// 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;
}
// Evaluates the textlineiness of a ColPartition. Uses EvaluateBox below,
// but uses the median top/bottom for horizontal and median left/right for
// vertical instead of the bounding box edges.
// Evaluates for both horizontal and vertical and returns the best result,
// with a positive value for horizontal and a negative value for vertical.
int TextlineProjection::EvaluateColPartition(const ColPartition& part,
const DENORM* denorm,
bool debug) const {
if (part.IsSingleton())
return EvaluateBox(part.bounding_box(), denorm, debug);
// Test vertical orientation.
TBOX box = part.bounding_box();
// Use the partition median for left/right.
box.set_left(part.median_left());
box.set_right(part.median_right());
int vresult = EvaluateBox(box, denorm, debug);
// Test horizontal orientation.
box = part.bounding_box();
// Use the partition median for top/bottom.
box.set_top(part.median_top());
box.set_bottom(part.median_bottom());
int hresult = EvaluateBox(box, denorm, debug);
if (debug) {
tprintf("Partition hresult=%d, vresult=%d from:", hresult, vresult);
part.bounding_box().print();
part.Print();
}
return hresult >= -vresult ? hresult : vresult;
}
// Swaps the outlines of *this and next if needed to keep the centers in
// increasing x.
void TBLOB::CorrectBlobOrder(TBLOB* next) {
TBOX box = bounding_box();
TBOX next_box = next->bounding_box();
if (box.x_middle() > next_box.x_middle()) {
Swap(&outlines, &next->outlines);
}
}
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());
}
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());
}
}
}
}
}
// 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;
}
/**********************************************************************
* 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());
}
// Returns an Imagedata containing the image of the given box,
// and ground truth boxes/truth text if available in the input.
// The image is not normalized in any way.
ImageData* Tesseract::GetLineData(const TBOX& line_box,
const GenericVector<TBOX>& boxes,
const GenericVector<STRING>& texts,
int start_box, int end_box,
const BLOCK& block) {
TBOX revised_box;
ImageData* image_data = GetRectImage(line_box, block, kImagePadding,
&revised_box);
if (image_data == NULL) return NULL;
image_data->set_page_number(applybox_page);
// Copy the boxes and shift them so they are relative to the image.
FCOORD block_rotation(block.re_rotation().x(), -block.re_rotation().y());
ICOORD shift = -revised_box.botleft();
GenericVector<TBOX> line_boxes;
GenericVector<STRING> line_texts;
for (int b = start_box; b < end_box; ++b) {
TBOX box = boxes[b];
box.rotate(block_rotation);
box.move(shift);
line_boxes.push_back(box);
line_texts.push_back(texts[b]);
}
GenericVector<int> page_numbers;
page_numbers.init_to_size(line_boxes.size(), applybox_page);
image_data->AddBoxes(line_boxes, line_texts, page_numbers);
return image_data;
}
void fixspace_dbg(WERD_RES *word) {
TBOX box = word->word->bounding_box ();
BOOL8 show_map_detail = FALSE;
inT16 i;
box.print ();
#ifndef SECURE_NAMES
tprintf (" \"%s\" ", word->best_choice->string ().string ());
tprintf ("Blob count: %d (word); %d/%d (outword)\n",
word->word->gblob_list ()->length (),
word->outword->gblob_list ()->length (),
word->outword->rej_blob_list ()->length ());
word->reject_map.print (debug_fp);
tprintf ("\n");
if (show_map_detail) {
tprintf ("\"%s\"\n", word->best_choice->string ().string ());
for (i = 0; word->best_choice->string ()[i] != '\0'; i++) {
tprintf ("**** \"%c\" ****\n", word->best_choice->string ()[i]);
word->reject_map[i].full_print (debug_fp);
}
}
tprintf ("Tess Accepted: %s\n", word->tess_accepted ? "TRUE" : "FALSE");
tprintf ("Done flag: %s\n\n", word->done ? "TRUE" : "FALSE");
#endif
}
// 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;
}
/** Handles a click event in a display window. */
void StrokeWidth::HandleClick(int x, int y) {
BBGrid<BLOBNBOX, BLOBNBOX_CLIST, BLOBNBOX_C_IT>::HandleClick(x, y);
// Run a radial search for blobs that overlap.
BlobGridSearch radsearch(this);
radsearch.StartRadSearch(x, y, 1);
BLOBNBOX* neighbour;
FCOORD click(static_cast<float>(x), static_cast<float>(y));
while ((neighbour = radsearch.NextRadSearch()) != NULL) {
TBOX nbox = neighbour->bounding_box();
if (nbox.contains(click) && neighbour->cblob() != NULL) {
PrintBoxWidths(neighbour);
if (neighbour->neighbour(BND_LEFT) != NULL)
PrintBoxWidths(neighbour->neighbour(BND_LEFT));
if (neighbour->neighbour(BND_RIGHT) != NULL)
PrintBoxWidths(neighbour->neighbour(BND_RIGHT));
if (neighbour->neighbour(BND_ABOVE) != NULL)
PrintBoxWidths(neighbour->neighbour(BND_ABOVE));
if (neighbour->neighbour(BND_BELOW) != NULL)
PrintBoxWidths(neighbour->neighbour(BND_BELOW));
int gaps[BND_COUNT];
neighbour->NeighbourGaps(gaps);
tprintf("Left gap=%d, right=%d, above=%d, below=%d, horz=%d, vert=%d\n"
"Good= %d %d %d %d\n",
gaps[BND_LEFT], gaps[BND_RIGHT],
gaps[BND_ABOVE], gaps[BND_BELOW],
neighbour->horz_possible(),
neighbour->vert_possible(),
neighbour->good_stroke_neighbour(BND_LEFT),
neighbour->good_stroke_neighbour(BND_RIGHT),
neighbour->good_stroke_neighbour(BND_ABOVE),
neighbour->good_stroke_neighbour(BND_BELOW));
break;
}
}
}
// 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;
}
// Computes the DENORMS for bl(baseline) and cn(character) normalization
// during feature extraction. The input denorm describes the current state
// of the blob, which is usually a baseline-normalized word.
// The Transforms setup are as follows:
// Baseline Normalized (bl) Output:
// We center the grapheme by aligning the x-coordinate of its centroid with
// x=128 and leaving the already-baseline-normalized y as-is.
//
// Character Normalized (cn) Output:
// We align the grapheme's centroid at the origin and scale it
// asymmetrically in x and y so that the 2nd moments are a standard value
// (51.2) ie the result is vaguely square.
// If classify_nonlinear_norm is true:
// A non-linear normalization is setup that attempts to evenly distribute
// edges across x and y.
//
// Some of the fields of fx_info are also setup:
// Length: Total length of outline.
// Rx: Rounded y second moment. (Reversed by convention.)
// Ry: rounded x second moment.
// Xmean: Rounded x center of mass of the blob.
// Ymean: Rounded y center of mass of the blob.
void Classify::SetupBLCNDenorms(const TBLOB& blob, bool nonlinear_norm,
DENORM* bl_denorm, DENORM* cn_denorm,
INT_FX_RESULT_STRUCT* fx_info) {
// Compute 1st and 2nd moments of the original outline.
FCOORD center, second_moments;
int length = blob.ComputeMoments(¢er, &second_moments);
if (fx_info != nullptr) {
fx_info->Length = length;
fx_info->Rx = IntCastRounded(second_moments.y());
fx_info->Ry = IntCastRounded(second_moments.x());
fx_info->Xmean = IntCastRounded(center.x());
fx_info->Ymean = IntCastRounded(center.y());
}
// Setup the denorm for Baseline normalization.
bl_denorm->SetupNormalization(nullptr, nullptr, &blob.denorm(), center.x(), 128.0f,
1.0f, 1.0f, 128.0f, 128.0f);
// Setup the denorm for character normalization.
if (nonlinear_norm) {
GenericVector<GenericVector<int> > x_coords;
GenericVector<GenericVector<int> > y_coords;
TBOX box;
blob.GetPreciseBoundingBox(&box);
box.pad(1, 1);
blob.GetEdgeCoords(box, &x_coords, &y_coords);
cn_denorm->SetupNonLinear(&blob.denorm(), box, UINT8_MAX, UINT8_MAX,
0.0f, 0.0f, x_coords, y_coords);
} else {
cn_denorm->SetupNormalization(nullptr, nullptr, &blob.denorm(),
center.x(), center.y(),
51.2f / second_moments.x(),
51.2f / second_moments.y(),
128.0f, 128.0f);
}
}
// 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++;
}
}
/// Helper to compute the dispute resolution metric.
/// Disputed blob resolution. The aim is to give the blob to the most
/// appropriate boxfile box. Most of the time it is obvious, but if
/// two boxfile boxes overlap significantly it is not. If a small boxfile
/// box takes most of the blob, and a large boxfile box does too, then
/// we want the small boxfile box to get it, but if the small box
/// is much smaller than the blob, we don't want it to get it.
/// Details of the disputed blob resolution:
/// Given a box with area A, and a blob with area B, with overlap area C,
/// then the miss metric is (A-C)(B-C)/(AB) and the box with minimum
/// miss metric gets the blob.
static double BoxMissMetric(const TBOX& box1, const TBOX& box2) {
int overlap_area = box1.intersection(box2).area();
double miss_metric = box1.area()- overlap_area;
miss_metric /= box1.area();
miss_metric *= box2.area() - overlap_area;
miss_metric /= box2.area();
return miss_metric;
}
// Build a fake outline, given just a bounding box and append to the list.
void C_OUTLINE::FakeOutline(const TBOX& box, C_OUTLINE_LIST* outlines) {
C_OUTLINE_IT ol_it(outlines);
// Make a C_OUTLINE from the bounds. This is a bit of a hack,
// as there is no outline, just a bounding box, but it works nicely.
CRACKEDGE start;
start.pos = box.topleft();
C_OUTLINE* outline = new C_OUTLINE(&start, box.topleft(), box.botright(), 0);
ol_it.add_to_end(outline);
}
// Return the partner of this TabVector if the vector qualifies as
// being a vertical text line, otherwise NULL.
TabVector* TabVector::VerticalTextlinePartner() {
if (!partners_.singleton())
return NULL;
TabVector_C_IT partner_it(&partners_);
TabVector* partner = partner_it.data();
BLOBNBOX_C_IT box_it1(&boxes_);
BLOBNBOX_C_IT box_it2(&partner->boxes_);
// Count how many boxes are also in the other list.
// At the same time, gather the mean width and median vertical gap.
if (textord_debug_tabfind > 1) {
Print("Testing for vertical text");
partner->Print(" partner");
}
int num_matched = 0;
int num_unmatched = 0;
int total_widths = 0;
int width = startpt().x() - partner->startpt().x();
if (width < 0)
width = -width;
STATS gaps(0, width * 2);
BLOBNBOX* prev_bbox = NULL;
box_it2.mark_cycle_pt();
for (box_it1.mark_cycle_pt(); !box_it1.cycled_list(); box_it1.forward()) {
BLOBNBOX* bbox = box_it1.data();
TBOX box = bbox->bounding_box();
if (prev_bbox != NULL) {
gaps.add(box.bottom() - prev_bbox->bounding_box().top(), 1);
}
while (!box_it2.cycled_list() && box_it2.data() != bbox &&
box_it2.data()->bounding_box().bottom() < box.bottom()) {
box_it2.forward();
}
if (!box_it2.cycled_list() && box_it2.data() == bbox &&
bbox->region_type() >= BRT_UNKNOWN &&
(prev_bbox == NULL || prev_bbox->region_type() >= BRT_UNKNOWN))
++num_matched;
else
++num_unmatched;
total_widths += box.width();
prev_bbox = bbox;
}
if (num_unmatched + num_matched == 0) return NULL;
double avg_width = total_widths * 1.0 / (num_unmatched + num_matched);
double max_gap = textord_tabvector_vertical_gap_fraction * avg_width;
int min_box_match = static_cast<int>((num_matched + num_unmatched) *
textord_tabvector_vertical_box_ratio);
bool is_vertical = (gaps.get_total() > 0 &&
num_matched >= min_box_match &&
gaps.median() <= max_gap);
if (textord_debug_tabfind > 1) {
tprintf("gaps=%d, matched=%d, unmatched=%d, min_match=%d "
"median gap=%.2f, width=%.2f max_gap=%.2f Vertical=%s\n",
gaps.get_total(), num_matched, num_unmatched, min_box_match,
gaps.median(), avg_width, max_gap, is_vertical?"Yes":"No");
}
return (is_vertical) ? partner : NULL;
}
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;
}
}
// 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);
}
// TODO(mezhirov) delete this function and replace with word->bounding_box()
static TBOX c_blob_list_get_bbox(C_BLOB_LIST *cblobs) {
TBOX result;
C_BLOB_IT c_it(cblobs);
for (c_it.mark_cycle_pt(); !c_it.cycled_list(); c_it.forward()) {
C_BLOB *blob = c_it.data();
//bboxes.push(tessy_rectangle(blob->bounding_box()));
result.bounding_union(blob->bounding_box());
}
return result;
}
// Helper gets the image of a rectangle, using the block.re_rotation() if
// needed to get to the image, and rotating the result back to horizontal
// layout. (CJK characters will be on their left sides) The vertical text flag
// is set in the returned ImageData if the text was originally vertical, which
// can be used to invoke a different CJK recognition engine. The revised_box
// is also returned to enable calculation of output bounding boxes.
ImageData* Tesseract::GetRectImage(const TBOX& box, const BLOCK& block,
int padding, TBOX* revised_box) const {
TBOX wbox = box;
wbox.pad(padding, padding);
*revised_box = wbox;
// Number of clockwise 90 degree rotations needed to get back to tesseract
// coords from the clipped image.
int num_rotations = 0;
if (block.re_rotation().y() > 0.0f)
num_rotations = 1;
else if (block.re_rotation().x() < 0.0f)
num_rotations = 2;
else if (block.re_rotation().y() < 0.0f)
num_rotations = 3;
// Handle two cases automatically: 1 the box came from the block, 2 the box
// came from a box file, and refers to the image, which the block may not.
if (block.bounding_box().major_overlap(*revised_box))
revised_box->rotate(block.re_rotation());
// Now revised_box always refers to the image.
// BestPix is never colormapped, but may be of any depth.
Pix* pix = BestPix();
int width = pixGetWidth(pix);
int height = pixGetHeight(pix);
TBOX image_box(0, 0, width, height);
// Clip to image bounds;
*revised_box &= image_box;
if (revised_box->null_box()) return NULL;
Box* clip_box = boxCreate(revised_box->left(), height - revised_box->top(),
revised_box->width(), revised_box->height());
Pix* box_pix = pixClipRectangle(pix, clip_box, NULL);
if (box_pix == NULL) return NULL;
boxDestroy(&clip_box);
if (num_rotations > 0) {
Pix* rot_pix = pixRotateOrth(box_pix, num_rotations);
pixDestroy(&box_pix);
box_pix = rot_pix;
}
// Convert sub-8-bit images to 8 bit.
int depth = pixGetDepth(box_pix);
if (depth < 8) {
Pix* grey;
grey = pixConvertTo8(box_pix, false);
pixDestroy(&box_pix);
box_pix = grey;
}
bool vertical_text = false;
if (num_rotations > 0) {
// Rotated the clipped revised box back to internal coordinates.
FCOORD rotation(block.re_rotation().x(), -block.re_rotation().y());
revised_box->rotate(rotation);
if (num_rotations != 2)
vertical_text = true;
}
return new ImageData(vertical_text, box_pix);
}
// brief Get a bounding box of a PBLOB.
// TODO(mezhirov) delete this function and replace with blob->bounding_box()
static TBOX pblob_get_bbox(PBLOB *blob) {
OUTLINE_LIST *outlines = blob->out_list();
OUTLINE_IT it(outlines);
TBOX result;
for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) {
OUTLINE *outline = it.data();
outline->compute_bb();
result.bounding_union(outline->bounding_box());
}
return result;
}
// 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());
}
}
}
// 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();
}
// 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;
}
// 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));
}
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);
}
}
/**********************************************************************
* 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);
}
// Returns the bounding box including the desired combination of upper and
// lower noise/diacritic elements.
TBOX WERD::restricted_bounding_box(bool upper_dots, bool lower_dots) const {
TBOX box = true_bounding_box();
int bottom = box.bottom();
int top = box.top();
// This is a read-only iteration of the rejected blobs.
C_BLOB_IT it(const_cast<C_BLOB_LIST*>(&rej_cblobs));
for (it.mark_cycle_pt(); !it.cycled_list(); it.forward()) {
TBOX dot_box = it.data()->bounding_box();
if ((upper_dots || dot_box.bottom() <= top) &&
(lower_dots || dot_box.top() >= bottom)) {
box += dot_box;
}
}
return box;
}
// 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;
}
}
}
