/// 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;
}
// 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);
}
}
}
// 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());
}
}
}
// 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();
}