integer coordinates of the binary image steps.
ICOORD pos = outline->position_at_index(start_index);
FCOORD origin(box.left(), box.bottom());
// f_pos is a floating-point version of pos that offers improved edge
// positioning using greyscale information or smoothing of edge steps.
FCOORD f_pos = outline->sub_pixel_pos_at_index(pos, start_index);
// pos_normed is f_pos after the appropriate normalization, and relative
// to origin.
// prev_normed is the previous value of pos_normed.
FCOORD prev_normed;
denorm.NormTransform(root_denorm, f_pos, &prev_normed);
prev_normed -= origin;
for (int index = start_index; index < end_index; ++index) {
ICOORD step = outline->step(index % step_length);
// Only use the point if its edge strength is positive. This excludes
// points that don't provide useful information, eg
// ___________
// |___________
// The vertical step provides only noisy, damaging information, as even
// with a greyscale image, the positioning of the edge there may be a
// fictitious extrapolation, so previous processing has eliminated it.
if (outline->edge_strength_at_index(index % step_length) > 0) {
FCOORD f_pos = outline->sub_pixel_pos_at_index(pos,
index % step_length);
FCOORD pos_normed;
denorm.NormTransform(root_denorm, f_pos, &pos_normed);
pos_normed -= origin;
// Accumulate the information that is selected by the caller.
if (bounding_box != NULL) {
SegmentBBox(pos_normed, prev_normed, bounding_box);
}
if (accumulator != NULL) {
SegmentLLSQ(pos_normed, prev_normed, accumulator);
}
if (x_coords != NULL && y_coords != NULL) {
SegmentCoords(pos_normed, prev_normed, x_limit, y_limit,
x_coords, y_coords);
}
prev_normed = pos_normed;
}
pos += step;
}
} else {
// There is no outline, so we are forced to use the polygonal approximation.
const EDGEPT* endpt = lastpt->next;
const EDGEPT* pt = startpt;
do {
FCOORD next_pos(pt->next->pos.x - box.left(),
pt->next->pos.y - box.bottom());
FCOORD pos(pt->pos.x - box.left(), pt->pos.y - box.bottom());
if (bounding_box != NULL) {
SegmentBBox(next_pos, pos, bounding_box);
}
if (accumulator != NULL) {
SegmentLLSQ(next_pos, pos, accumulator);
}
if (x_coords != NULL && y_coords != NULL) {
SegmentCoords(next_pos, pos, x_limit, y_limit, x_coords, y_coords);
}
} while ((pt = pt->next) != endpt);
}
}
