// Returns the maximum strokewidth in the given binary image by doubling
// the maximum of the distance function.
static int MaxStrokeWidth(Pix* pix) {
Pix* dist_pix = pixDistanceFunction(pix, 4, 8, L_BOUNDARY_BG);
int width = pixGetWidth(dist_pix);
int height = pixGetHeight(dist_pix);
int wpl = pixGetWpl(dist_pix);
l_uint32* data = pixGetData(dist_pix);
// Find the maximum value in the distance image.
int max_dist = 0;
for (int y = 0; y < height; ++y) {
for (int x = 0; x < width; ++x) {
int pixel = GET_DATA_BYTE(data, x);
if (pixel > max_dist)
max_dist = pixel;
}
data += wpl;
}
pixDestroy(&dist_pix);
return max_dist * 2;
}
/*!
* \brief pixFindStrokeWidth()
*
* \param[in] pixs 1 bpp
* \param[in] thresh fractional count threshold relative to distance 1
* \param[in] tab8 [optional] table for counting fg pixels; can be NULL
* \param[out] *pwidth estimated width of the strokes
* \param[out] *pnahisto [optional] histo of pixel distances from bg
* \return 0 if OK, 1 on error
*
* <pre>
* Notes:
* (1) This uses two methods to estimate the stroke width:
* (a) half the fg boundary length
* (b) a value derived from the histogram of the fg distance transform
* (2) Distance is measured in 8-connected
* (3) %thresh is the minimum fraction N(dist=d)/N(dist=1) of pixels
* required to determine if the pixels at distance d are above
* the noise. It is typically about 0.15.
* </pre>
*/
l_int32
pixFindStrokeWidth(PIX *pixs,
l_float32 thresh,
l_int32 *tab8,
l_float32 *pwidth,
NUMA **pnahisto)
{
l_int32 i, n, count, length, first, last;
l_int32 *tab;
l_float32 width1, width2, ratio, extra;
l_float32 *fa;
NUMA *na1, *na2;
PIX *pix1;
PROCNAME("pixFindStrokeWidth");
if (!pwidth)
return ERROR_INT("&width not defined", procName, 1);
*pwidth = 0;
if (!pixs)
return ERROR_INT("pixs not defined", procName, 1);
tab = (tab8) ? tab8 : makePixelSumTab8();
/* ------- Method 1: via boundary length ------- */
/* The computed stroke length is a bit larger than that actual
* length, because of the addition of the 'caps' at the
* stroke ends. Therefore the computed width is a bit
* smaller than the average width. */
pixFindStrokeLength(pixs, tab8, &length);
pixCountPixels(pixs, &count, tab8);
width1 = (l_float32)count / (l_float32)length;
/* ------- Method 2: via distance transform ------- */
/* First get the histogram of distances */
pix1 = pixDistanceFunction(pixs, 8, 8, L_BOUNDARY_BG);
na1 = pixGetGrayHistogram(pix1, 1);
pixDestroy(&pix1);
numaGetNonzeroRange(na1, 0.1, &first, &last);
na2 = numaClipToInterval(na1, 0, last);
numaWriteStream(stderr, na2);
/* Find the bucket with the largest distance whose contents
* exceed the threshold. */
fa = numaGetFArray(na2, L_NOCOPY);
n = numaGetCount(na2);
for (i = n - 1; i > 0; i--) {
ratio = fa[i] / fa[1];
if (ratio > thresh) break;
}
/* Let the last skipped bucket contribute to the stop bucket.
* This is the 'extra' term below. The result may be a slight
* over-correction, so the computed width may be a bit larger
* than the average width. */
extra = (i < n - 1) ? fa[i + 1] / fa[1] : 0;
width2 = 2.0 * (i - 1.0 + ratio + extra);
fprintf(stderr, "width1 = %5.2f, width2 = %5.2f\n", width1, width2);
/* Average the two results */
*pwidth = (width1 + width2) / 2.0;
if (!tab8) LEPT_FREE(tab);
numaDestroy(&na1);
if (pnahisto)
*pnahisto = na2;
else
numaDestroy(&na2);
return 0;
}
static void
TestDistance(PIXA *pixa,
PIX *pixs,
l_int32 conn,
l_int32 depth,
l_int32 bc,
l_int32 *pcount,
L_REGPARAMS *rp)
{
PIX *pixt1, *pixt2, *pixt3, *pixt4, *pixt5;
/* Test the distance function and display */
pixInvert(pixs, pixs);
pixt1 = pixDistanceFunction(pixs, conn, depth, bc);
regTestWritePixAndCheck(pixt1, IFF_PNG, pcount, rp);
pixSaveTiled(pixt1, pixa, 1, 1, 20, 0);
pixInvert(pixs, pixs);
pixt2 = pixMaxDynamicRange(pixt1, L_LOG_SCALE);
regTestWritePixAndCheck(pixt2, IFF_JFIF_JPEG, pcount, rp);
pixSaveTiled(pixt2, pixa, 1, 0, 20, 0);
pixDestroy(&pixt1);
pixDestroy(&pixt2);
/* Test the distance function and display with contour rendering */
pixInvert(pixs, pixs);
pixt1 = pixDistanceFunction(pixs, conn, depth, bc);
regTestWritePixAndCheck(pixt1, IFF_PNG, pcount, rp);
pixSaveTiled(pixt1, pixa, 1, 1, 20, 0);
pixInvert(pixs, pixs);
pixt2 = pixRenderContours(pixt1, 2, 4, 1); /* binary output */
regTestWritePixAndCheck(pixt2, IFF_PNG, pcount, rp);
pixSaveTiled(pixt2, pixa, 1, 0, 20, 0);
pixt3 = pixRenderContours(pixt1, 2, 4, depth);
pixt4 = pixMaxDynamicRange(pixt3, L_LINEAR_SCALE);
regTestWritePixAndCheck(pixt4, IFF_JFIF_JPEG, pcount, rp);
pixSaveTiled(pixt4, pixa, 1, 0, 20, 0);
pixt5 = pixMaxDynamicRange(pixt3, L_LOG_SCALE);
regTestWritePixAndCheck(pixt5, IFF_JFIF_JPEG, pcount, rp);
pixSaveTiled(pixt5, pixa, 1, 0, 20, 0);
pixDestroy(&pixt1);
pixDestroy(&pixt2);
pixDestroy(&pixt3);
pixDestroy(&pixt4);
pixDestroy(&pixt5);
/* Label all pixels in each c.c. with a color equal to the
* max distance of any pixel within that c.c. from the bg.
* Note that we've normalized so the dynamic range extends
* to 255. For the image here, each unit of distance is
* represented by about 21 grayscale units. The largest
* distance is 12. */
if (depth == 8) {
pixt1 = pixDistanceFunction(pixs, conn, depth, bc);
pixt4 = pixMaxDynamicRange(pixt1, L_LOG_SCALE);
regTestWritePixAndCheck(pixt4, IFF_JFIF_JPEG, pcount, rp);
pixSaveTiled(pixt4, pixa, 1, 1, 20, 0);
pixt2 = pixCreateTemplate(pixt1);
pixSetMasked(pixt2, pixs, 255);
regTestWritePixAndCheck(pixt2, IFF_JFIF_JPEG, pcount, rp);
pixSaveTiled(pixt2, pixa, 1, 0, 20, 0);
pixSeedfillGray(pixt1, pixt2, 4);
pixt3 = pixMaxDynamicRange(pixt1, L_LINEAR_SCALE);
regTestWritePixAndCheck(pixt3, IFF_JFIF_JPEG, pcount, rp);
pixSaveTiled(pixt3, pixa, 1, 0, 20, 0);
pixDestroy(&pixt1);
pixDestroy(&pixt2);
pixDestroy(&pixt3);
pixDestroy(&pixt4);
}
return;
}
/**********************************************************************
* SetBlobStrokeWidth
*
* Set the horizontal and vertical stroke widths in the blob.
**********************************************************************/
void SetBlobStrokeWidth(Pix* pix, BLOBNBOX* blob) {
// Cut the blob rectangle into a Pix.
int pix_height = pixGetHeight(pix);
const TBOX& box = blob->bounding_box();
int width = box.width();
int height = box.height();
Box* blob_pix_box = boxCreate(box.left(), pix_height - box.top(),
width, height);
Pix* pix_blob = pixClipRectangle(pix, blob_pix_box, nullptr);
boxDestroy(&blob_pix_box);
Pix* dist_pix = pixDistanceFunction(pix_blob, 4, 8, L_BOUNDARY_BG);
pixDestroy(&pix_blob);
// Compute the stroke widths.
uint32_t* data = pixGetData(dist_pix);
int wpl = pixGetWpl(dist_pix);
// Horizontal width of stroke.
STATS h_stats(0, width + 1);
for (int y = 0; y < height; ++y) {
uint32_t* pixels = data + y*wpl;
int prev_pixel = 0;
int pixel = GET_DATA_BYTE(pixels, 0);
for (int x = 1; x < width; ++x) {
int next_pixel = GET_DATA_BYTE(pixels, x);
// We are looking for a pixel that is equal to its vertical neighbours,
// yet greater than its left neighbour.
if (prev_pixel < pixel &&
(y == 0 || pixel == GET_DATA_BYTE(pixels - wpl, x - 1)) &&
(y == height - 1 || pixel == GET_DATA_BYTE(pixels + wpl, x - 1))) {
if (pixel > next_pixel) {
// Single local max, so an odd width.
h_stats.add(pixel * 2 - 1, 1);
} else if (pixel == next_pixel && x + 1 < width &&
pixel > GET_DATA_BYTE(pixels, x + 1)) {
// Double local max, so an even width.
h_stats.add(pixel * 2, 1);
}
}
prev_pixel = pixel;
pixel = next_pixel;
}
}
// Vertical width of stroke.
STATS v_stats(0, height + 1);
for (int x = 0; x < width; ++x) {
int prev_pixel = 0;
int pixel = GET_DATA_BYTE(data, x);
for (int y = 1; y < height; ++y) {
uint32_t* pixels = data + y*wpl;
int next_pixel = GET_DATA_BYTE(pixels, x);
// We are looking for a pixel that is equal to its horizontal neighbours,
// yet greater than its upper neighbour.
if (prev_pixel < pixel &&
(x == 0 || pixel == GET_DATA_BYTE(pixels - wpl, x - 1)) &&
(x == width - 1 || pixel == GET_DATA_BYTE(pixels - wpl, x + 1))) {
if (pixel > next_pixel) {
// Single local max, so an odd width.
v_stats.add(pixel * 2 - 1, 1);
} else if (pixel == next_pixel && y + 1 < height &&
pixel > GET_DATA_BYTE(pixels + wpl, x)) {
// Double local max, so an even width.
v_stats.add(pixel * 2, 1);
}
}
prev_pixel = pixel;
pixel = next_pixel;
}
}
pixDestroy(&dist_pix);
// Store the horizontal and vertical width in the blob, keeping both
// widths if there is enough information, otherwse only the one with
// the most samples.
// If there are insufficient samples, store zero, rather than using
// 2*area/perimeter, as the numbers that gives do not match the numbers
// from the distance method.
if (h_stats.get_total() >= (width + height) / 4) {
blob->set_horz_stroke_width(h_stats.ile(0.5f));
if (v_stats.get_total() >= (width + height) / 4)
blob->set_vert_stroke_width(v_stats.ile(0.5f));
else
blob->set_vert_stroke_width(0.0f);
} else {
if (v_stats.get_total() >= (width + height) / 4 ||
v_stats.get_total() > h_stats.get_total()) {
blob->set_horz_stroke_width(0.0f);
blob->set_vert_stroke_width(v_stats.ile(0.5f));
} else {
blob->set_horz_stroke_width(h_stats.get_total() > 2 ? h_stats.ile(0.5f)
: 0.0f);
blob->set_vert_stroke_width(0.0f);
}
}
}
/**********************************************************************
* SetBlobStrokeWidth
*
* Set the horizontal and vertical stroke widths in the blob.
**********************************************************************/
void SetBlobStrokeWidth(bool debug, BLOBNBOX* blob) {
#ifdef HAVE_LIBLEPT
// Cut the blob rectangle into a Pix.
// TODO(rays) make the page_image a Pix so this is more direct.
const TBOX& box = blob->bounding_box();
IMAGE blob_im;
int width = box.width();
int height = box.height();
blob_im.create(width, height, 1);
copy_sub_image(&page_image, box.left(), box.bottom(), width, height,
&blob_im, 0, 0, false);
Pix* pix = blob_im.ToPix();
Pix* dist_pix = pixDistanceFunction(pix, 4, 8, L_BOUNDARY_BG);
if (debug) {
pixWrite("cutpix.png", pix, IFF_PNG);
pixWrite("distpix.png", dist_pix, IFF_PNG);
}
pixDestroy(&pix);
// Compute the stroke widths.
uinT32* data = pixGetData(dist_pix);
int wpl = pixGetWpl(dist_pix);
// Horizontal width of stroke.
STATS h_stats(0, width + 1);
for (int y = 0; y < height; ++y) {
uinT32* pixels = data + y*wpl;
int prev_pixel = 0;
int pixel = GET_DATA_BYTE(pixels, 0);
for (int x = 1; x < width; ++x) {
int next_pixel = GET_DATA_BYTE(pixels, x);
// We are looking for a pixel that is equal to its vertical neighbours,
// yet greater than its left neighbour.
if (prev_pixel < pixel &&
(y == 0 || pixel == GET_DATA_BYTE(pixels - wpl, x - 1)) &&
(y == height - 1 || pixel == GET_DATA_BYTE(pixels + wpl, x - 1))) {
if (pixel > next_pixel) {
// Single local max, so an odd width.
h_stats.add(pixel * 2 - 1, 1);
} else if (pixel == next_pixel && x + 1 < width &&
pixel > GET_DATA_BYTE(pixels, x + 1)) {
// Double local max, so an even width.
h_stats.add(pixel * 2, 1);
}
}
prev_pixel = pixel;
pixel = next_pixel;
}
}
if (debug) {
h_stats.print(stderr, true);
}
// Vertical width of stroke.
STATS v_stats(0, height + 1);
for (int x = 0; x < width; ++x) {
int prev_pixel = 0;
int pixel = GET_DATA_BYTE(data, x);
for (int y = 1; y < height; ++y) {
uinT32* pixels = data + y*wpl;
int next_pixel = GET_DATA_BYTE(pixels, x);
// We are looking for a pixel that is equal to its horizontal neighbours,
// yet greater than its upper neighbour.
if (prev_pixel < pixel &&
(x == 0 || pixel == GET_DATA_BYTE(pixels - wpl, x - 1)) &&
(x == width - 1 || pixel == GET_DATA_BYTE(pixels - wpl, x + 1))) {
if (pixel > next_pixel) {
// Single local max, so an odd width.
v_stats.add(pixel * 2 - 1, 1);
} else if (pixel == next_pixel && y + 1 < height &&
pixel > GET_DATA_BYTE(pixels + wpl, x)) {
// Double local max, so an even width.
v_stats.add(pixel * 2, 1);
}
}
prev_pixel = pixel;
pixel = next_pixel;
}
}
if (debug) {
v_stats.print(stderr, true);
}
pixDestroy(&dist_pix);
// Store the horizontal and vertical width in the blob, keeping both
// widths if there is enough information, otherwse only the one with
// the most samples.
// If there are insufficent samples, store zero, rather than using
// 2*area/perimeter, as the numbers that gives do not match the numbers
// from the distance method.
if (debug) {
tprintf("box=%d,%d->%d,%d, hcount=%d, vcount=%d, target=%d\n",
box.left(), box.bottom(), box.right(), box.top(),
h_stats.get_total(), v_stats.get_total(), (width+height) /4);
tprintf("hstats median=%f, lq=%f, uq=%f, sd=%f\n",
h_stats.median(), h_stats.ile(0.25f), h_stats.ile(0.75f),
h_stats.sd());
tprintf("vstats median=%f, lq=%f, uq=%f, sd=%f\n",
v_stats.median(), v_stats.ile(0.25f), v_stats.ile(0.75f),
v_stats.sd());
}
if (h_stats.get_total() >= (width + height) / 4) {
blob->set_horz_stroke_width(h_stats.ile(0.5f));
if (v_stats.get_total() >= (width + height) / 4)
blob->set_vert_stroke_width(v_stats.ile(0.5f));
else
blob->set_vert_stroke_width(0.0f);
} else {
if (v_stats.get_total() >= (width + height) / 4 ||
v_stats.get_total() > h_stats.get_total()) {
blob->set_horz_stroke_width(0.0f);
blob->set_vert_stroke_width(v_stats.ile(0.5f));
} else {
blob->set_horz_stroke_width(h_stats.get_total() > 2 ? h_stats.ile(0.5f)
: 0.0f);
blob->set_vert_stroke_width(0.0f);
}
}
#else
// Without leptonica present, use the 2*area/perimeter as an approximation.
float width = 2.0f * blob->cblob()->area();
width /= blob->cblob()->perimeter();
blob->set_horz_stroke_width(width);
blob->set_vert_stroke_width(width);
#endif
}
