/*!
* fpixBuildHorizontalDisparity()
*
* Input: fpixv (vertical disparity model)
* factor (conversion factor for vertical disparity slope;
* use 0 for default)
* &extraw (<return> extra width to be added to dewarped pix)
* Return: fpixh, or null on error
*
* Notes:
* (1) This takes the difference in vertical disparity at top
* and bottom of the image, and converts it to an assumed
* horizontal disparity.
*/
FPIX *
fpixBuildHorizontalDisparity(FPIX *fpixv,
l_float32 factor,
l_int32 *pextraw)
{
l_int32 w, h, i, j, fw, wpl, maxloc;
l_float32 val1, val2, vdisp, vdisp0, maxval;
l_float32 *data, *line, *fadiff;
NUMA *nadiff;
FPIX *fpixh;
PROCNAME("fpixBuildHorizontalDisparity");
if (!fpixv)
return (FPIX *)ERROR_PTR("fpixv not defined", procName, NULL);
if (!pextraw)
return (FPIX *)ERROR_PTR("&extraw not defined", procName, NULL);
if (factor == 0.0)
factor = DEFAULT_SLOPE_FACTOR;
/* Estimate horizontal disparity from the vertical disparity
* difference between the top and bottom, normalized to the
* image height. Add the maximum value to the width of the
* output image, so that all src pixels can be mapped
* into the dest. */
fpixGetDimensions(fpixv, &w, &h);
nadiff = numaCreate(w);
for (j = 0; j < w; j++) {
fpixGetPixel(fpixv, j, 0, &val1);
fpixGetPixel(fpixv, j, h - 1, &val2);
vdisp = factor * (val2 - val1) / (l_float32)h;
if (j == 0) vdisp0 = vdisp;
vdisp = vdisp0 - vdisp;
numaAddNumber(nadiff, vdisp);
}
numaGetMax(nadiff, &maxval, &maxloc);
*pextraw = (l_int32)(maxval + 0.5);
fw = w + *pextraw;
fpixh = fpixCreate(fw, h);
data = fpixGetData(fpixh);
wpl = fpixGetWpl(fpixh);
fadiff = numaGetFArray(nadiff, L_NOCOPY);
for (i = 0; i < h; i++) {
line = data + i * wpl;
for (j = 0; j < fw; j++) {
if (j < maxloc) /* this may not work for even pages */
line[j] = fadiff[j];
else /* keep it at the max value the rest of the way across */
line[j] = maxval;
}
}
numaDestroy(&nadiff);
return fpixh;
}
/*!
* \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;
}
/*!
* dewarpBuildModel()
*
* Input: dew
* debugflag (1 for debugging output)
* Return: 0 if OK, 1 on error
*
* Notes:
* (1) This is the basic function that builds the vertical
* disparity array, which allows determination of the
* src pixel in the input image corresponding to each
* dest pixel in the dewarped image.
* (2) The method is as follows:
* * Estimate the centers of all the long textlines and
* fit a LS quadratic to each one. This smooths the curves.
* * Sample each curve at a regular interval, find the y-value
* of the flat point on each curve, and subtract the sampled
* curve value from this value. This is the vertical
* disparity.
* * Fit a LS quadratic to each set of vertically aligned
* disparity samples. This smooths the disparity values
* in the vertical direction. Then resample at the same
* regular interval, We now have a regular grid of smoothed
* vertical disparity valuels.
* * Interpolate this grid to get a full resolution disparity
* map. This can be applied directly to the src image
* pixels to dewarp the image in the vertical direction,
* making all textlines horizontal.
*/
l_int32
dewarpBuildModel(L_DEWARP *dew,
l_int32 debugflag)
{
char *tempname;
l_int32 i, j, nlines, nx, ny, sampling;
l_float32 c0, c1, c2, x, y, flaty, val;
l_float32 *faflats;
NUMA *nax, *nafit, *nacurve, *nacurves, *naflat, *naflats, *naflatsi;
PIX *pixs, *pixt1, *pixt2;
PTA *pta, *ptad;
PTAA *ptaa1, *ptaa2, *ptaa3, *ptaa4, *ptaa5, *ptaa6, *ptaa7;
FPIX *fpix1, *fpix2, *fpix3;
PROCNAME("dewarpBuildModel");
if (!dew)
return ERROR_INT("dew not defined", procName, 1);
pixs = dew->pixs;
if (debugflag) {
pixDisplayWithTitle(pixs, 0, 0, "pixs", 1);
pixWriteTempfile("/tmp", "pixs.png", pixs, IFF_PNG, NULL);
}
/* Make initial estimate of centers of textlines */
ptaa1 = pixGetTextlineCenters(pixs, DEBUG_TEXTLINE_CENTERS);
if (debugflag) {
pixt1 = pixConvertTo32(pixs);
pixt2 = pixDisplayPtaa(pixt1, ptaa1);
pixWriteTempfile("/tmp", "lines1.png", pixt2, IFF_PNG, NULL);
pixDestroy(&pixt1);
pixDestroy(&pixt2);
}
/* Remove all lines that are not near the length
* of the longest line. */
ptaa2 = ptaaRemoveShortLines(pixs, ptaa1, 0.8, DEBUG_SHORT_LINES);
if (debugflag) {
pixt1 = pixConvertTo32(pixs);
pixt2 = pixDisplayPtaa(pixt1, ptaa2);
pixWriteTempfile("/tmp", "lines2.png", pixt2, IFF_PNG, NULL);
pixDestroy(&pixt1);
pixDestroy(&pixt2);
}
nlines = ptaaGetCount(ptaa2);
if (nlines < dew->minlines)
return ERROR_INT("insufficient lines to build model", procName, 1);
/* Do quadratic fit to smooth each line. A single quadratic
* over the entire width of the line appears to be sufficient.
* Quartics tend to overfit to noise. Each line is thus
* represented by three coefficients: c2 * x^2 + c1 * x + c0.
* Using the coefficients, sample each fitted curve uniformly
* across the full width of the image. */
sampling = dew->sampling;
nx = dew->nx;
ny = dew->ny;
ptaa3 = ptaaCreate(nlines);
nacurve = numaCreate(nlines); /* stores curvature coeff c2 */
for (i = 0; i < nlines; i++) { /* for each line */
pta = ptaaGetPta(ptaa2, i, L_CLONE);
ptaGetQuadraticLSF(pta, &c2, &c1, &c0, NULL);
numaAddNumber(nacurve, c2);
ptad = ptaCreate(nx);
for (j = 0; j < nx; j++) { /* uniformly sampled in x */
x = j * sampling;
applyQuadraticFit(c2, c1, c0, x, &y);
ptaAddPt(ptad, x, y);
}
ptaaAddPta(ptaa3, ptad, L_INSERT);
ptaDestroy(&pta);
}
if (debugflag) {
ptaa4 = ptaaCreate(nlines);
for (i = 0; i < nlines; i++) {
pta = ptaaGetPta(ptaa2, i, L_CLONE);
ptaGetArrays(pta, &nax, NULL);
ptaGetQuadraticLSF(pta, NULL, NULL, NULL, &nafit);
ptad = ptaCreateFromNuma(nax, nafit);
ptaaAddPta(ptaa4, ptad, L_INSERT);
ptaDestroy(&pta);
numaDestroy(&nax);
numaDestroy(&nafit);
}
pixt1 = pixConvertTo32(pixs);
pixt2 = pixDisplayPtaa(pixt1, ptaa4);
pixWriteTempfile("/tmp", "lines3.png", pixt2, IFF_PNG, NULL);
pixDestroy(&pixt1);
pixDestroy(&pixt2);
ptaaDestroy(&ptaa4);
}
/* Find and save the flat points in each curve. */
naflat = numaCreate(nlines);
for (i = 0; i < nlines; i++) {
pta = ptaaGetPta(ptaa3, i, L_CLONE);
numaGetFValue(nacurve, i, &c2);
if (c2 <= 0) /* flat point at bottom; max value of y in curve */
ptaGetRange(pta, NULL, NULL, NULL, &flaty);
else /* flat point at top; min value of y in curve */
ptaGetRange(pta, NULL, NULL, &flaty, NULL);
numaAddNumber(naflat, flaty);
ptaDestroy(&pta);
}
/* Sort the lines in ptaa3 by their position */
naflatsi = numaGetSortIndex(naflat, L_SORT_INCREASING);
naflats = numaSortByIndex(naflat, naflatsi);
nacurves = numaSortByIndex(nacurve, naflatsi);
dew->naflats = naflats;
dew->nacurves = nacurves;
ptaa4 = ptaaSortByIndex(ptaa3, naflatsi);
numaDestroy(&naflat);
numaDestroy(&nacurve);
numaDestroy(&naflatsi);
if (debugflag) {
tempname = genTempFilename("/tmp", "naflats.na", 0);
numaWrite(tempname, naflats);
FREE(tempname);
}
/* Convert the sampled points in ptaa3 to a sampled disparity with
* with respect to the flat point in the curve. */
ptaa5 = ptaaCreate(nlines);
for (i = 0; i < nlines; i++) {
pta = ptaaGetPta(ptaa4, i, L_CLONE);
numaGetFValue(naflats, i, &flaty);
ptad = ptaCreate(nx);
for (j = 0; j < nx; j++) {
ptaGetPt(pta, j, &x, &y);
ptaAddPt(ptad, x, flaty - y);
}
ptaaAddPta(ptaa5, ptad, L_INSERT);
ptaDestroy(&pta);
}
if (debugflag) {
tempname = genTempFilename("/tmp", "ptaa5.ptaa", 0);
ptaaWrite(tempname, ptaa5, 0);
FREE(tempname);
}
/* Generate a ptaa taking vertical 'columns' from ptaa5.
* We want to fit the vertical disparity on the column to the
* vertical position of the line, which we call 'y' here and
* obtain from naflats. */
ptaa6 = ptaaCreate(nx);
faflats = numaGetFArray(naflats, L_NOCOPY);
for (j = 0; j < nx; j++) {
pta = ptaCreate(nlines);
for (i = 0; i < nlines; i++) {
y = faflats[i];
ptaaGetPt(ptaa5, i, j, NULL, &val); /* disparity value */
ptaAddPt(pta, y, val);
}
ptaaAddPta(ptaa6, pta, L_INSERT);
}
if (debugflag) {
tempname = genTempFilename("/tmp", "ptaa6.ptaa", 0);
ptaaWrite(tempname, ptaa6, 0);
FREE(tempname);
}
/* Do quadratic fit vertically on a subset of pixel columns
* for the vertical displacement, which identifies the
* src pixel(s) for each dest pixel. Sample the displacement
* on a regular grid in the vertical direction. */
ptaa7 = ptaaCreate(nx); /* uniformly sampled across full height of image */
for (j = 0; j < nx; j++) { /* for each column */
pta = ptaaGetPta(ptaa6, j, L_CLONE);
ptaGetQuadraticLSF(pta, &c2, &c1, &c0, NULL);
ptad = ptaCreate(ny);
for (i = 0; i < ny; i++) { /* uniformly sampled in y */
y = i * sampling;
applyQuadraticFit(c2, c1, c0, y, &val);
ptaAddPt(ptad, y, val);
}
ptaaAddPta(ptaa7, ptad, L_INSERT);
ptaDestroy(&pta);
}
if (debugflag) {
tempname = genTempFilename("/tmp", "ptaa7.ptaa", 0);
ptaaWrite(tempname, ptaa7, 0);
FREE(tempname);
}
/* Save the result in a fpix at the specified subsampling */
fpix1 = fpixCreate(nx, ny);
for (i = 0; i < ny; i++) {
for (j = 0; j < nx; j++) {
ptaaGetPt(ptaa7, j, i, NULL, &val);
fpixSetPixel(fpix1, j, i, val);
}
}
dew->sampvdispar = fpix1;
/* Generate a full res fpix for vertical dewarping. We require that
* the size of this fpix is at least as big as the input image. */
fpix2 = fpixScaleByInteger(fpix1, sampling);
dew->fullvdispar = fpix2;
if (debugflag) {
pixt1 = fpixRenderContours(fpix2, -2., 2.0, 0.2);
pixWriteTempfile("/tmp", "vert-contours.png", pixt1, IFF_PNG, NULL);
pixDisplay(pixt1, 1000, 0);
pixDestroy(&pixt1);
}
/* Generate full res and sampled fpix for horizontal dewarping. This
* works to the extent that the line curvature is due to bending
* out of the plane normal to the camera, and not wide-angle
* "fishbowl" distortion. Also generate the sampled horizontal
* disparity array. */
if (dew->applyhoriz) {
fpix3 = fpixBuildHorizontalDisparity(fpix2, 0, &dew->extraw);
dew->fullhdispar = fpix3;
dew->samphdispar = fpixSampledDisparity(fpix3, dew->sampling);
if (debugflag) {
pixt1 = fpixRenderContours(fpix3, -2., 2.0, 0.2);
pixWriteTempfile("/tmp", "horiz-contours.png", pixt1,
IFF_PNG, NULL);
pixDisplay(pixt1, 1000, 0);
pixDestroy(&pixt1);
}
}
dew->success = 1;
ptaaDestroy(&ptaa1);
ptaaDestroy(&ptaa2);
ptaaDestroy(&ptaa3);
ptaaDestroy(&ptaa4);
ptaaDestroy(&ptaa5);
ptaaDestroy(&ptaa6);
ptaaDestroy(&ptaa7);
return 0;
}
