/*!
* 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);
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);
boxaWrite(tempname, bahm);
FREE(tempname);
tempname = genTempFilename("/tmp", "textmask.boxa", 0);
boxaWrite(tempname, batm);
FREE(tempname);
tempname = genTempFilename("/tmp", "textblock.boxa", 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;
}
/*!
* pixMorphCompSequence()
*
* Input: pixs
* sequence (string specifying sequence)
* dispsep (controls debug display of each result in the sequence:
* 0: no output
* > 0: gives horizontal separation in pixels between
* successive displays
* < 0: pdf output; abs(dispsep) is used for naming)
* Return: pixd, or null on error
*
* Notes:
* (1) This does rasterop morphology on binary images, using composite
* operations for extra speed on large Sels.
* (2) Safe closing is used atomically. However, if you implement a
* closing as a sequence with a dilation followed by an
* erosion, it will not be safe, and to ensure that you have
* no boundary effects you must add a border in advance and
* remove it at the end.
* (3) For other usage details, see the notes for pixMorphSequence().
* (4) The sequence string is formatted as follows:
* - An arbitrary number of operations, each separated
* by a '+' character. White space is ignored.
* - Each operation begins with a case-independent character
* specifying the operation:
* d or D (dilation)
* e or E (erosion)
* o or O (opening)
* c or C (closing)
* r or R (rank binary reduction)
* x or X (replicative binary expansion)
* b or B (add a border of 0 pixels of this size)
* - The args to the morphological operations are bricks of hits,
* and are formatted as a.b, where a and b are horizontal and
* vertical dimensions, rsp.
* - The args to the reduction are a sequence of up to 4 integers,
* each from 1 to 4.
* - The arg to the expansion is a power of two, in the set
* {2, 4, 8, 16}.
*/
PIX *
pixMorphCompSequence(PIX *pixs,
const char *sequence,
l_int32 dispsep)
{
char *rawop, *op, *fname;
char buf[256];
l_int32 nops, i, j, nred, fact, w, h, x, y, border, pdfout;
l_int32 level[4];
PIX *pixt1, *pixt2;
PIXA *pixa;
SARRAY *sa;
PROCNAME("pixMorphCompSequence");
if (!pixs)
return (PIX *)ERROR_PTR("pixs not defined", procName, NULL);
if (!sequence)
return (PIX *)ERROR_PTR("sequence not defined", procName, NULL);
/* Split sequence into individual operations */
sa = sarrayCreate(0);
sarraySplitString(sa, sequence, "+");
nops = sarrayGetCount(sa);
pdfout = (dispsep < 0) ? 1 : 0;
if (!morphSequenceVerify(sa)) {
sarrayDestroy(&sa);
return (PIX *)ERROR_PTR("sequence not valid", procName, NULL);
}
/* Parse and operate */
pixa = NULL;
if (pdfout) {
pixa = pixaCreate(0);
pixaAddPix(pixa, pixs, L_CLONE);
snprintf(buf, sizeof(buf), "/tmp/seq_output_%d.pdf", L_ABS(dispsep));
fname = genPathname(buf, NULL);
}
border = 0;
pixt1 = pixCopy(NULL, pixs);
pixt2 = NULL;
x = y = 0;
for (i = 0; i < nops; i++) {
rawop = sarrayGetString(sa, i, 0);
op = stringRemoveChars(rawop, " \n\t");
switch (op[0])
{
case 'd':
case 'D':
sscanf(&op[1], "%d.%d", &w, &h);
pixt2 = pixDilateCompBrick(NULL, pixt1, w, h);
pixSwapAndDestroy(&pixt1, &pixt2);
break;
case 'e':
case 'E':
sscanf(&op[1], "%d.%d", &w, &h);
pixt2 = pixErodeCompBrick(NULL, pixt1, w, h);
pixSwapAndDestroy(&pixt1, &pixt2);
break;
case 'o':
case 'O':
sscanf(&op[1], "%d.%d", &w, &h);
pixOpenCompBrick(pixt1, pixt1, w, h);
break;
case 'c':
case 'C':
sscanf(&op[1], "%d.%d", &w, &h);
pixCloseSafeCompBrick(pixt1, pixt1, w, h);
break;
case 'r':
case 'R':
nred = strlen(op) - 1;
for (j = 0; j < nred; j++)
level[j] = op[j + 1] - '0';
for (j = nred; j < 4; j++)
level[j] = 0;
pixt2 = pixReduceRankBinaryCascade(pixt1, level[0], level[1],
level[2], level[3]);
pixSwapAndDestroy(&pixt1, &pixt2);
break;
case 'x':
case 'X':
sscanf(&op[1], "%d", &fact);
pixt2 = pixExpandReplicate(pixt1, fact);
pixSwapAndDestroy(&pixt1, &pixt2);
break;
case 'b':
case 'B':
sscanf(&op[1], "%d", &border);
pixt2 = pixAddBorder(pixt1, border, 0);
pixSwapAndDestroy(&pixt1, &pixt2);
break;
default:
/* All invalid ops are caught in the first pass */
break;
}
FREE(op);
/* Debug output */
if (dispsep > 0) {
pixDisplay(pixt1, x, y);
x += dispsep;
}
if (pdfout)
pixaAddPix(pixa, pixt1, L_COPY);
}
if (border > 0) {
pixt2 = pixRemoveBorder(pixt1, border);
pixSwapAndDestroy(&pixt1, &pixt2);
}
if (pdfout) {
pixaConvertToPdf(pixa, 0, 1.0, L_FLATE_ENCODE, 0, fname, fname);
FREE(fname);
pixaDestroy(&pixa);
}
sarrayDestroy(&sa);
return pixt1;
}
/*!
* \brief pixUpDownDetectGeneralDwa()
*
* \param[in] pixs 1 bpp, deskewed, English text
* \param[out] pconf confidence that text is rightside-up
* \param[in] mincount min number of up + down; use 0 for default
* \param[in] npixels number of pixels removed from each side of word box
* \param[in] debug 1 for debug output; 0 otherwise
* \return 0 if OK, 1 on error
*
* <pre>
* Notes:
* (1) See the notes in pixUpDownDetectGeneral() for usage.
* </pre>
*/
l_int32
pixUpDownDetectGeneralDwa(PIX *pixs,
l_float32 *pconf,
l_int32 mincount,
l_int32 npixels,
l_int32 debug)
{
char flipsel1[] = "flipsel1";
char flipsel2[] = "flipsel2";
char flipsel3[] = "flipsel3";
char flipsel4[] = "flipsel4";
l_int32 countup, countdown, nmax;
l_float32 nup, ndown;
PIX *pixt, *pix0, *pix1, *pix2, *pix3, *pixm;
PROCNAME("pixUpDownDetectGeneralDwa");
if (!pconf)
return ERROR_INT("&conf not defined", procName, 1);
*pconf = 0.0;
if (!pixs || pixGetDepth(pixs) != 1)
return ERROR_INT("pixs not defined or not 1 bpp", procName, 1);
if (mincount == 0)
mincount = DEFAULT_MIN_UP_DOWN_COUNT;
if (npixels < 0)
npixels = 0;
lept_mkdir("lept/orient");
/* One of many reasonable pre-filtering sequences: (1, 8) and (30, 1).
* This closes holes in x-height characters and joins them at
* the x-height. There is more noise in the descender detection
* from this, but it works fairly well. */
pixt = pixMorphSequenceDwa(pixs, "c1.8 + c30.1", 0);
/* Be sure to add the border before the flip DWA operations! */
pix0 = pixAddBorderGeneral(pixt, ADDED_BORDER, ADDED_BORDER,
ADDED_BORDER, ADDED_BORDER, 0);
pixDestroy(&pixt);
/* Optionally, make a mask of the word bounding boxes, shortening
* each of them by a fixed amount at each end. */
pixm = NULL;
if (npixels > 0) {
l_int32 i, nbox, x, y, w, h;
BOX *box;
BOXA *boxa;
pix1 = pixMorphSequenceDwa(pix0, "o10.1", 0);
boxa = pixConnComp(pix1, NULL, 8);
pixm = pixCreateTemplate(pix1);
pixDestroy(&pix1);
nbox = boxaGetCount(boxa);
for (i = 0; i < nbox; i++) {
box = boxaGetBox(boxa, i, L_CLONE);
boxGetGeometry(box, &x, &y, &w, &h);
if (w > 2 * npixels)
pixRasterop(pixm, x + npixels, y - 6, w - 2 * npixels, h + 13,
PIX_SET, NULL, 0, 0);
boxDestroy(&box);
}
boxaDestroy(&boxa);
}
/* Find the ascenders and optionally filter with pixm.
* For an explanation of the procedure used for counting the result
* of the HMT, see comments in pixUpDownDetectGeneral(). */
pix1 = pixFlipFHMTGen(NULL, pix0, flipsel1);
pix2 = pixFlipFHMTGen(NULL, pix0, flipsel2);
pixOr(pix1, pix1, pix2);
if (pixm)
pixAnd(pix1, pix1, pixm);
pix3 = pixReduceRankBinaryCascade(pix1, 1, 1, 0, 0);
pixCountPixels(pix3, &countup, NULL);
pixDestroy(&pix1);
pixDestroy(&pix2);
pixDestroy(&pix3);
/* Find the ascenders and optionally filter with pixm. */
pix1 = pixFlipFHMTGen(NULL, pix0, flipsel3);
pix2 = pixFlipFHMTGen(NULL, pix0, flipsel4);
pixOr(pix1, pix1, pix2);
if (pixm)
pixAnd(pix1, pix1, pixm);
pix3 = pixReduceRankBinaryCascade(pix1, 1, 1, 0, 0);
pixCountPixels(pix3, &countdown, NULL);
pixDestroy(&pix1);
pixDestroy(&pix2);
pixDestroy(&pix3);
/* Evaluate statistically, generating a confidence that is
* related to the probability with a gaussian distribution. */
nup = (l_float32)(countup);
ndown = (l_float32)(countdown);
nmax = L_MAX(countup, countdown);
if (nmax > mincount)
*pconf = 2. * ((nup - ndown) / sqrt(nup + ndown));
if (debug) {
if (pixm) pixWriteDebug("/tmp/lept/orient/pixm2.png", pixm, IFF_PNG);
fprintf(stderr, "nup = %7.3f, ndown = %7.3f, conf = %7.3f\n",
nup, ndown, *pconf);
if (*pconf > DEFAULT_MIN_UP_DOWN_CONF)
fprintf(stderr, "Text is rightside-up\n");
if (*pconf < -DEFAULT_MIN_UP_DOWN_CONF)
fprintf(stderr, "Text is upside-down\n");
}
pixDestroy(&pix0);
pixDestroy(&pixm);
return 0;
}
/*!
* \brief pixMirrorDetect()
*
* \param[in] pixs 1 bpp, deskewed, English text
* \param[out] pconf confidence that text is not LR mirror reversed
* \param[in] mincount min number of left + right; use 0 for default
* \param[in] debug 1 for debug output; 0 otherwise
* \return 0 if OK, 1 on error
*
* <pre>
* Notes:
* (1) For this test, it is necessary that the text is horizontally
* oriented, with ascenders going up.
* (2) conf is the normalized difference between the number of
* right and left facing characters with ascenders.
* Left-facing are {d}; right-facing are {b, h, k}.
* At least that was the expectation. In practice, we can
* really just say that it is the normalized difference in
* hits using two specific hit-miss filters, textsel1 and textsel2,
* after the image has been suitably pre-filtered so that
* these filters are effective. See (4) for what's really happening.
* (3) A large positive conf value indicates normal text, whereas
* a large negative conf value means the page is mirror reversed.
* (4) The implementation is a bit tricky. The general idea is
* to fill the x-height part of characters, but not the space
* between them, before doing the HMT. This is done by
* finding pixels added using two different operations -- a
* horizontal close and a vertical dilation -- and adding
* the intersection of these sets to the original. It turns
* out that the original intuition about the signal was largely
* in error: much of the signal for right-facing characters
* comes from the lower part of common x-height characters, like
* the e and c, that remain open after these operations.
* So it's important that the operations to close the x-height
* parts of the characters are purposely weakened sufficiently
* to allow these characters to remain open. The wonders
* of morphology!
* </pre>
*/
l_int32
pixMirrorDetect(PIX *pixs,
l_float32 *pconf,
l_int32 mincount,
l_int32 debug)
{
l_int32 count1, count2, nmax;
l_float32 nleft, nright;
PIX *pix0, *pix1, *pix2, *pix3;
SEL *sel1, *sel2;
PROCNAME("pixMirrorDetect");
if (!pconf)
return ERROR_INT("&conf not defined", procName, 1);
*pconf = 0.0;
if (!pixs || pixGetDepth(pixs) != 1)
return ERROR_INT("pixs not defined or not 1 bpp", procName, 1);
if (mincount == 0)
mincount = DEFAULT_MIN_MIRROR_FLIP_COUNT;
sel1 = selCreateFromString(textsel1, 5, 6, NULL);
sel2 = selCreateFromString(textsel2, 5, 6, NULL);
/* Fill x-height characters but not space between them, sort of. */
pix3 = pixMorphCompSequence(pixs, "d1.30", 0);
pixXor(pix3, pix3, pixs);
pix0 = pixMorphCompSequence(pixs, "c15.1", 0);
pixXor(pix0, pix0, pixs);
pixAnd(pix0, pix0, pix3);
pixOr(pix0, pix0, pixs);
pixDestroy(&pix3);
/* Filter the right-facing characters. */
pix1 = pixHMT(NULL, pix0, sel1);
pix3 = pixReduceRankBinaryCascade(pix1, 1, 1, 0, 0);
pixCountPixels(pix3, &count1, NULL);
pixDebugFlipDetect("/tmp/lept/orient/right.png", pixs, pix1, debug);
pixDestroy(&pix1);
pixDestroy(&pix3);
/* Filter the left-facing characters. */
pix2 = pixHMT(NULL, pix0, sel2);
pix3 = pixReduceRankBinaryCascade(pix2, 1, 1, 0, 0);
pixCountPixels(pix3, &count2, NULL);
pixDebugFlipDetect("/tmp/lept/orient/left.png", pixs, pix2, debug);
pixDestroy(&pix2);
pixDestroy(&pix3);
nright = (l_float32)count1;
nleft = (l_float32)count2;
nmax = L_MAX(count1, count2);
pixDestroy(&pix0);
selDestroy(&sel1);
selDestroy(&sel2);
if (nmax > mincount)
*pconf = 2. * ((nright - nleft) / sqrt(nright + nleft));
if (debug) {
fprintf(stderr, "nright = %f, nleft = %f\n", nright, nleft);
if (*pconf > DEFAULT_MIN_MIRROR_FLIP_CONF)
fprintf(stderr, "Text is not mirror reversed\n");
if (*pconf < -DEFAULT_MIN_MIRROR_FLIP_CONF)
fprintf(stderr, "Text is mirror reversed\n");
}
return 0;
}
/*!
* \brief pixMirrorDetectDwa()
*
* \param[in] pixs 1 bpp, deskewed, English text
* \param[out] pconf confidence that text is not LR mirror reversed
* \param[in] mincount min number of left + right; use 0 for default
* \param[in] debug 1 for debug output; 0 otherwise
* \return 0 if OK, 1 on error
*
* <pre>
* Notes:
* (1) We assume the text is horizontally oriented, with
* ascenders going up.
* (2) See notes in pixMirrorDetect().
* </pre>
*/
l_int32
pixMirrorDetectDwa(PIX *pixs,
l_float32 *pconf,
l_int32 mincount,
l_int32 debug)
{
char flipsel1[] = "flipsel1";
char flipsel2[] = "flipsel2";
l_int32 count1, count2, nmax;
l_float32 nleft, nright;
PIX *pix0, *pix1, *pix2, *pix3;
PROCNAME("pixMirrorDetectDwa");
if (!pconf)
return ERROR_INT("&conf not defined", procName, 1);
*pconf = 0.0;
if (!pixs || pixGetDepth(pixs) != 1)
return ERROR_INT("pixs not defined or not 1 bpp", procName, 1);
if (mincount == 0)
mincount = DEFAULT_MIN_MIRROR_FLIP_COUNT;
/* Fill x-height characters but not space between them, sort of. */
pix3 = pixMorphSequenceDwa(pixs, "d1.30", 0);
pixXor(pix3, pix3, pixs);
pix0 = pixMorphSequenceDwa(pixs, "c15.1", 0);
pixXor(pix0, pix0, pixs);
pixAnd(pix0, pix0, pix3);
pixOr(pix3, pix0, pixs);
pixDestroy(&pix0);
pix0 = pixAddBorderGeneral(pix3, ADDED_BORDER, ADDED_BORDER,
ADDED_BORDER, ADDED_BORDER, 0);
pixDestroy(&pix3);
/* Filter the right-facing characters. */
pix1 = pixFlipFHMTGen(NULL, pix0, flipsel1);
pix3 = pixReduceRankBinaryCascade(pix1, 1, 1, 0, 0);
pixCountPixels(pix3, &count1, NULL);
pixDestroy(&pix1);
pixDestroy(&pix3);
/* Filter the left-facing characters. */
pix2 = pixFlipFHMTGen(NULL, pix0, flipsel2);
pix3 = pixReduceRankBinaryCascade(pix2, 1, 1, 0, 0);
pixCountPixels(pix3, &count2, NULL);
pixDestroy(&pix2);
pixDestroy(&pix3);
pixDestroy(&pix0);
nright = (l_float32)count1;
nleft = (l_float32)count2;
nmax = L_MAX(count1, count2);
if (nmax > mincount)
*pconf = 2. * ((nright - nleft) / sqrt(nright + nleft));
if (debug) {
fprintf(stderr, "nright = %f, nleft = %f\n", nright, nleft);
if (*pconf > DEFAULT_MIN_MIRROR_FLIP_CONF)
fprintf(stderr, "Text is not mirror reversed\n");
if (*pconf < -DEFAULT_MIN_MIRROR_FLIP_CONF)
fprintf(stderr, "Text is mirror reversed\n");
}
return 0;
}
/*!
* pixGetWordsInTextlines()
*
* Input: pixs (1 bpp, typ. 300 ppi)
* reduction (1 for input res; 2 for 2x reduction of input res)
* minwidth, minheight (of saved components; smaller are discarded)
* maxwidth, maxheight (of saved components; larger are discarded)
* &boxad (<return> word boxes sorted in textline line order)
* &pixad (<return> word images sorted in textline line order)
* &naindex (<return> index of textline for each word)
* Return: 0 if OK, 1 on error
*
* Notes:
* (1) The input should be at a resolution of about 300 ppi.
* The word masks and word images can be computed at either
* 150 ppi or 300 ppi. For the former, set reduction = 2.
* (2) The four size constraints on saved components are all
* scaled by @reduction.
* (3) The result are word images (and their b.b.), extracted in
* textline order, at either full res or 2x reduction,
* and with a numa giving the textline index for each word.
* (4) The pixa and boxa interfaces should make this type of
* application simple to put together. The steps are:
* - optionally reduce by 2x
* - generate first estimate of word masks
* - get b.b. of these, and remove the small and big ones
* - extract pixa of the word images, using the b.b.
* - sort actual word images in textline order (2d)
* - flatten them to a pixa (1d), saving the textline index
* for each pix
* (5) In an actual application, it may be desirable to pre-filter
* the input image to remove large components, to extract
* single columns of text, and to deskew them. For example,
* to remove both large components and small noisy components
* that can interfere with the statistics used to estimate
* parameters for segmenting by words, but still retain text lines,
* the following image preprocessing can be done:
* Pix *pixt = pixMorphSequence(pixs, "c40.1", 0);
* Pix *pixf = pixSelectBySize(pixt, 0, 60, 8,
* L_SELECT_HEIGHT, L_SELECT_IF_LT, NULL);
* pixAnd(pixf, pixf, pixs); // the filtered image
* The closing turns text lines into long blobs, but does not
* significantly increase their height. But if there are many
* small connected components in a dense texture, this is likely
* to generate tall components that will be eliminated in pixf.
*/
l_int32
pixGetWordsInTextlines(PIX *pixs,
l_int32 reduction,
l_int32 minwidth,
l_int32 minheight,
l_int32 maxwidth,
l_int32 maxheight,
BOXA **pboxad,
PIXA **ppixad,
NUMA **pnai) {
l_int32 maxdil;
BOXA *boxa1, *boxad;
BOXAA *baa;
NUMA *nai;
NUMAA *naa;
PIXA *pixa1, *pixad;
PIX *pix1;
PIXAA *paa;
PROCNAME("pixGetWordsInTextlines");
if (!pboxad || !ppixad || !pnai)
return ERROR_INT("&boxad, &pixad, &nai not all defined", procName, 1);
*pboxad = NULL;
*ppixad = NULL;
*pnai = NULL;
if (!pixs)
return ERROR_INT("pixs not defined", procName, 1);
if (reduction != 1 && reduction != 2)
return ERROR_INT("reduction not in {1,2}", procName, 1);
if (reduction == 1) {
pix1 = pixClone(pixs);
maxdil = 18;
} else { /* reduction == 2 */
pix1 = pixReduceRankBinaryCascade(pixs, 1, 0, 0, 0);
maxdil = 9;
}
/* Get the bounding boxes of the words from the word mask. */
pixWordBoxesByDilation(pix1, maxdil, minwidth, minheight,
maxwidth, maxheight, &boxa1, NULL);
/* Generate a pixa of the word images */
pixa1 = pixaCreateFromBoxa(pix1, boxa1, NULL); /* mask over each word */
/* Sort the bounding boxes of these words by line. We use the
* index mapping to allow identical sorting of the pixa. */
baa = boxaSort2d(boxa1, &naa, -1, -1, 4);
paa = pixaSort2dByIndex(pixa1, naa, L_CLONE);
/* Flatten the word paa */
pixad = pixaaFlattenToPixa(paa, &nai, L_CLONE);
boxad = pixaGetBoxa(pixad, L_COPY);
*pnai = nai;
*pboxad = boxad;
*ppixad = pixad;
pixDestroy(&pix1);
pixaDestroy(&pixa1);
boxaDestroy(&boxa1);
boxaaDestroy(&baa);
pixaaDestroy(&paa);
numaaDestroy(&naa);
return 0;
}
l_int32
DoPageSegmentation(PIX *pixs, /* should be at least 300 ppi */
l_int32 which) /* 1, 2, 3, 4 */
{
char buf[256];
l_int32 zero;
BOXA *boxatm, *boxahm;
PIX *pixr; /* image reduced to 150 ppi */
PIX *pixhs; /* image of halftone seed, 150 ppi */
PIX *pixm; /* image of mask of components, 150 ppi */
PIX *pixhm1; /* image of halftone mask, 150 ppi */
PIX *pixhm2; /* image of halftone mask, 300 ppi */
PIX *pixht; /* image of halftone components, 150 ppi */
PIX *pixnht; /* image without halftone components, 150 ppi */
PIX *pixi; /* inverted image, 150 ppi */
PIX *pixvws; /* image of vertical whitespace, 150 ppi */
PIX *pixtm1; /* image of closed textlines, 150 ppi */
PIX *pixtm2; /* image of refined text line mask, 150 ppi */
PIX *pixtm3; /* image of refined text line mask, 300 ppi */
PIX *pixtb1; /* image of text block mask, 150 ppi */
PIX *pixtb2; /* image of text block mask, 300 ppi */
PIX *pixnon; /* image of non-text or halftone, 150 ppi */
PIX *pixt1, *pixt2, *pixt3;
PIXA *pixa;
PIXCMAP *cmap;
PTAA *ptaa;
l_int32 ht_flag = 0;
l_int32 ws_flag = 0;
l_int32 text_flag = 0;
l_int32 block_flag = 0;
PROCNAME("DoPageSegmentation");
if (which == 1)
ht_flag = 1;
else if (which == 2)
ws_flag = 1;
else if (which == 3)
text_flag = 1;
else if (which == 4)
block_flag = 1;
else
return ERROR_INT("invalid parameter: not in [1...4]", procName, 1);
pixDisplayWrite(NULL, -1);
/* Reduce to 150 ppi */
pixt1 = pixScaleToGray2(pixs);
pixDisplayWriteFormat(pixt1, L_MAX(ws_flag, L_MAX(ht_flag, block_flag)),
IFF_PNG);
if (which == 1) pixWrite("/tmp/lept/orig.gray.150.png", pixt1, IFF_PNG);
pixDestroy(&pixt1);
pixr = pixReduceRankBinaryCascade(pixs, 1, 0, 0, 0);
/* Get seed for halftone parts */
pixt1 = pixReduceRankBinaryCascade(pixr, 4, 4, 3, 0);
pixt2 = pixOpenBrick(NULL, pixt1, 5, 5);
pixhs = pixExpandBinaryPower2(pixt2, 8);
pixDisplayWriteFormat(pixhs, ht_flag, IFF_PNG);
if (which == 1) pixWrite("/tmp/lept/htseed.150.png", pixhs, IFF_PNG);
pixDestroy(&pixt1);
pixDestroy(&pixt2);
/* Get mask for connected regions */
pixm = pixCloseSafeBrick(NULL, pixr, 4, 4);
pixDisplayWriteFormat(pixm, ht_flag, IFF_PNG);
if (which == 1) pixWrite("/tmp/lept/ccmask.150.png", pixm, IFF_PNG);
/* Fill seed into mask to get halftone mask */
pixhm1 = pixSeedfillBinary(NULL, pixhs, pixm, 4);
pixDisplayWriteFormat(pixhm1, ht_flag, IFF_PNG);
if (which == 1) pixWrite("/tmp/lept/htmask.150.png", pixhm1, IFF_PNG);
pixhm2 = pixExpandBinaryPower2(pixhm1, 2);
/* Extract halftone stuff */
pixht = pixAnd(NULL, pixhm1, pixr);
if (which == 1) pixWrite("/tmp/lept/ht.150.png", pixht, IFF_PNG);
/* Extract non-halftone stuff */
pixnht = pixXor(NULL, pixht, pixr);
pixDisplayWriteFormat(pixnht, text_flag, IFF_PNG);
if (which == 1) pixWrite("/tmp/lept/text.150.png", pixnht, IFF_PNG);
pixZero(pixht, &zero);
if (zero)
fprintf(stderr, "No halftone parts found\n");
else
fprintf(stderr, "Halftone parts found\n");
/* Get bit-inverted image */
pixi = pixInvert(NULL, pixnht);
if (which == 1) pixWrite("/tmp/lept/invert.150.png", pixi, IFF_PNG);
pixDisplayWriteFormat(pixi, ws_flag, IFF_PNG);
/* The whitespace mask will break textlines where there
* is a large amount of white space below or above.
* We can prevent this by identifying regions of the
* inverted image that have large horizontal (bigger than
* the separation between columns) and significant
* vertical extent (bigger than the separation between
* textlines), and subtracting this from the whitespace mask. */
pixt1 = pixMorphCompSequence(pixi, "o80.60", 0);
pixt2 = pixSubtract(NULL, pixi, pixt1);
pixDisplayWriteFormat(pixt2, ws_flag, IFF_PNG);
pixDestroy(&pixt1);
/* Identify vertical whitespace by opening inverted image */
pixt3 = pixOpenBrick(NULL, pixt2, 5, 1); /* removes thin vertical lines */
pixvws = pixOpenBrick(NULL, pixt3, 1, 200); /* gets long vertical lines */
pixDisplayWriteFormat(pixvws, L_MAX(text_flag, ws_flag), IFF_PNG);
if (which == 1) pixWrite("/tmp/lept/vertws.150.png", pixvws, IFF_PNG);
pixDestroy(&pixt2);
pixDestroy(&pixt3);
/* Get proto (early processed) text line mask. */
/* First close the characters and words in the textlines */
pixtm1 = pixCloseSafeBrick(NULL, pixnht, 30, 1);
pixDisplayWriteFormat(pixtm1, text_flag, IFF_PNG);
if (which == 1) pixWrite("/tmp/lept/textmask1.150.png", pixtm1, IFF_PNG);
/* Next open back up the vertical whitespace corridors */
pixtm2 = pixSubtract(NULL, pixtm1, pixvws);
if (which == 1) pixWrite("/tmp/lept/textmask2.150.png", pixtm2, IFF_PNG);
/* Do a small opening to remove noise */
pixOpenBrick(pixtm2, pixtm2, 3, 3);
pixDisplayWriteFormat(pixtm2, text_flag, IFF_PNG);
if (which == 1) pixWrite("/tmp/lept/textmask3.150.png", pixtm2, IFF_PNG);
pixtm3 = pixExpandBinaryPower2(pixtm2, 2);
/* Join pixels vertically to make text block mask */
pixtb1 = pixMorphSequence(pixtm2, "c1.10 + o4.1", 0);
pixDisplayWriteFormat(pixtb1, block_flag, IFF_PNG);
if (which == 1) pixWrite("/tmp/lept/textblock1.150.png", pixtb1, IFF_PNG);
/* Solidify the textblock mask and remove noise:
* (1) For each c.c., close the blocks and dilate slightly
* to form a solid mask.
* (2) Small horizontal closing between components
* (3) Open the white space between columns, again
* (4) Remove small components */
pixt1 = pixMorphSequenceByComponent(pixtb1, "c30.30 + d3.3", 8, 0, 0, NULL);
pixCloseSafeBrick(pixt1, pixt1, 10, 1);
pixDisplayWriteFormat(pixt1, block_flag, IFF_PNG);
pixt2 = pixSubtract(NULL, pixt1, pixvws);
pixt3 = pixSelectBySize(pixt2, 25, 5, 8, L_SELECT_IF_BOTH,
L_SELECT_IF_GTE, NULL);
pixDisplayWriteFormat(pixt3, block_flag, IFF_PNG);
if (which == 1) pixWrite("/tmp/lept/textblock2.150.png", pixt3, IFF_PNG);
pixtb2 = pixExpandBinaryPower2(pixt3, 2);
pixDestroy(&pixt1);
pixDestroy(&pixt2);
pixDestroy(&pixt3);
/* Identify the outlines of each textblock */
ptaa = pixGetOuterBordersPtaa(pixtb2);
pixt1 = pixRenderRandomCmapPtaa(pixtb2, ptaa, 1, 8, 1);
cmap = pixGetColormap(pixt1);
pixcmapResetColor(cmap, 0, 130, 130, 130); /* set interior to gray */
if (which == 1) pixWrite("/tmp/lept/textblock3.300.png", pixt1, IFF_PNG);
pixDisplayWithTitle(pixt1, 480, 360, "textblock mask with outlines", DFLAG);
ptaaDestroy(&ptaa);
pixDestroy(&pixt1);
/* Fill line mask (as seed) into the original */
pixt1 = pixSeedfillBinary(NULL, pixtm3, pixs, 8);
pixOr(pixtm3, pixtm3, pixt1);
pixDestroy(&pixt1);
if (which == 1) pixWrite("/tmp/lept/textmask.300.png", pixtm3, IFF_PNG);
pixDisplayWithTitle(pixtm3, 480, 360, "textline mask 4", DFLAG);
/* Fill halftone mask (as seed) into the original */
pixt1 = pixSeedfillBinary(NULL, pixhm2, pixs, 8);
pixOr(pixhm2, pixhm2, pixt1);
pixDestroy(&pixt1);
if (which == 1) pixWrite("/tmp/lept/htmask.300.png", pixhm2, IFF_PNG);
pixDisplayWithTitle(pixhm2, 520, 390, "halftonemask 2", DFLAG);
/* Find objects that are neither text nor halftones */
pixt1 = pixSubtract(NULL, pixs, pixtm3); /* remove text pixels */
pixnon = pixSubtract(NULL, pixt1, pixhm2); /* remove halftone pixels */
if (which == 1) pixWrite("/tmp/lept/other.300.png", pixnon, IFF_PNG);
pixDisplayWithTitle(pixnon, 540, 420, "other stuff", DFLAG);
pixDestroy(&pixt1);
/* Write out b.b. for text line mask and halftone mask components */
boxatm = pixConnComp(pixtm3, NULL, 4);
boxahm = pixConnComp(pixhm2, NULL, 8);
if (which == 1) boxaWrite("/tmp/lept/textmask.boxa", boxatm);
if (which == 1) boxaWrite("/tmp/lept/htmask.boxa", boxahm);
pixa = pixaReadFiles("/tmp/lept/display", "file");
pixt1 = pixaDisplayTiledAndScaled(pixa, 8, 250, 4, 0, 25, 2);
snprintf(buf, sizeof(buf), "/tmp/lept/segout.%d.png", which);
pixWrite(buf, pixt1, IFF_PNG);
pixDestroy(&pixt1);
pixaDestroy(&pixa);
/* clean up to test with valgrind */
pixDestroy(&pixr);
pixDestroy(&pixhs);
pixDestroy(&pixm);
pixDestroy(&pixhm1);
pixDestroy(&pixhm2);
pixDestroy(&pixht);
pixDestroy(&pixnht);
pixDestroy(&pixi);
pixDestroy(&pixvws);
pixDestroy(&pixtm1);
pixDestroy(&pixtm2);
pixDestroy(&pixtm3);
pixDestroy(&pixtb1);
pixDestroy(&pixtb2);
pixDestroy(&pixnon);
boxaDestroy(&boxatm);
boxaDestroy(&boxahm);
return 0;
}
/*!
* \brief pixFindSkewSweepAndSearchScorePivot()
*
* \param[in] pixs 1 bpp
* \param[out] pangle angle required to deskew; in degrees
* \param[out] pconf confidence given by ratio of max/min score
* \param[out] pendscore [optional] max score; use NULL to ignore
* \param[in] redsweep sweep reduction factor = 1, 2, 4 or 8
* \param[in] redsearch binary search reduction factor = 1, 2, 4 or 8;
* and must not exceed redsweep
* \param[in] sweepcenter angle about which sweep is performed; in degrees
* \param[in] sweeprange half the full range, taken about sweepcenter;
* in degrees
* \param[in] sweepdelta angle increment of sweep; in degrees
* \param[in] minbsdelta min binary search increment angle; in degrees
* \param[in] pivot L_SHEAR_ABOUT_CORNER, L_SHEAR_ABOUT_CENTER
* \return 0 if OK, 1 on error or if angle measurment not valid
*
* <pre>
* Notes:
* (1) See notes in pixFindSkewSweepAndSearchScore().
* (2) This allows choice of shear pivoting from either the UL corner
* or the center. For small angles, the ability to discriminate
* angles is better with shearing from the UL corner. However,
* for large angles (say, greater than 20 degrees), it is better
* to shear about the center because a shear from the UL corner
* loses too much of the image.
* </pre>
*/
l_int32
pixFindSkewSweepAndSearchScorePivot(PIX *pixs,
l_float32 *pangle,
l_float32 *pconf,
l_float32 *pendscore,
l_int32 redsweep,
l_int32 redsearch,
l_float32 sweepcenter,
l_float32 sweeprange,
l_float32 sweepdelta,
l_float32 minbsdelta,
l_int32 pivot)
{
l_int32 ret, bzero, i, nangles, n, ratio, maxindex, minloc;
l_int32 width, height;
l_float32 deg2rad, theta, delta;
l_float32 sum, maxscore, maxangle;
l_float32 centerangle, leftcenterangle, rightcenterangle;
l_float32 lefttemp, righttemp;
l_float32 bsearchscore[5];
l_float32 minscore, minthresh;
l_float32 rangeleft;
NUMA *natheta, *nascore;
PIX *pixsw, *pixsch, *pixt1, *pixt2;
PROCNAME("pixFindSkewSweepAndSearchScorePivot");
if (pendscore) *pendscore = 0.0;
if (pangle) *pangle = 0.0;
if (pconf) *pconf = 0.0;
if (!pangle || !pconf)
return ERROR_INT("&angle and/or &conf not defined", procName, 1);
if (!pixs || pixGetDepth(pixs) != 1)
return ERROR_INT("pixs not defined or not 1 bpp", procName, 1);
if (redsweep != 1 && redsweep != 2 && redsweep != 4 && redsweep != 8)
return ERROR_INT("redsweep must be in {1,2,4,8}", procName, 1);
if (redsearch != 1 && redsearch != 2 && redsearch != 4 && redsearch != 8)
return ERROR_INT("redsearch must be in {1,2,4,8}", procName, 1);
if (redsearch > redsweep)
return ERROR_INT("redsearch must not exceed redsweep", procName, 1);
if (pivot != L_SHEAR_ABOUT_CORNER && pivot != L_SHEAR_ABOUT_CENTER)
return ERROR_INT("invalid pivot", procName, 1);
deg2rad = 3.1415926535 / 180.;
ret = 0;
/* Generate reduced image for binary search, if requested */
if (redsearch == 1)
pixsch = pixClone(pixs);
else if (redsearch == 2)
pixsch = pixReduceRankBinaryCascade(pixs, 1, 0, 0, 0);
else if (redsearch == 4)
pixsch = pixReduceRankBinaryCascade(pixs, 1, 1, 0, 0);
else /* redsearch == 8 */
pixsch = pixReduceRankBinaryCascade(pixs, 1, 1, 2, 0);
pixZero(pixsch, &bzero);
if (bzero) {
pixDestroy(&pixsch);
return 1;
}
/* Generate reduced image for sweep, if requested */
ratio = redsweep / redsearch;
if (ratio == 1) {
pixsw = pixClone(pixsch);
} else { /* ratio > 1 */
if (ratio == 2)
pixsw = pixReduceRankBinaryCascade(pixsch, 1, 0, 0, 0);
else if (ratio == 4)
pixsw = pixReduceRankBinaryCascade(pixsch, 1, 2, 0, 0);
else /* ratio == 8 */
pixsw = pixReduceRankBinaryCascade(pixsch, 1, 2, 2, 0);
}
pixt1 = pixCreateTemplate(pixsw);
if (ratio == 1)
pixt2 = pixClone(pixt1);
else
pixt2 = pixCreateTemplate(pixsch);
nangles = (l_int32)((2. * sweeprange) / sweepdelta + 1);
natheta = numaCreate(nangles);
nascore = numaCreate(nangles);
if (!pixsch || !pixsw) {
ret = ERROR_INT("pixsch and pixsw not both made", procName, 1);
goto cleanup;
}
if (!pixt1 || !pixt2) {
ret = ERROR_INT("pixt1 and pixt2 not both made", procName, 1);
goto cleanup;
}
if (!natheta || !nascore) {
ret = ERROR_INT("natheta and nascore not both made", procName, 1);
goto cleanup;
}
/* Do sweep */
rangeleft = sweepcenter - sweeprange;
for (i = 0; i < nangles; i++) {
theta = rangeleft + i * sweepdelta; /* degrees */
/* Shear pix and put the result in pixt1 */
if (pivot == L_SHEAR_ABOUT_CORNER)
pixVShearCorner(pixt1, pixsw, deg2rad * theta, L_BRING_IN_WHITE);
else
pixVShearCenter(pixt1, pixsw, deg2rad * theta, L_BRING_IN_WHITE);
/* Get score */
pixFindDifferentialSquareSum(pixt1, &sum);
#if DEBUG_PRINT_SCORES
L_INFO("sum(%7.2f) = %7.0f\n", procName, theta, sum);
#endif /* DEBUG_PRINT_SCORES */
/* Save the result in the output arrays */
numaAddNumber(nascore, sum);
numaAddNumber(natheta, theta);
}
/* Find the largest of the set (maxscore at maxangle) */
numaGetMax(nascore, &maxscore, &maxindex);
numaGetFValue(natheta, maxindex, &maxangle);
#if DEBUG_PRINT_SWEEP
L_INFO(" From sweep: angle = %7.3f, score = %7.3f\n", procName,
maxangle, maxscore);
#endif /* DEBUG_PRINT_SWEEP */
#if DEBUG_PLOT_SCORES
/* Plot the sweep result -- the scores versus rotation angle --
* using gnuplot with GPLOT_LINES (lines connecting data points). */
{GPLOT *gplot;
gplot = gplotCreate("sweep_output", GPLOT_PNG,
"Sweep. Variance of difference of ON pixels vs. angle",
"angle (deg)", "score");
gplotAddPlot(gplot, natheta, nascore, GPLOT_LINES, "plot1");
gplotAddPlot(gplot, natheta, nascore, GPLOT_POINTS, "plot2");
gplotMakeOutput(gplot);
gplotDestroy(&gplot);
}
#endif /* DEBUG_PLOT_SCORES */
/* Check if the max is at the end of the sweep. */
n = numaGetCount(natheta);
if (maxindex == 0 || maxindex == n - 1) {
L_WARNING("max found at sweep edge\n", procName);
goto cleanup;
}
/* Empty the numas for re-use */
numaEmpty(nascore);
numaEmpty(natheta);
/* Do binary search to find skew angle.
* First, set up initial three points. */
centerangle = maxangle;
if (pivot == L_SHEAR_ABOUT_CORNER) {
pixVShearCorner(pixt2, pixsch, deg2rad * centerangle, L_BRING_IN_WHITE);
pixFindDifferentialSquareSum(pixt2, &bsearchscore[2]);
pixVShearCorner(pixt2, pixsch, deg2rad * (centerangle - sweepdelta),
L_BRING_IN_WHITE);
pixFindDifferentialSquareSum(pixt2, &bsearchscore[0]);
pixVShearCorner(pixt2, pixsch, deg2rad * (centerangle + sweepdelta),
L_BRING_IN_WHITE);
pixFindDifferentialSquareSum(pixt2, &bsearchscore[4]);
} else {
pixVShearCenter(pixt2, pixsch, deg2rad * centerangle, L_BRING_IN_WHITE);
pixFindDifferentialSquareSum(pixt2, &bsearchscore[2]);
pixVShearCenter(pixt2, pixsch, deg2rad * (centerangle - sweepdelta),
L_BRING_IN_WHITE);
pixFindDifferentialSquareSum(pixt2, &bsearchscore[0]);
pixVShearCenter(pixt2, pixsch, deg2rad * (centerangle + sweepdelta),
L_BRING_IN_WHITE);
pixFindDifferentialSquareSum(pixt2, &bsearchscore[4]);
}
numaAddNumber(nascore, bsearchscore[2]);
numaAddNumber(natheta, centerangle);
numaAddNumber(nascore, bsearchscore[0]);
numaAddNumber(natheta, centerangle - sweepdelta);
numaAddNumber(nascore, bsearchscore[4]);
numaAddNumber(natheta, centerangle + sweepdelta);
/* Start the search */
delta = 0.5 * sweepdelta;
while (delta >= minbsdelta)
{
/* Get the left intermediate score */
leftcenterangle = centerangle - delta;
if (pivot == L_SHEAR_ABOUT_CORNER)
pixVShearCorner(pixt2, pixsch, deg2rad * leftcenterangle,
L_BRING_IN_WHITE);
else
pixVShearCenter(pixt2, pixsch, deg2rad * leftcenterangle,
L_BRING_IN_WHITE);
pixFindDifferentialSquareSum(pixt2, &bsearchscore[1]);
numaAddNumber(nascore, bsearchscore[1]);
numaAddNumber(natheta, leftcenterangle);
/* Get the right intermediate score */
rightcenterangle = centerangle + delta;
if (pivot == L_SHEAR_ABOUT_CORNER)
pixVShearCorner(pixt2, pixsch, deg2rad * rightcenterangle,
L_BRING_IN_WHITE);
else
pixVShearCenter(pixt2, pixsch, deg2rad * rightcenterangle,
L_BRING_IN_WHITE);
pixFindDifferentialSquareSum(pixt2, &bsearchscore[3]);
numaAddNumber(nascore, bsearchscore[3]);
numaAddNumber(natheta, rightcenterangle);
/* Find the maximum of the five scores and its location.
* Note that the maximum must be in the center
* three values, not in the end two. */
maxscore = bsearchscore[1];
maxindex = 1;
for (i = 2; i < 4; i++) {
if (bsearchscore[i] > maxscore) {
maxscore = bsearchscore[i];
maxindex = i;
}
}
/* Set up score array to interpolate for the next iteration */
lefttemp = bsearchscore[maxindex - 1];
righttemp = bsearchscore[maxindex + 1];
bsearchscore[2] = maxscore;
bsearchscore[0] = lefttemp;
bsearchscore[4] = righttemp;
/* Get new center angle and delta for next iteration */
centerangle = centerangle + delta * (maxindex - 2);
delta = 0.5 * delta;
}
*pangle = centerangle;
#if DEBUG_PRINT_SCORES
L_INFO(" Binary search score = %7.3f\n", procName, bsearchscore[2]);
#endif /* DEBUG_PRINT_SCORES */
if (pendscore) /* save if requested */
*pendscore = bsearchscore[2];
/* Return the ratio of Max score over Min score
* as a confidence value. Don't trust if the Min score
* is too small, which can happen if the image is all black
* with only a few white pixels interspersed. In that case,
* we get a contribution from the top and bottom edges when
* vertically sheared, but this contribution becomes zero when
* the shear angle is zero. For zero shear angle, the only
* contribution will be from the white pixels. We expect that
* the signal goes as the product of the (height * width^2),
* so we compute a (hopefully) normalized minimum threshold as
* a function of these dimensions. */
numaGetMin(nascore, &minscore, &minloc);
width = pixGetWidth(pixsch);
height = pixGetHeight(pixsch);
minthresh = MINSCORE_THRESHOLD_CONSTANT * width * width * height;
#if DEBUG_THRESHOLD
L_INFO(" minthresh = %10.2f, minscore = %10.2f\n", procName,
minthresh, minscore);
L_INFO(" maxscore = %10.2f\n", procName, maxscore);
#endif /* DEBUG_THRESHOLD */
if (minscore > minthresh)
*pconf = maxscore / minscore;
else
*pconf = 0.0;
/* Don't trust it if too close to the edge of the sweep
* range or if maxscore is small */
if ((centerangle > rangeleft + 2 * sweeprange - sweepdelta) ||
(centerangle < rangeleft + sweepdelta) ||
(maxscore < MIN_VALID_MAXSCORE))
*pconf = 0.0;
#if DEBUG_PRINT_BINARY
fprintf(stderr, "Binary search: angle = %7.3f, score ratio = %6.2f\n",
*pangle, *pconf);
fprintf(stderr, " max score = %8.0f\n", maxscore);
#endif /* DEBUG_PRINT_BINARY */
#if DEBUG_PLOT_SCORES
/* Plot the result -- the scores versus rotation angle --
* using gnuplot with GPLOT_POINTS. Because the data
* points are not ordered by theta (increasing or decreasing),
* using GPLOT_LINES would be confusing! */
{GPLOT *gplot;
gplot = gplotCreate("search_output", GPLOT_PNG,
"Binary search. Variance of difference of ON pixels vs. angle",
"angle (deg)", "score");
gplotAddPlot(gplot, natheta, nascore, GPLOT_POINTS, "plot1");
gplotMakeOutput(gplot);
gplotDestroy(&gplot);
}
#endif /* DEBUG_PLOT_SCORES */
cleanup:
pixDestroy(&pixsw);
pixDestroy(&pixsch);
pixDestroy(&pixt1);
pixDestroy(&pixt2);
numaDestroy(&nascore);
numaDestroy(&natheta);
return ret;
}
/*!
* \brief pixFindSkewSweep()
*
* \param[in] pixs 1 bpp
* \param[out] pangle angle required to deskew, in degrees
* \param[in] reduction factor = 1, 2, 4 or 8
* \param[in] sweeprange half the full range; assumed about 0; in degrees
* \param[in] sweepdelta angle increment of sweep; in degrees
* \return 0 if OK, 1 on error or if angle measurment not valid
*
* <pre>
* Notes:
* (1) This examines the 'score' for skew angles with equal intervals.
* (2) Caller must check the return value for validity of the result.
* </pre>
*/
l_int32
pixFindSkewSweep(PIX *pixs,
l_float32 *pangle,
l_int32 reduction,
l_float32 sweeprange,
l_float32 sweepdelta)
{
l_int32 ret, bzero, i, nangles;
l_float32 deg2rad, theta;
l_float32 sum, maxscore, maxangle;
NUMA *natheta, *nascore;
PIX *pix, *pixt;
PROCNAME("pixFindSkewSweep");
if (!pangle)
return ERROR_INT("&angle not defined", procName, 1);
*pangle = 0.0;
if (!pixs)
return ERROR_INT("pixs not defined", procName, 1);
if (pixGetDepth(pixs) != 1)
return ERROR_INT("pixs not 1 bpp", procName, 1);
if (reduction != 1 && reduction != 2 && reduction != 4 && reduction != 8)
return ERROR_INT("reduction must be in {1,2,4,8}", procName, 1);
deg2rad = 3.1415926535 / 180.;
ret = 0;
/* Generate reduced image, if requested */
if (reduction == 1)
pix = pixClone(pixs);
else if (reduction == 2)
pix = pixReduceRankBinaryCascade(pixs, 1, 0, 0, 0);
else if (reduction == 4)
pix = pixReduceRankBinaryCascade(pixs, 1, 1, 0, 0);
else /* reduction == 8 */
pix = pixReduceRankBinaryCascade(pixs, 1, 1, 2, 0);
pixZero(pix, &bzero);
if (bzero) {
pixDestroy(&pix);
return 1;
}
nangles = (l_int32)((2. * sweeprange) / sweepdelta + 1);
natheta = numaCreate(nangles);
nascore = numaCreate(nangles);
pixt = pixCreateTemplate(pix);
if (!pix || !pixt) {
ret = ERROR_INT("pix and pixt not both made", procName, 1);
goto cleanup;
}
if (!natheta || !nascore) {
ret = ERROR_INT("natheta and nascore not both made", procName, 1);
goto cleanup;
}
for (i = 0; i < nangles; i++) {
theta = -sweeprange + i * sweepdelta; /* degrees */
/* Shear pix about the UL corner and put the result in pixt */
pixVShearCorner(pixt, pix, deg2rad * theta, L_BRING_IN_WHITE);
/* Get score */
pixFindDifferentialSquareSum(pixt, &sum);
#if DEBUG_PRINT_SCORES
L_INFO("sum(%7.2f) = %7.0f\n", procName, theta, sum);
#endif /* DEBUG_PRINT_SCORES */
/* Save the result in the output arrays */
numaAddNumber(nascore, sum);
numaAddNumber(natheta, theta);
}
/* Find the location of the maximum (i.e., the skew angle)
* by fitting the largest data point and its two neighbors
* to a quadratic, using lagrangian interpolation. */
numaFitMax(nascore, &maxscore, natheta, &maxangle);
*pangle = maxangle;
#if DEBUG_PRINT_SWEEP
L_INFO(" From sweep: angle = %7.3f, score = %7.3f\n", procName,
maxangle, maxscore);
#endif /* DEBUG_PRINT_SWEEP */
#if DEBUG_PLOT_SCORES
/* Plot the result -- the scores versus rotation angle --
* using gnuplot with GPLOT_LINES (lines connecting data points).
* The GPLOT data structure is first created, with the
* appropriate data incorporated from the two input NUMAs,
* and then the function gplotMakeOutput() uses gnuplot to
* generate the output plot. This can be either a .png file
* or a .ps file, depending on whether you use GPLOT_PNG
* or GPLOT_PS. */
{GPLOT *gplot;
gplot = gplotCreate("sweep_output", GPLOT_PNG,
"Sweep. Variance of difference of ON pixels vs. angle",
"angle (deg)", "score");
gplotAddPlot(gplot, natheta, nascore, GPLOT_LINES, "plot1");
gplotAddPlot(gplot, natheta, nascore, GPLOT_POINTS, "plot2");
gplotMakeOutput(gplot);
gplotDestroy(&gplot);
}
#endif /* DEBUG_PLOT_SCORES */
cleanup:
pixDestroy(&pix);
pixDestroy(&pixt);
numaDestroy(&nascore);
numaDestroy(&natheta);
return ret;
}
