// Internal version of EvaluateBox returns the unclipped gradients as well
// as the result of EvaluateBox.
// hgrad1 and hgrad2 are the gradients for the horizontal textline.
int TextlineProjection::EvaluateBoxInternal(const TBOX& box,
const DENORM* denorm, bool debug,
int* hgrad1, int* hgrad2,
int* vgrad1, int* vgrad2) const {
int top_gradient = BestMeanGradientInRow(denorm, box.left(), box.right(),
box.top(), true);
int bottom_gradient = -BestMeanGradientInRow(denorm, box.left(), box.right(),
box.bottom(), false);
int left_gradient = BestMeanGradientInColumn(denorm, box.left(), box.bottom(),
box.top(), true);
int right_gradient = -BestMeanGradientInColumn(denorm, box.right(),
box.bottom(), box.top(),
false);
int top_clipped = MAX(top_gradient, 0);
int bottom_clipped = MAX(bottom_gradient, 0);
int left_clipped = MAX(left_gradient, 0);
int right_clipped = MAX(right_gradient, 0);
if (debug) {
tprintf("Gradients: top = %d, bottom = %d, left= %d, right= %d for box:",
top_gradient, bottom_gradient, left_gradient, right_gradient);
box.print();
}
int result = MAX(top_clipped, bottom_clipped) -
MAX(left_clipped, right_clipped);
if (hgrad1 != NULL && hgrad2 != NULL) {
*hgrad1 = top_gradient;
*hgrad2 = bottom_gradient;
}
if (vgrad1 != NULL && vgrad2 != NULL) {
*vgrad1 = left_gradient;
*vgrad2 = right_gradient;
}
return result;
}
/**********************************************************************
* blobs_widths
*
* Compute the widths of a list of blobs. Return an array of the widths
* and gaps.
**********************************************************************/
WIDTH_RECORD *blobs_widths(TBLOB *blobs) { /*blob to compute on */
WIDTH_RECORD *width_record;
TPOINT topleft; /*bounding box */
TPOINT botright;
int i = 0;
int blob_end;
int num_blobs = count_blobs (blobs);
/* Get memory */
width_record = (WIDTH_RECORD *) memalloc (sizeof (int) * num_blobs * 2);
width_record->num_chars = num_blobs;
TBOX bbox = blobs->bounding_box();
width_record->widths[i++] = bbox.width();
/* First width */
blob_end = bbox.right();
for (TBLOB* blob = blobs->next; blob != NULL; blob = blob->next) {
TBOX curbox = blob->bounding_box();
width_record->widths[i++] = curbox.left() - blob_end;
width_record->widths[i++] = curbox.width();
blob_end = curbox.right();
}
return width_record;
}
// Find a set of blobs that are aligned in the given vertical
// direction with the given blob. Returns a list of aligned
// blobs and the number in the list.
// For other parameters see FindAlignedBlob below.
int AlignedBlob::AlignTabs(const AlignedBlobParams& params,
bool top_to_bottom, BLOBNBOX* bbox,
BLOBNBOX_CLIST* good_points, int* end_y) {
int ptcount = 0;
BLOBNBOX_C_IT it(good_points);
TBOX box = bbox->bounding_box();
int x_start = params.right_tab ? box.right() : box.left();
while (bbox != NULL) {
// Add the blob to the list if the appropriate side is a tab candidate,
// or if we are working on a ragged tab.
if (((params.right_tab && bbox->right_tab_type() != TT_NONE) ||
(!params.right_tab && bbox->left_tab_type() != TT_NONE) ||
params.ragged) &&
(it.empty() || it.data() != bbox)) {
if (top_to_bottom)
it.add_before_then_move(bbox);
else
it.add_after_then_move(bbox);
++ptcount;
}
// Find the next blob that is aligned with the current one.
// FindAlignedBlob guarantees that forward progress will be made in the
// top_to_bottom direction, and therefore eventually it will return NULL,
// making this while (bbox != NULL) loop safe.
bbox = FindAlignedBlob(params, top_to_bottom, bbox, x_start, end_y);
if (bbox != NULL) {
box = bbox->bounding_box();
if (!params.ragged)
x_start = params.right_tab ? box.right() : box.left();
}
}
return ptcount;
}
// 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;
}
QString printTBOX(TBOX box,int height, bool eol)
{
if (eol)
return QString ("Bounding box=(%1,%2)->(%3,%4)\n").arg(box.left())
.arg(height - box.top()).arg(box.right()).arg(height - box.bottom());
else
return QString ("Bounding box=(%1,%2)->(%3,%4)").arg(box.left())
.arg(height - box.top()).arg(box.right()).arg(height - box.bottom());
}
void make_illegal_segment( //find segmentation
FPSEGPT_LIST *prev_list, //previous segments
TBOX blob_box, //bounding box
BLOBNBOX_IT blob_it, //iterator
int16_t region_index, //number of segment
int16_t pitch, //pitch estimate
int16_t pitch_error, //tolerance
FPSEGPT_LIST *seg_list //output list
) {
int16_t x; //current coord
int16_t min_x = 0; //in this region
int16_t max_x = 0;
int16_t offset; //dist to edge
FPSEGPT *segpt; //segment point
FPSEGPT *prevpt; //previous point
float best_cost; //best path
FPSEGPT_IT segpt_it = seg_list;//iterator
//previous points
FPSEGPT_IT prevpt_it = prev_list;
best_cost = FLT_MAX;
for (prevpt_it.mark_cycle_pt (); !prevpt_it.cycled_list ();
prevpt_it.forward ()) {
prevpt = prevpt_it.data ();
if (prevpt->cost_function () < best_cost) {
//find least
best_cost = prevpt->cost_function ();
min_x = prevpt->position ();
max_x = min_x; //limits on coords
}
else if (prevpt->cost_function () == best_cost) {
max_x = prevpt->position ();
}
}
min_x += pitch - pitch_error;
max_x += pitch + pitch_error;
for (x = min_x; x <= max_x; x++) {
while (x > blob_box.right ()) {
blob_box = box_next (&blob_it);
}
offset = x - blob_box.left ();
if (blob_box.right () - x < offset)
offset = blob_box.right () - x;
segpt = new FPSEGPT (x, FALSE, offset,
region_index, pitch, pitch_error, prev_list);
if (segpt->previous () != nullptr) {
ASSERT_HOST (offset >= 0);
fprintf (stderr, "made fake at %d\n", x);
//make one up
segpt_it.add_after_then_move (segpt);
segpt->faked = TRUE;
segpt->fake_count++;
}
else
delete segpt;
}
}
bool PIXROW::bad_box( //return true if box exceeds image
int xsize,
int ysize) const {
TBOX bbox = bounding_box ();
if (bbox.left () < 0 || bbox.right () > xsize
|| bbox.top () > ysize || bbox.bottom () < 0) {
tprintf("Box (%d,%d)->(%d,%d) bad compared to %d,%d\n",
bbox.left(),bbox.bottom(), bbox.right(), bbox.top(),
xsize, ysize);
return true;
}
return false;
}
static void PrintBoxWidths(BLOBNBOX* neighbour) {
TBOX nbox = neighbour->bounding_box();
tprintf("Box (%d,%d)->(%d,%d): h-width=%.1f, v-width=%.1f p-width=%1.f\n",
nbox.left(), nbox.bottom(), nbox.right(), nbox.top(),
neighbour->horz_stroke_width(), neighbour->vert_stroke_width(),
2.0 * neighbour->cblob()->area()/neighbour->cblob()->perimeter());
}
// Inserts a list of blobs into the projection.
// Rotation is a multiple of 90 degrees to get from blob coords to
// nontext_map coords, nontext_map_box is the bounds of the nontext_map.
// Blobs are spread horizontally or vertically according to their internal
// flags, but the spreading is truncated by set pixels in the nontext_map
// and also by the horizontal rule line limits on the blobs.
void TextlineProjection::ProjectBlobs(BLOBNBOX_LIST* blobs,
const FCOORD& rotation,
const TBOX& nontext_map_box,
Pix* nontext_map) {
BLOBNBOX_IT blob_it(blobs);
for (blob_it.mark_cycle_pt(); !blob_it.cycled_list(); blob_it.forward()) {
BLOBNBOX* blob = blob_it.data();
TBOX bbox = blob->bounding_box();
ICOORD middle((bbox.left() + bbox.right()) / 2,
(bbox.bottom() + bbox.top()) / 2);
bool spreading_horizontally = PadBlobBox(blob, &bbox);
// Rotate to match the nontext_map.
bbox.rotate(rotation);
middle.rotate(rotation);
if (rotation.x() == 0.0f)
spreading_horizontally = !spreading_horizontally;
// Clip to the image before applying the increments.
bbox &= nontext_map_box; // This is in-place box intersection.
// Check for image pixels before spreading.
TruncateBoxToMissNonText(middle.x(), middle.y(), spreading_horizontally,
nontext_map, &bbox);
if (bbox.area() > 0) {
IncrementRectangle8Bit(bbox);
}
}
}
// Generates a TrainingSample from a TBLOB. Extracts features and sets
// the bounding box, so classifiers that operate on the image can work.
// TODO(rays) Make BlobToTrainingSample a member of Classify now that
// the FlexFx and FeatureDescription code have been removed and LearnBlob
// is now a member of Classify.
TrainingSample* BlobToTrainingSample(
const TBLOB& blob, bool nonlinear_norm, INT_FX_RESULT_STRUCT* fx_info,
GenericVector<INT_FEATURE_STRUCT>* bl_features) {
GenericVector<INT_FEATURE_STRUCT> cn_features;
Classify::ExtractFeatures(blob, nonlinear_norm, bl_features,
&cn_features, fx_info, nullptr);
// TODO(rays) Use blob->PreciseBoundingBox() instead.
TBOX box = blob.bounding_box();
TrainingSample* sample = nullptr;
int num_features = fx_info->NumCN;
if (num_features > 0) {
sample = TrainingSample::CopyFromFeatures(*fx_info, box, &cn_features[0],
num_features);
}
if (sample != nullptr) {
// Set the bounding box (in original image coordinates) in the sample.
TPOINT topleft, botright;
topleft.x = box.left();
topleft.y = box.top();
botright.x = box.right();
botright.y = box.bottom();
TPOINT original_topleft, original_botright;
blob.denorm().DenormTransform(nullptr, topleft, &original_topleft);
blob.denorm().DenormTransform(nullptr, botright, &original_botright);
sample->set_bounding_box(TBOX(original_topleft.x, original_botright.y,
original_botright.x, original_topleft.y));
}
return sample;
}
void OL_BUCKETS::extract_children( // recursive count
C_OUTLINE *outline, // parent outline
C_OUTLINE_IT *it // destination iterator
) {
inT16 xmin, xmax; // coord limits
inT16 ymin, ymax;
inT16 xindex, yindex; // current bucket
TBOX olbox;
C_OUTLINE_IT child_it; // search iterator
olbox = outline->bounding_box();
xmin =(olbox.left() - bl.x()) / BUCKETSIZE;
xmax =(olbox.right() - bl.x()) / BUCKETSIZE;
ymin =(olbox.bottom() - bl.y()) / BUCKETSIZE;
ymax =(olbox.top() - bl.y()) / BUCKETSIZE;
for (yindex = ymin; yindex <= ymax; yindex++) {
for (xindex = xmin; xindex <= xmax; xindex++) {
child_it.set_to_list(&buckets[yindex * bxdim + xindex]);
for (child_it.mark_cycle_pt(); !child_it.cycled_list();
child_it.forward()) {
if (*child_it.data() < *outline) {
it->add_after_then_move(child_it.extract());
}
}
}
}
}
/**********************************************************************
* render_segmentation
*
* Create a list of line segments that represent the list of chunks
* using the correct segmentation that was supplied as input.
**********************************************************************/
void render_segmentation(ScrollView *window,
TBLOB *chunks,
SEARCH_STATE segmentation) {
TBLOB *blob;
C_COL color = Black;
int char_num = -1;
int chunks_left = 0;
TBOX bbox;
if (chunks) bbox = chunks->bounding_box();
for (blob = chunks; blob != NULL; blob = blob->next) {
bbox += blob->bounding_box();
if (chunks_left-- == 0) {
color = color_list[++char_num % NUM_COLORS];
if (char_num < segmentation[0])
chunks_left = segmentation[char_num + 1];
else
chunks_left = MAX_INT32;
}
render_outline(window, blob->outlines, color);
}
window->ZoomToRectangle(bbox.left(), bbox.top(),
bbox.right(), bbox.bottom());
}
// This function takes tif/box pair of files and runs recognition on the image,
// while making sure that the word bounds that tesseract identified roughly
// match to those specified by the input box file. For each word (ngram in a
// single bounding box from the input box file) it outputs the ocred result,
// the correct label, rating and certainty.
void Tesseract::recog_training_segmented(const STRING &fname,
PAGE_RES *page_res,
volatile ETEXT_DESC *monitor,
FILE *output_file) {
STRING box_fname = fname;
const char *lastdot = strrchr(box_fname.string(), '.');
if (lastdot != NULL) box_fname[lastdot - box_fname.string()] = '\0';
box_fname += ".box";
// read_next_box() will close box_file
FILE *box_file = open_file(box_fname.string(), "r");
PAGE_RES_IT page_res_it;
page_res_it.page_res = page_res;
page_res_it.restart_page();
char label[kBoxReadBufSize];
// Process all the words on this page.
TBOX tbox; // tesseract-identified box
TBOX bbox; // box from the box file
bool keep_going;
int line_number = 0;
do {
keep_going = read_t(&page_res_it, &tbox);
keep_going &= read_b(applybox_page, &line_number, box_file, label, &bbox);
// Align bottom left points of the TBOXes.
while (keep_going &&
!NearlyEqual<int>(tbox.bottom(), bbox.bottom(), kMaxBoxEdgeDiff)) {
keep_going = (bbox.bottom() < tbox.bottom()) ?
read_t(&page_res_it, &tbox) :
read_b(applybox_page, &line_number, box_file, label, &bbox);
}
while (keep_going &&
!NearlyEqual<int>(tbox.left(), bbox.left(), kMaxBoxEdgeDiff)) {
keep_going = (bbox.left() > tbox.left()) ? read_t(&page_res_it, &tbox) :
read_b(applybox_page, &line_number, box_file, label, &bbox);
}
// OCR the word if top right points of the TBOXes are similar.
if (keep_going &&
NearlyEqual<int>(tbox.right(), bbox.right(), kMaxBoxEdgeDiff) &&
NearlyEqual<int>(tbox.top(), bbox.top(), kMaxBoxEdgeDiff)) {
ambigs_classify_and_output(page_res_it.prev_word(),
page_res_it.prev_row(),
page_res_it.prev_block(),
label, output_file);
}
} while (keep_going);
}
// Extract the OCR results, costs (penalty points for uncertainty),
// and the bounding boxes of the characters.
static void extract_result(ELIST_ITERATOR *out,
PAGE_RES* page_res) {
PAGE_RES_IT page_res_it(page_res);
int word_count = 0;
while (page_res_it.word() != NULL) {
WERD_RES *word = page_res_it.word();
const char *str = word->best_choice->string().string();
const char *len = word->best_choice->lengths().string();
if (word_count)
add_space(out);
TBOX bln_rect;
PBLOB_LIST *blobs = word->outword->blob_list();
PBLOB_IT it(blobs);
int n = strlen(len);
TBOX** boxes_to_fix = new TBOX*[n];
for (int i = 0; i < n; i++) {
PBLOB *blob = it.data();
TBOX current = pblob_get_bbox(blob);
bln_rect.bounding_union(current);
TESS_CHAR *tc = new TESS_CHAR(rating_to_cost(word->best_choice->rating()),
str, *len);
tc->box = current;
boxes_to_fix[i] = &tc->box;
out->add_after_then_move(tc);
it.forward();
str += *len;
len++;
}
// Find the word bbox before normalization.
// Here we can't use the C_BLOB bboxes directly,
// since connected letters are not yet cut.
TBOX real_rect = c_blob_list_get_bbox(word->word->cblob_list());
// Denormalize boxes by transforming the bbox of the whole bln word
// into the denorm bbox (`real_rect') of the whole word.
double x_stretch = double(real_rect.width()) / bln_rect.width();
double y_stretch = double(real_rect.height()) / bln_rect.height();
for (int j = 0; j < n; j++) {
TBOX *box = boxes_to_fix[j];
int x0 = int(real_rect.left() +
x_stretch * (box->left() - bln_rect.left()) + 0.5);
int x1 = int(real_rect.left() +
x_stretch * (box->right() - bln_rect.left()) + 0.5);
int y0 = int(real_rect.bottom() +
y_stretch * (box->bottom() - bln_rect.bottom()) + 0.5);
int y1 = int(real_rect.bottom() +
y_stretch * (box->top() - bln_rect.bottom()) + 0.5);
*box = TBOX(ICOORD(x0, y0), ICOORD(x1, y1));
}
delete [] boxes_to_fix;
page_res_it.forward();
word_count++;
}
}
// Creates a box file string from a unichar string, TBOX and page number.
void MakeBoxFileStr(const char* unichar_str, const TBOX& box, int page_num,
STRING* box_str) {
*box_str = unichar_str;
box_str->add_str_int(" ", box.left());
box_str->add_str_int(" ", box.bottom());
box_str->add_str_int(" ", box.right());
box_str->add_str_int(" ", box.top());
box_str->add_str_int(" ", page_num);
}
// Find a set of blobs that are aligned in the given vertical
// direction with the given blob. Returns a list of aligned
// blobs and the number in the list.
// For other parameters see FindAlignedBlob below.
int AlignedBlob::AlignTabs(const AlignedBlobParams& params,
bool top_to_bottom, BLOBNBOX* bbox,
BLOBNBOX_CLIST* good_points, int* end_y) {
int ptcount = 0;
BLOBNBOX_C_IT it(good_points);
TBOX box = bbox->bounding_box();
bool debug = WithinTestRegion(2, box.left(), box.bottom());
if (debug) {
tprintf("Starting alignment run at blob:");
box.print();
}
int x_start = params.right_tab ? box.right() : box.left();
while (bbox != nullptr) {
// Add the blob to the list if the appropriate side is a tab candidate,
// or if we are working on a ragged tab.
TabType type = params.right_tab ? bbox->right_tab_type()
: bbox->left_tab_type();
if (((type != TT_NONE && type != TT_MAYBE_RAGGED) || params.ragged) &&
(it.empty() || it.data() != bbox)) {
if (top_to_bottom)
it.add_before_then_move(bbox);
else
it.add_after_then_move(bbox);
++ptcount;
}
// Find the next blob that is aligned with the current one.
// FindAlignedBlob guarantees that forward progress will be made in the
// top_to_bottom direction, and therefore eventually it will return nullptr,
// making this while (bbox != nullptr) loop safe.
bbox = FindAlignedBlob(params, top_to_bottom, bbox, x_start, end_y);
if (bbox != nullptr) {
box = bbox->bounding_box();
if (!params.ragged)
x_start = params.right_tab ? box.right() : box.left();
}
}
if (debug) {
tprintf("Alignment run ended with %d pts at blob:", ptcount);
box.print();
}
return ptcount;
}
// Initialize from box coordinates.
POLY_BLOCK::POLY_BLOCK(const TBOX& box, PolyBlockType t) {
vertices.clear();
ICOORDELT_IT v = &vertices;
v.move_to_first();
v.add_to_end(new ICOORDELT(box.left(), box.top()));
v.add_to_end(new ICOORDELT(box.left(), box.bottom()));
v.add_to_end(new ICOORDELT(box.right(), box.bottom()));
v.add_to_end(new ICOORDELT(box.right(), box.top()));
compute_bb();
type = t;
}
// Normalizes the blob for classification only if needed.
// (Normally this means a non-zero classify rotation.)
// If no Normalization is needed, then NULL is returned, and the input blob
// can be used directly. Otherwise a new TBLOB is returned which must be
// deleted after use.
TBLOB* TBLOB::ClassifyNormalizeIfNeeded() const {
TBLOB* rotated_blob = NULL;
// If necessary, copy the blob and rotate it. The rotation is always
// +/- 90 degrees, as 180 was already taken care of.
if (denorm_.block() != NULL &&
denorm_.block()->classify_rotation().y() != 0.0) {
TBOX box = bounding_box();
int x_middle = (box.left() + box.right()) / 2;
int y_middle = (box.top() + box.bottom()) / 2;
rotated_blob = new TBLOB(*this);
const FCOORD& rotation = denorm_.block()->classify_rotation();
// Move the rotated blob back to the same y-position so that we
// can still distinguish similar glyphs with differeny y-position.
float target_y = kBlnBaselineOffset +
(rotation.y() > 0 ? x_middle - box.left() : box.right() - x_middle);
rotated_blob->Normalize(NULL, &rotation, &denorm_, x_middle, y_middle,
1.0f, 1.0f, 0.0f, target_y,
denorm_.inverse(), denorm_.pix());
}
return rotated_blob;
}
// Returns true if any cell value in the given rectangle is zero.
bool IntGrid::AnyZeroInRect(const TBOX& rect) const {
int min_x, min_y, max_x, max_y;
GridCoords(rect.left(), rect.bottom(), &min_x, &min_y);
GridCoords(rect.right(), rect.top(), &max_x, &max_y);
for (int y = min_y; y <= max_y; ++y) {
for (int x = min_x; x <= max_x; ++x) {
if (GridCellValue(x, y) == 0)
return true;
}
}
return false;
}
/**
* @name start_seam_list
*
* Initialize a list of seams that match the original number of blobs
* present in the starting segmentation. Each of the seams created
* by this routine have location information only.
*/
void start_seam_list(TWERD *word, GenericVector<SEAM*>* seam_array) {
seam_array->truncate(0);
TPOINT location;
for (int b = 1; b < word->NumBlobs(); ++b) {
TBOX bbox = word->blobs[b - 1]->bounding_box();
TBOX nbox = word->blobs[b]->bounding_box();
location.x = (bbox.right() + nbox.left()) / 2;
location.y = (bbox.bottom() + bbox.top() + nbox.bottom() + nbox.top()) / 4;
seam_array->push_back(new SEAM(0.0f, location, NULL, NULL, NULL));
}
}
// Setup for a baseline normalization. If there are segs, then they
// are used, otherwise, if there is a row, that is used, otherwise the
// bottom of the word_box is used for the baseline.
void DENORM::SetupBLNormalize(const BLOCK* block, const ROW* row,
float x_height, const TBOX& word_box,
int num_segs, const DENORM_SEG* segs) {
float scale = kBlnXHeight / x_height;
float x_origin = (word_box.left() + word_box.right()) / 2.0f;
float y_origin = 0.0f;
if (num_segs == 0 && row == NULL) {
y_origin = word_box.bottom();
}
SetupNormalization(block, row, NULL, NULL, segs, num_segs,
x_origin, y_origin, scale, scale,
0.0f, static_cast<float>(kBlnBaselineOffset));
}
void show_point(PAGE_RES* page_res, float x, float y) {
FCOORD pt(x, y);
PAGE_RES_IT pr_it(page_res);
const int kBufsize = 512;
char msg[kBufsize];
char *msg_ptr = msg;
msg_ptr += sprintf(msg_ptr, "Pt:(%0.3f, %0.3f) ", x, y);
for (WERD_RES* word = pr_it.word(); word != NULL; word = pr_it.forward()) {
if (pr_it.row() != pr_it.prev_row() &&
pr_it.row()->row->bounding_box().contains(pt)) {
msg_ptr += sprintf(msg_ptr, "BL(x)=%0.3f ",
pr_it.row()->row->base_line(x));
}
if (word->word->bounding_box().contains(pt)) {
TBOX box = word->word->bounding_box();
msg_ptr += sprintf(msg_ptr, "Wd(%d, %d)/(%d, %d) ",
box.left(), box.bottom(),
box.right(), box.top());
C_BLOB_IT cblob_it(word->word->cblob_list());
for (cblob_it.mark_cycle_pt();
!cblob_it.cycled_list();
cblob_it.forward()) {
C_BLOB* cblob = cblob_it.data();
box = cblob->bounding_box();
if (box.contains(pt)) {
msg_ptr += sprintf(msg_ptr,
"CBlb(%d, %d)/(%d, %d) ",
box.left(), box.bottom(),
box.right(), box.top());
}
}
}
}
image_win->AddMessage(msg);
}
/**
* Returns the bounding rectangle of the current object at the given level in
* the coordinates of the working image that is pix_binary().
* See comment on coordinate system above.
* Returns false if there is no such object at the current position.
*/
bool PageIterator::BoundingBoxInternal(PageIteratorLevel level,
int* left, int* top,
int* right, int* bottom) const {
if (Empty(level))
return false;
TBOX box;
PARA *para = NULL;
switch (level) {
case RIL_BLOCK:
box = it_->block()->block->bounding_box();
break;
case RIL_PARA:
para = it_->row()->row->para();
// explicit fall-through.
case RIL_TEXTLINE:
box = it_->row()->row->bounding_box();
break;
case RIL_WORD:
box = it_->word()->word->bounding_box();
break;
case RIL_SYMBOL:
if (cblob_it_ == NULL)
box = it_->word()->box_word->BlobBox(blob_index_);
else
box = cblob_it_->data()->bounding_box();
}
if (level == RIL_PARA) {
PageIterator other = *this;
other.Begin();
do {
if (other.it_->block() &&
other.it_->block()->block == it_->block()->block &&
other.it_->row() && other.it_->row()->row &&
other.it_->row()->row->para() == para) {
box = box.bounding_union(other.it_->row()->row->bounding_box());
}
} while (other.Next(RIL_TEXTLINE));
}
if (level != RIL_SYMBOL || cblob_it_ != NULL)
box.rotate(it_->block()->block->re_rotation());
// Now we have a box in tesseract coordinates relative to the image rectangle,
// we have to convert the coords to a top-down system.
const int pix_height = pixGetHeight(tesseract_->pix_binary());
const int pix_width = pixGetWidth(tesseract_->pix_binary());
*left = ClipToRange(static_cast<int>(box.left()), 0, pix_width);
*top = ClipToRange(pix_height - box.top(), 0, pix_height);
*right = ClipToRange(static_cast<int>(box.right()), *left, pix_width);
*bottom = ClipToRange(pix_height - box.bottom(), *top, pix_height);
return true;
}
// Helper function to add 1 to a rectangle in source image coords to the
// internal projection pix_.
void TextlineProjection::IncrementRectangle8Bit(const TBOX& box) {
int scaled_left = ImageXToProjectionX(box.left());
int scaled_top = ImageYToProjectionY(box.top());
int scaled_right = ImageXToProjectionX(box.right());
int scaled_bottom = ImageYToProjectionY(box.bottom());
int wpl = pixGetWpl(pix_);
uint32_t* data = pixGetData(pix_) + scaled_top * wpl;
for (int y = scaled_top; y <= scaled_bottom; ++y) {
for (int x = scaled_left; x <= scaled_right; ++x) {
int pixel = GET_DATA_BYTE(data, x);
if (pixel < 255)
SET_DATA_BYTE(data, x, pixel + 1);
}
data += wpl;
}
}
// Returns true if more than half the area of the rect is covered by grid
// cells that are over the threshold.
bool IntGrid::RectMostlyOverThreshold(const TBOX& rect, int threshold) const {
int min_x, min_y, max_x, max_y;
GridCoords(rect.left(), rect.bottom(), &min_x, &min_y);
GridCoords(rect.right(), rect.top(), &max_x, &max_y);
int total_area = 0;
for (int y = min_y; y <= max_y; ++y) {
for (int x = min_x; x <= max_x; ++x) {
int value = GridCellValue(x, y);
if (value > threshold) {
TBOX cell_box(x * gridsize_, y * gridsize_,
(x + 1) * gridsize_, (y + 1) * gridsize_);
cell_box &= rect; // This is in-place box intersection.
total_area += cell_box.area();
}
}
}
return total_area * 2 > rect.area();
}
// Display the tab codes of the BLOBNBOXes in this grid.
ScrollView* AlignedBlob::DisplayTabs(const char* window_name,
ScrollView* tab_win) {
#ifndef GRAPHICS_DISABLED
if (tab_win == NULL)
tab_win = MakeWindow(0, 50, window_name);
// For every tab in the grid, display it.
GridSearch<BLOBNBOX, BLOBNBOX_CLIST, BLOBNBOX_C_IT> gsearch(this);
gsearch.StartFullSearch();
BLOBNBOX* bbox;
while ((bbox = gsearch.NextFullSearch()) != NULL) {
TBOX box = bbox->bounding_box();
int left_x = box.left();
int right_x = box.right();
int top_y = box.top();
int bottom_y = box.bottom();
TabType tabtype = bbox->left_tab_type();
if (tabtype != TT_NONE) {
if (tabtype == TT_UNCONFIRMED)
tab_win->Pen(ScrollView::BLUE);
else if (tabtype == TT_CONFIRMED)
tab_win->Pen(ScrollView::GREEN);
else if (tabtype == TT_FAKE)
tab_win->Pen(ScrollView::YELLOW);
else
tab_win->Pen(ScrollView::GREY);
tab_win->Line(left_x, top_y, left_x, bottom_y);
}
tabtype = bbox->right_tab_type();
if (tabtype != TT_NONE) {
if (tabtype == TT_UNCONFIRMED)
tab_win->Pen(ScrollView::MAGENTA);
else if (tabtype == TT_CONFIRMED)
tab_win->Pen(ScrollView::RED);
else if (tabtype == TT_FAKE)
tab_win->Pen(ScrollView::ORANGE);
else
tab_win->Pen(ScrollView::GREY);
tab_win->Line(right_x, top_y, right_x, bottom_y);
}
}
tab_win->Update();
#endif
return tab_win;
}
// Displays the segmentation state of *this (if not the same as the last
// one displayed) and waits for a click in the window.
void WERD_CHOICE::DisplaySegmentation(TWERD* word) {
#ifndef GRAPHICS_DISABLED
// Number of different colors to draw with.
const int kNumColors = 6;
static ScrollView *segm_window = NULL;
// Check the state against the static prev_drawn_state.
static GenericVector<int> prev_drawn_state;
bool already_done = prev_drawn_state.size() == length_;
if (!already_done) prev_drawn_state.init_to_size(length_, 0);
for (int i = 0; i < length_; ++i) {
if (prev_drawn_state[i] != state_[i]) {
already_done = false;
}
prev_drawn_state[i] = state_[i];
}
if (already_done || word->blobs.empty()) return;
// Create the window if needed.
if (segm_window == NULL) {
segm_window = new ScrollView("Segmentation", 5, 10, 500, 256,
2000.0, 256.0, true);
} else {
segm_window->Clear();
}
TBOX bbox;
int blob_index = 0;
for (int c = 0; c < length_; ++c) {
ScrollView::Color color =
static_cast<ScrollView::Color>(c % kNumColors + 3);
for (int i = 0; i < state_[c]; ++i, ++blob_index) {
TBLOB* blob = word->blobs[blob_index];
bbox += blob->bounding_box();
blob->plot(segm_window, color, color);
}
}
segm_window->ZoomToRectangle(bbox.left(), bbox.top(),
bbox.right(), bbox.bottom());
segm_window->Update();
window_wait(segm_window);
#endif
}
/**
* Returns the baseline of the current object at the given level.
* The baseline is the line that passes through (x1, y1) and (x2, y2).
* WARNING: with vertical text, baselines may be vertical!
*/
bool PageIterator::Baseline(PageIteratorLevel level,
int* x1, int* y1, int* x2, int* y2) const {
if (it_->word() == NULL) return false; // Already at the end!
ROW* row = it_->row()->row;
WERD* word = it_->word()->word;
TBOX box = (level == RIL_WORD || level == RIL_SYMBOL)
? word->bounding_box()
: row->bounding_box();
int left = box.left();
ICOORD startpt(left, static_cast<inT16>(row->base_line(left) + 0.5));
int right = box.right();
ICOORD endpt(right, static_cast<inT16>(row->base_line(right) + 0.5));
// Rotate to image coordinates and convert to global image coords.
startpt.rotate(it_->block()->block->re_rotation());
endpt.rotate(it_->block()->block->re_rotation());
*x1 = startpt.x() / scale_ + rect_left_;
*y1 = (rect_height_ - startpt.y()) / scale_ + rect_top_;
*x2 = endpt.x() / scale_ + rect_left_;
*y2 = (rect_height_ - endpt.y()) / scale_ + rect_top_;
return true;
}
// 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();
}
}
/**********************************************************************
* 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);
}
}