/*!
* 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;
}
int main(int argc,
char **argv)
{
char *filein, *fileout;
l_int32 x, y, n, i;
PIX *pixs;
PTA *pta;
PTAA *ptaa, *ptaa2, *ptaa3;
static char mainName[] = "cornertest";
if (argc != 3)
return ERROR_INT(" Syntax: cornertest filein fileout", mainName, 1);
filein = argv[1];
fileout = argv[2];
if ((pixs = pixRead(filein)) == NULL)
return ERROR_INT("pixs not made", mainName, 1);
/* Clean noise in LR corner of witten.tif */
pixSetPixel(pixs, 2252, 3051, 0);
pixSetPixel(pixs, 2252, 3050, 0);
pixSetPixel(pixs, 2251, 3050, 0);
pta = pixFindCornerPixels(pixs);
ptaWriteStream(stderr, pta, 1);
/* Test pta and ptaa I/O */
#if 1
ptaa = ptaaCreate(3);
ptaaAddPta(ptaa, pta, L_COPY);
ptaaAddPta(ptaa, pta, L_COPY);
ptaaAddPta(ptaa, pta, L_COPY);
ptaaWriteStream(stderr, ptaa, 1);
ptaaWrite("/tmp/junkptaa", ptaa, 1);
ptaa2 = ptaaRead("/tmp/junkptaa");
ptaaWrite("/tmp/junkptaa2", ptaa2, 1);
ptaaWrite("/tmp/junkptaa3", ptaa, 0);
ptaa3 = ptaaRead("/tmp/junkptaa3");
ptaaWrite("/tmp/junkptaa4", ptaa3, 0);
ptaaDestroy(&ptaa);
ptaaDestroy(&ptaa2);
ptaaDestroy(&ptaa3);
#endif
/* mark corner pixels */
n = ptaGetCount(pta);
for (i = 0; i < n; i++) {
ptaGetIPt(pta, i, &x, &y);
pixRenderLine(pixs, x - LINE_SIZE, y, x + LINE_SIZE, y, 5,
L_FLIP_PIXELS);
pixRenderLine(pixs, x, y - LINE_SIZE, x, y + LINE_SIZE, 5,
L_FLIP_PIXELS);
}
pixWrite(fileout, pixs, IFF_PNG);
pixDestroy(&pixs);
ptaDestroy(&pta);
ptaDestroy(&pta);
return 0;
}
/*!
* pixGetRegionsBinary()
*
* Input: pixs (1 bpp, assumed to be 300 to 400 ppi)
* &pixhm (<optional return> halftone mask)
* &pixtm (<optional return> textline mask)
* &pixtb (<optional return> textblock mask)
* debug (flag: set to 1 for debug output)
* Return: 0 if OK, 1 on error
*
* Notes:
* (1) It is best to deskew the image before segmenting.
* (2) The debug flag enables a number of outputs. These
* are included to show how to generate and save/display
* these results.
*/
l_int32
pixGetRegionsBinary(PIX *pixs,
PIX **ppixhm,
PIX **ppixtm,
PIX **ppixtb,
l_int32 debug)
{
char *tempname;
l_int32 htfound, tlfound;
PIX *pixr, *pixt1, *pixt2;
PIX *pixtext; /* text pixels only */
PIX *pixhm2; /* halftone mask; 2x reduction */
PIX *pixhm; /* halftone mask; */
PIX *pixtm2; /* textline mask; 2x reduction */
PIX *pixtm; /* textline mask */
PIX *pixvws; /* vertical white space mask */
PIX *pixtb2; /* textblock mask; 2x reduction */
PIX *pixtbf2; /* textblock mask; 2x reduction; small comps filtered */
PIX *pixtb; /* textblock mask */
PROCNAME("pixGetRegionsBinary");
if (ppixhm) *ppixhm = NULL;
if (ppixtm) *ppixtm = NULL;
if (ppixtb) *ppixtb = NULL;
if (!pixs)
return ERROR_INT("pixs not defined", procName, 1);
if (pixGetDepth(pixs) != 1)
return ERROR_INT("pixs not 1 bpp", procName, 1);
/* 2x reduce, to 150 -200 ppi */
pixr = pixReduceRankBinaryCascade(pixs, 1, 0, 0, 0);
pixDisplayWrite(pixr, debug);
/* Get the halftone mask */
pixhm2 = pixGenHalftoneMask(pixr, &pixtext, &htfound, debug);
/* Get the textline mask from the text pixels */
pixtm2 = pixGenTextlineMask(pixtext, &pixvws, &tlfound, debug);
/* Get the textblock mask from the textline mask */
pixtb2 = pixGenTextblockMask(pixtm2, pixvws, debug);
pixDestroy(&pixr);
pixDestroy(&pixtext);
pixDestroy(&pixvws);
/* Remove small components from the mask, where a small
* component is defined as one with both width and height < 60 */
pixtbf2 = pixSelectBySize(pixtb2, 60, 60, 4, L_SELECT_IF_EITHER,
L_SELECT_IF_GTE, NULL);
pixDestroy(&pixtb2);
pixDisplayWriteFormat(pixtbf2, debug, IFF_PNG);
/* Expand all masks to full resolution, and do filling or
* small dilations for better coverage. */
pixhm = pixExpandReplicate(pixhm2, 2);
pixt1 = pixSeedfillBinary(NULL, pixhm, pixs, 8);
pixOr(pixhm, pixhm, pixt1);
pixDestroy(&pixt1);
pixDisplayWriteFormat(pixhm, debug, IFF_PNG);
pixt1 = pixExpandReplicate(pixtm2, 2);
pixtm = pixDilateBrick(NULL, pixt1, 3, 3);
pixDestroy(&pixt1);
pixDisplayWriteFormat(pixtm, debug, IFF_PNG);
pixt1 = pixExpandReplicate(pixtbf2, 2);
pixtb = pixDilateBrick(NULL, pixt1, 3, 3);
pixDestroy(&pixt1);
pixDisplayWriteFormat(pixtb, debug, IFF_PNG);
pixDestroy(&pixhm2);
pixDestroy(&pixtm2);
pixDestroy(&pixtbf2);
/* Debug: identify objects that are neither text nor halftone image */
if (debug) {
pixt1 = pixSubtract(NULL, pixs, pixtm); /* remove text pixels */
pixt2 = pixSubtract(NULL, pixt1, pixhm); /* remove halftone pixels */
pixDisplayWriteFormat(pixt2, 1, IFF_PNG);
pixDestroy(&pixt1);
pixDestroy(&pixt2);
}
/* Debug: display textline components with random colors */
if (debug) {
l_int32 w, h;
BOXA *boxa;
PIXA *pixa;
boxa = pixConnComp(pixtm, &pixa, 8);
pixGetDimensions(pixtm, &w, &h, NULL);
pixt1 = pixaDisplayRandomCmap(pixa, w, h);
pixcmapResetColor(pixGetColormap(pixt1), 0, 255, 255, 255);
pixDisplay(pixt1, 100, 100);
pixDisplayWriteFormat(pixt1, 1, IFF_PNG);
pixaDestroy(&pixa);
boxaDestroy(&boxa);
pixDestroy(&pixt1);
}
/* Debug: identify the outlines of each textblock */
if (debug) {
PIXCMAP *cmap;
PTAA *ptaa;
ptaa = pixGetOuterBordersPtaa(pixtb);
tempname = genTempFilename("/tmp", "tb_outlines.ptaa", 0, 0);
ptaaWrite(tempname, ptaa, 1);
FREE(tempname);
pixt1 = pixRenderRandomCmapPtaa(pixtb, ptaa, 1, 16, 1);
cmap = pixGetColormap(pixt1);
pixcmapResetColor(cmap, 0, 130, 130, 130);
pixDisplay(pixt1, 500, 100);
pixDisplayWriteFormat(pixt1, 1, IFF_PNG);
pixDestroy(&pixt1);
ptaaDestroy(&ptaa);
}
/* Debug: get b.b. for all mask components */
if (debug) {
BOXA *bahm, *batm, *batb;
bahm = pixConnComp(pixhm, NULL, 4);
batm = pixConnComp(pixtm, NULL, 4);
batb = pixConnComp(pixtb, NULL, 4);
tempname = genTempFilename("/tmp", "htmask.boxa", 0, 0);
boxaWrite(tempname, bahm);
FREE(tempname);
tempname = genTempFilename("/tmp", "textmask.boxa", 0, 0);
boxaWrite(tempname, batm);
FREE(tempname);
tempname = genTempFilename("/tmp", "textblock.boxa", 0, 0);
boxaWrite(tempname, batb);
FREE(tempname);
boxaDestroy(&bahm);
boxaDestroy(&batm);
boxaDestroy(&batb);
}
if (ppixhm)
*ppixhm = pixhm;
else
pixDestroy(&pixhm);
if (ppixtm)
*ppixtm = pixtm;
else
pixDestroy(&pixtm);
if (ppixtb)
*ppixtb = pixtb;
else
pixDestroy(&pixtb);
return 0;
}
