// Helper to remove an enclosing circle from an image.
// If there isn't one, then the image will most likely get badly mangled.
// The returned pix must be pixDestroyed after use. NULL may be returned
// if the image doesn't meet the trivial conditions that it uses to determine
// success.
static Pix* RemoveEnclosingCircle(Pix* pixs) {
Pix* pixsi = pixInvert(NULL, pixs);
Pix* pixc = pixCreateTemplate(pixs);
pixSetOrClearBorder(pixc, 1, 1, 1, 1, PIX_SET);
pixSeedfillBinary(pixc, pixc, pixsi, 4);
pixInvert(pixc, pixc);
pixDestroy(&pixsi);
Pix* pixt = pixAnd(NULL, pixs, pixc);
l_int32 max_count;
pixCountConnComp(pixt, 8, &max_count);
// The count has to go up before we start looking for the minimum.
l_int32 min_count = MAX_INT32;
Pix* pixout = NULL;
for (int i = 1; i < kMaxCircleErosions; i++) {
pixDestroy(&pixt);
pixErodeBrick(pixc, pixc, 3, 3);
pixt = pixAnd(NULL, pixs, pixc);
l_int32 count;
pixCountConnComp(pixt, 8, &count);
if (i == 1 || count > max_count) {
max_count = count;
min_count = count;
} else if (i > 1 && count < min_count) {
min_count = count;
pixDestroy(&pixout);
pixout = pixCopy(NULL, pixt); // Save the best.
} else if (count >= min_count) {
break; // We have passed by the best.
}
}
pixDestroy(&pixt);
pixDestroy(&pixc);
return pixout;
}
/*!
* pixaClipToPix()
*
* Input: pixas
* pixs
* Return: pixad, or null on error
*
* Notes:
* (1) This is intended for use in situations where pixas
* was originally generated from the input pixs.
* (2) Returns a pixad where each pix in pixas is ANDed
* with its associated region of the input pixs. This
* region is specified by the the box that is associated
* with the pix.
* (3) In a typical application of this function, pixas has
* a set of region masks, so this generates a pixa of
* the parts of pixs that correspond to each region
* mask component, along with the bounding box for
* the region.
*/
PIXA *
pixaClipToPix(PIXA *pixas,
PIX *pixs)
{
l_int32 i, n;
BOX *box;
PIX *pix, *pixc;
PIXA *pixad;
PROCNAME("pixaClipToPix");
if (!pixas)
return (PIXA *)ERROR_PTR("pixas not defined", procName, NULL);
if (!pixs)
return (PIXA *)ERROR_PTR("pixs not defined", procName, NULL);
n = pixaGetCount(pixas);
if ((pixad = pixaCreate(n)) == NULL)
return (PIXA *)ERROR_PTR("pixad not made", procName, NULL);
for (i = 0; i < n; i++) {
pix = pixaGetPix(pixas, i, L_CLONE);
box = pixaGetBox(pixas, i, L_COPY);
pixc = pixClipRectangle(pixs, box, NULL);
pixAnd(pixc, pixc, pix);
pixaAddPix(pixad, pixc, L_INSERT);
pixaAddBox(pixad, box, L_INSERT);
pixDestroy(&pix);
}
return pixad;
}
l_int32 main(int argc,
char **argv)
{
l_int32 irval, igval, ibval;
l_float32 rval, gval, bval, fract, fgfract;
L_BMF *bmf;
BOX *box;
BOXA *boxa;
FPIX *fpix;
PIX *pixs, *pix1, *pix2, *pix3, *pix4, *pix5, *pix6, *pix7;
PIX *pix8, *pix9, *pix10, *pix11, *pix12, *pix13, *pix14, *pix15;
PIXA *pixa;
L_REGPARAMS *rp;
if (regTestSetup(argc, argv, &rp))
return 1;
pixa = pixaCreate(0);
pixs = pixRead("breviar38.150.jpg");
/* pixs = pixRead("breviar32.150.jpg"); */
pixaAddPix(pixa, pixs, L_CLONE);
regTestWritePixAndCheck(rp, pixs, IFF_JFIF_JPEG); /* 0 */
pixDisplayWithTitle(pixs, 0, 0, "Input image", rp->display);
/* Extract the blue component, which is small in all the text
* regions, including in the highlight color region */
pix1 = pixGetRGBComponent(pixs, COLOR_BLUE);
pixaAddPix(pixa, pix1, L_CLONE);
regTestWritePixAndCheck(rp, pix1, IFF_JFIF_JPEG); /* 1 */
pixDisplayWithTitle(pix1, 200, 0, "Blue component", rp->display);
/* Do a background normalization, with the background set to
* approximately 200 */
pix2 = pixBackgroundNormSimple(pix1, NULL, NULL);
pixaAddPix(pixa, pix2, L_COPY);
regTestWritePixAndCheck(rp, pix2, IFF_JFIF_JPEG); /* 2 */
pixDisplayWithTitle(pix2, 400, 0, "BG normalized to 200", rp->display);
/* Do a linear transform on the gray pixels, with 50 going to
* black and 160 going to white. 50 is sufficiently low to
* make both the red and black print quite dark. Quantize
* to a few equally spaced gray levels. This is the image
* to which highlight color will be applied. */
pixGammaTRC(pix2, pix2, 1.0, 50, 160);
pix3 = pixThresholdOn8bpp(pix2, 7, 1);
pixaAddPix(pixa, pix3, L_CLONE);
regTestWritePixAndCheck(rp, pix3, IFF_JFIF_JPEG); /* 3 */
pixDisplayWithTitle(pix3, 600, 0, "Basic quantized with white bg",
rp->display);
/* Identify the regions of red text. First, make a mask
* consisting of all pixels such that (R-B)/B is larger
* than 2.0. This will have all the red, plus a lot of
* the dark pixels. */
fpix = pixComponentFunction(pixs, 1.0, 0.0, -1.0, 0.0, 0.0, 1.0);
pix4 = fpixThresholdToPix(fpix, 2.0);
pixInvert(pix4, pix4); /* red plus some dark text */
pixaAddPix(pixa, pix4, L_CLONE);
regTestWritePixAndCheck(rp, pix4, IFF_PNG); /* 4 */
pixDisplayWithTitle(pix4, 800, 0, "Red plus dark pixels", rp->display);
/* Make a mask consisting of all the red and background pixels */
pix5 = pixGetRGBComponent(pixs, COLOR_RED);
pix6 = pixThresholdToBinary(pix5, 128);
pixInvert(pix6, pix6); /* red plus background (white) */
/* Intersect the two masks to get a mask consisting of pixels
* that are almost certainly red. This is the seed. */
pix7 = pixAnd(NULL, pix4, pix6); /* red only (seed) */
pixaAddPix(pixa, pix7, L_COPY);
regTestWritePixAndCheck(rp, pix7, IFF_PNG); /* 5 */
pixDisplayWithTitle(pix7, 0, 600, "Seed for red color", rp->display);
/* Make the clipping mask by thresholding the image with
* the background cleaned to white. */
pix8 = pixThresholdToBinary(pix2, 230); /* mask */
pixaAddPix(pixa, pix8, L_CLONE);
regTestWritePixAndCheck(rp, pix8, IFF_PNG); /* 6 */
pixDisplayWithTitle(pix8, 200, 600, "Clipping mask for red components",
rp->display);
/* Fill into the mask from the seed */
pixSeedfillBinary(pix7, pix7, pix8, 8); /* filled: red plus touching */
regTestWritePixAndCheck(rp, pix7, IFF_PNG); /* 7 */
pixDisplayWithTitle(pix7, 400, 600, "Red component mask filled",
rp->display);
/* Remove long horizontal and vertical lines from the filled result */
pix9 = pixMorphSequence(pix7, "o40.1", 0);
pixSubtract(pix7, pix7, pix9); /* remove long horizontal lines */
pixDestroy(&pix9);
pix9 = pixMorphSequence(pix7, "o1.40", 0);
pixSubtract(pix7, pix7, pix9); /* remove long vertical lines */
/* Close the regions to be colored */
pix10 = pixMorphSequence(pix7, "c5.1", 0);
pixaAddPix(pixa, pix10, L_CLONE);
regTestWritePixAndCheck(rp, pix10, IFF_PNG); /* 8 */
pixDisplayWithTitle(pix10, 600, 600,
"Components defining regions allowing coloring",
rp->display);
/* Sanity check on amount to be colored. Only accept images
* with less than 10% of all the pixels with highlight color */
pixForegroundFraction(pix10, &fgfract);
if (fgfract >= 0.10) {
L_INFO("too much highlighting: fract = %6.3f; removing it\n",
rp->testname, fgfract);
pixClearAll(pix10);
pixSetPixel(pix10, 0, 0, 1);
}
/* Get the bounding boxes of the regions to be colored */
boxa = pixConnCompBB(pix10, 8);
/* Get a color to paint that is representative of the
* actual highlight color in the image. Scale each
* color component up from the average by an amount necessary
* to saturate the red. Then divide the green and
* blue components by 2.0. */
pixGetAverageMaskedRGB(pixs, pix7, 0, 0, 1, L_MEAN_ABSVAL,
&rval, &gval, &bval);
fract = 255.0 / rval;
irval = lept_roundftoi(fract * rval);
igval = lept_roundftoi(fract * gval / 2.0);
ibval = lept_roundftoi(fract * bval / 2.0);
fprintf(stderr, "(r,g,b) = (%d,%d,%d)\n", irval, igval, ibval);
/* Color the quantized gray version in the selected regions */
pix11 = pixColorGrayRegions(pix3, boxa, L_PAINT_DARK, 220, irval,
igval, ibval);
pixaAddPix(pixa, pix11, L_CLONE);
regTestWritePixAndCheck(rp, pix11, IFF_PNG); /* 9 */
pixDisplayWithTitle(pix11, 800, 600, "Final colored result", rp->display);
pixaAddPix(pixa, pixs, L_CLONE);
/* Test colorization on gray and cmapped gray */
pix12 = pixColorGrayRegions(pix2, boxa, L_PAINT_DARK, 220, 0, 255, 0);
pixaAddPix(pixa, pix12, L_CLONE);
regTestWritePixAndCheck(rp, pix12, IFF_PNG); /* 10 */
pixDisplayWithTitle(pix12, 900, 600, "Colorizing boxa gray", rp->display);
box = boxCreate(200, 200, 250, 350);
pix13 = pixCopy(NULL, pix2);
pixColorGray(pix13, box, L_PAINT_DARK, 220, 0, 0, 255);
pixaAddPix(pixa, pix13, L_CLONE);
regTestWritePixAndCheck(rp, pix13, IFF_PNG); /* 11 */
pixDisplayWithTitle(pix13, 1000, 600, "Colorizing box gray", rp->display);
pix14 = pixThresholdTo4bpp(pix2, 6, 1);
pix15 = pixColorGrayRegions(pix14, boxa, L_PAINT_DARK, 220, 0, 0, 255);
pixaAddPix(pixa, pix15, L_CLONE);
regTestWritePixAndCheck(rp, pix15, IFF_PNG); /* 12 */
pixDisplayWithTitle(pix15, 1100, 600, "Colorizing boxa cmap", rp->display);
pixColorGrayCmap(pix14, box, L_PAINT_DARK, 0, 255, 255);
pixaAddPix(pixa, pix14, L_CLONE);
regTestWritePixAndCheck(rp, pix14, IFF_PNG); /* 13 */
pixDisplayWithTitle(pix14, 1200, 600, "Colorizing box cmap", rp->display);
boxDestroy(&box);
/* Generate a pdf of the intermediate results */
lept_mkdir("lept");
L_INFO("Writing to /tmp/lept/colorize.pdf\n", rp->testname);
pixaConvertToPdf(pixa, 90, 1.0, 0, 0, "Colorizing highlighted text",
"/tmp/lept/colorize.pdf");
pixaDestroy(&pixa);
fpixDestroy(&fpix);
boxDestroy(&box);
boxaDestroy(&boxa);
pixDestroy(&pixs);
pixDestroy(&pix1);
pixDestroy(&pix2);
pixDestroy(&pix3);
pixDestroy(&pix4);
pixDestroy(&pix5);
pixDestroy(&pix6);
pixDestroy(&pix7);
pixDestroy(&pix8);
pixDestroy(&pix9);
pixDestroy(&pix10);
pixDestroy(&pix11);
pixDestroy(&pix12);
pixDestroy(&pix13);
pixDestroy(&pix14);
pixDestroy(&pix15);
/* Test the color detector */
pixa = pixaCreate(7);
bmf = bmfCreate("./fonts", 4);
pix1 = TestForRedColor(rp, "brev06.75.jpg", 1, bmf); /* 14 */
pixaAddPix(pixa, pix1, L_INSERT);
pix1 = TestForRedColor(rp, "brev10.75.jpg", 0, bmf); /* 15 */
pixaAddPix(pixa, pix1, L_INSERT);
pix1 = TestForRedColor(rp, "brev14.75.jpg", 1, bmf); /* 16 */
pixaAddPix(pixa, pix1, L_INSERT);
pix1 = TestForRedColor(rp, "brev20.75.jpg", 1, bmf); /* 17 */
pixaAddPix(pixa, pix1, L_INSERT);
pix1 = TestForRedColor(rp, "brev36.75.jpg", 0, bmf); /* 18 */
pixaAddPix(pixa, pix1, L_INSERT);
pix1 = TestForRedColor(rp, "brev53.75.jpg", 1, bmf); /* 19 */
pixaAddPix(pixa, pix1, L_INSERT);
pix1 = TestForRedColor(rp, "brev56.75.jpg", 1, bmf); /* 20 */
pixaAddPix(pixa, pix1, L_INSERT);
/* Generate a pdf of the color detector results */
L_INFO("Writing to /tmp/lept/colordetect.pdf\n", rp->testname);
pixaConvertToPdf(pixa, 45, 1.0, 0, 0, "Color detection",
"/tmp/lept/colordetect.pdf");
pixaDestroy(&pixa);
bmfDestroy(&bmf);
return regTestCleanup(rp);
}
* that might be used *
* -------------------------------------------------------------------- */
#if 0
pixd = pixCreateTemplate(pixs);
pixd = pixDilate(NULL, pixs, sel);
pixd = pixErode(NULL, pixs, sel);
pixd = pixOpen(NULL, pixs, sel);
pixd = pixClose(NULL, pixs, sel);
pixDilate(pixd, pixs, sel);
pixErode(pixd, pixs, sel);
pixOpen(pixd, pixs, sel);
pixClose(pixd, pixs, sel);
pixAnd(pixd, pixd, pixs);
pixOr(pixd, pixd, pixs);
pixXor(pixd, pixd, pixs);
pixSubtract(pixd, pixd, pixs);
pixInvert(pixd, pixs);
pixd = pixAnd(NULL, pixd, pixs);
pixd = pixOr(NULL, pixd, pixs);
pixd = pixXor(NULL, pixd, pixs);
pixd = pixSubtract(NULL, pixd, pixs);
pixd = pixInvert(NULL, pixs);
pixInvert(pixs, pixs);
#endif /* 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 *pixm1; /* image of closed textlines, 150 ppi */
PIX *pixm2; /* image of refined text line mask, 150 ppi */
PIX *pixm3; /* image of refined text line mask, 300 ppi */
PIX *pixb1; /* image of text block mask, 150 ppi */
PIX *pixb2; /* image of text block mask, 300 ppi */
PIX *pixnon; /* image of non-text or halftone, 150 ppi */
PIX *pix1, *pix2, *pix3, *pix4;
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);
pixa = pixaCreate(0);
lept_mkdir("lept/livre");
/* Reduce to 150 ppi */
pix1 = pixScaleToGray2(pixs);
if (ws_flag || ht_flag || block_flag) pixaAddPix(pixa, pix1, L_COPY);
if (which == 1)
pixWrite("/tmp/lept/livre/orig.gray.150.png", pix1, IFF_PNG);
pixDestroy(&pix1);
pixr = pixReduceRankBinaryCascade(pixs, 1, 0, 0, 0);
/* Get seed for halftone parts */
pix1 = pixReduceRankBinaryCascade(pixr, 4, 4, 3, 0);
pix2 = pixOpenBrick(NULL, pix1, 5, 5);
pixhs = pixExpandBinaryPower2(pix2, 8);
if (ht_flag) pixaAddPix(pixa, pixhs, L_COPY);
if (which == 1)
pixWrite("/tmp/lept/livre/htseed.150.png", pixhs, IFF_PNG);
pixDestroy(&pix1);
pixDestroy(&pix2);
/* Get mask for connected regions */
pixm = pixCloseSafeBrick(NULL, pixr, 4, 4);
if (ht_flag) pixaAddPix(pixa, pixm, L_COPY);
if (which == 1)
pixWrite("/tmp/lept/livre/ccmask.150.png", pixm, IFF_PNG);
/* Fill seed into mask to get halftone mask */
pixhm1 = pixSeedfillBinary(NULL, pixhs, pixm, 4);
if (ht_flag) pixaAddPix(pixa, pixhm1, L_COPY);
if (which == 1) pixWrite("/tmp/lept/livre/htmask.150.png", pixhm1, IFF_PNG);
pixhm2 = pixExpandBinaryPower2(pixhm1, 2);
/* Extract halftone stuff */
pixht = pixAnd(NULL, pixhm1, pixr);
if (which == 1) pixWrite("/tmp/lept/livre/ht.150.png", pixht, IFF_PNG);
/* Extract non-halftone stuff */
pixnht = pixXor(NULL, pixht, pixr);
if (text_flag) pixaAddPix(pixa, pixnht, L_COPY);
if (which == 1) pixWrite("/tmp/lept/livre/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 (ws_flag) pixaAddPix(pixa, pixi, L_COPY);
if (which == 1) pixWrite("/tmp/lept/livre/invert.150.png", pixi, 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. */
pix1 = pixMorphCompSequence(pixi, "o80.60", 0);
pix2 = pixSubtract(NULL, pixi, pix1);
if (ws_flag) pixaAddPix(pixa, pix2, L_COPY);
pixDestroy(&pix1);
/* Identify vertical whitespace by opening inverted image */
pix3 = pixOpenBrick(NULL, pix2, 5, 1); /* removes thin vertical lines */
pixvws = pixOpenBrick(NULL, pix3, 1, 200); /* gets long vertical lines */
if (text_flag || ws_flag) pixaAddPix(pixa, pixvws, L_COPY);
if (which == 1) pixWrite("/tmp/lept/livre/vertws.150.png", pixvws, IFF_PNG);
pixDestroy(&pix2);
pixDestroy(&pix3);
/* Get proto (early processed) text line mask. */
/* First close the characters and words in the textlines */
pixm1 = pixCloseSafeBrick(NULL, pixnht, 30, 1);
if (text_flag) pixaAddPix(pixa, pixm1, L_COPY);
if (which == 1)
pixWrite("/tmp/lept/livre/textmask1.150.png", pixm1, IFF_PNG);
/* Next open back up the vertical whitespace corridors */
pixm2 = pixSubtract(NULL, pixm1, pixvws);
if (which == 1)
pixWrite("/tmp/lept/livre/textmask2.150.png", pixm2, IFF_PNG);
/* Do a small opening to remove noise */
pixOpenBrick(pixm2, pixm2, 3, 3);
if (text_flag) pixaAddPix(pixa, pixm2, L_COPY);
if (which == 1)
pixWrite("/tmp/lept/livre/textmask3.150.png", pixm2, IFF_PNG);
pixm3 = pixExpandBinaryPower2(pixm2, 2);
/* Join pixels vertically to make text block mask */
pixb1 = pixMorphSequence(pixm2, "c1.10 + o4.1", 0);
if (block_flag) pixaAddPix(pixa, pixb1, L_COPY);
if (which == 1)
pixWrite("/tmp/lept/livre/textblock1.150.png", pixb1, 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 */
pix1 = pixMorphSequenceByComponent(pixb1, "c30.30 + d3.3", 8, 0, 0, NULL);
pixCloseSafeBrick(pix1, pix1, 10, 1);
if (block_flag) pixaAddPix(pixa, pix1, L_COPY);
pix2 = pixSubtract(NULL, pix1, pixvws);
pix3 = pixSelectBySize(pix2, 25, 5, 8, L_SELECT_IF_BOTH,
L_SELECT_IF_GTE, NULL);
if (block_flag) pixaAddPix(pixa, pix3, L_COPY);
if (which == 1)
pixWrite("/tmp/lept/livre/textblock2.150.png", pix3, IFF_PNG);
pixb2 = pixExpandBinaryPower2(pix3, 2);
pixDestroy(&pix1);
pixDestroy(&pix2);
pixDestroy(&pix3);
/* Identify the outlines of each textblock */
ptaa = pixGetOuterBordersPtaa(pixb2);
pix1 = pixRenderRandomCmapPtaa(pixb2, ptaa, 1, 8, 1);
cmap = pixGetColormap(pix1);
pixcmapResetColor(cmap, 0, 130, 130, 130); /* set interior to gray */
if (which == 1)
pixWrite("/tmp/lept/livre/textblock3.300.png", pix1, IFF_PNG);
pixDisplayWithTitle(pix1, 480, 360, "textblock mask with outlines", DFLAG);
ptaaDestroy(&ptaa);
pixDestroy(&pix1);
/* Fill line mask (as seed) into the original */
pix1 = pixSeedfillBinary(NULL, pixm3, pixs, 8);
pixOr(pixm3, pixm3, pix1);
pixDestroy(&pix1);
if (which == 1)
pixWrite("/tmp/lept/livre/textmask.300.png", pixm3, IFF_PNG);
pixDisplayWithTitle(pixm3, 480, 360, "textline mask 4", DFLAG);
/* Fill halftone mask (as seed) into the original */
pix1 = pixSeedfillBinary(NULL, pixhm2, pixs, 8);
pixOr(pixhm2, pixhm2, pix1);
pixDestroy(&pix1);
if (which == 1)
pixWrite("/tmp/lept/livre/htmask.300.png", pixhm2, IFF_PNG);
pixDisplayWithTitle(pixhm2, 520, 390, "halftonemask 2", DFLAG);
/* Find objects that are neither text nor halftones */
pix1 = pixSubtract(NULL, pixs, pixm3); /* remove text pixels */
pixnon = pixSubtract(NULL, pix1, pixhm2); /* remove halftone pixels */
pixDestroy(&pix1);
if (which == 1)
pixWrite("/tmp/lept/livre/other.300.png", pixnon, IFF_PNG);
pixDisplayWithTitle(pixnon, 540, 420, "other stuff", DFLAG);
/* Write out b.b. for text line mask and halftone mask components */
boxatm = pixConnComp(pixm3, NULL, 4);
boxahm = pixConnComp(pixhm2, NULL, 8);
if (which == 1) {
boxaWrite("/tmp/lept/livre/textmask.boxa", boxatm);
boxaWrite("/tmp/lept/livre/htmask.boxa", boxahm);
}
pix1 = pixaDisplayTiledAndScaled(pixa, 8, 250, 4, 0, 25, 2);
pixDisplay(pix1, 0, 375 * (which - 1));
snprintf(buf, sizeof(buf), "/tmp/lept/livre/segout.%d.png", which);
pixWrite(buf, pix1, IFF_PNG);
pixDestroy(&pix1);
pixaDestroy(&pixa);
/* clean up to test with valgrind */
pixDestroy(&pixr);
pixDestroy(&pixhs);
pixDestroy(&pixm);
pixDestroy(&pixhm1);
pixDestroy(&pixhm2);
pixDestroy(&pixht);
pixDestroy(&pixi);
pixDestroy(&pixnht);
pixDestroy(&pixvws);
pixDestroy(&pixm1);
pixDestroy(&pixm2);
pixDestroy(&pixm3);
pixDestroy(&pixb1);
pixDestroy(&pixb2);
pixDestroy(&pixnon);
boxaDestroy(&boxatm);
boxaDestroy(&boxahm);
return 0;
}
/*!
* Note: this method is generally inferior to pixHasColorRegions(); it
* is retained as a reference only
*
* \brief pixFindColorRegionsLight()
*
* \param[in] pixs 32 bpp rgb
* \param[in] pixm [optional] 1 bpp mask image
* \param[in] factor subsample factor; integer >= 1
* \param[in] darkthresh threshold to eliminate dark pixels (e.g., text)
* from consideration; typ. 70; -1 for default.
* \param[in] lightthresh threshold for minimum gray value at 95% rank
* near white; typ. 220; -1 for default
* \param[in] mindiff minimum difference from 95% rank value, used
* to count darker pixels; typ. 50; -1 for default
* \param[in] colordiff minimum difference in (max - min) component to
* qualify as a color pixel; typ. 40; -1 for default
* \param[out] pcolorfract fraction of 'color' pixels found
* \param[out] pcolormask1 [optional] mask over background color, if any
* \param[out] pcolormask2 [optional] filtered mask over background color
* \param[out] pixadb [optional] debug intermediate results
* \return 0 if OK, 1 on error
*
* <pre>
* Notes:
* (1) This function tries to determine if there is a significant
* color or darker region on a scanned page image where part
* of the image is very close to "white". It will also allow
* extraction of small regions of lightly colored pixels.
* If the background is darker (and reddish), use instead
* pixHasColorRegions2().
* (2) If %pixm exists, only pixels under fg are considered. Typically,
* the inverse of %pixm would have fg pixels over a photograph.
* (3) There are four thresholds.
* * %darkthresh: ignore pixels darker than this (typ. fg text).
* We make a 1 bpp mask of these pixels, and then dilate it to
* remove all vestiges of fg from their vicinity.
* * %lightthresh: let val95 be the pixel value for which 95%
* of the non-masked pixels have a lower value (darker) of
* their min component. Then if val95 is darker than
* %lightthresh, the image is not considered to have a
* light bg, and this returns 0.0 for %colorfract.
* * %mindiff: we are interested in the fraction of pixels that
* have two conditions. The first is that their min component
* is at least %mindiff darker than val95.
* * %colordiff: the second condition is that the max-min diff
* of the pixel components exceeds %colordiff.
* (4) This returns in %pcolorfract the fraction of pixels that have
* both a min component that is at least %mindiff below that at the
* 95% rank value (where 100% rank is the lightest value), and
* a max-min diff that is at least %colordiff. Without the
* %colordiff constraint, gray pixels of intermediate value
* could get flagged by this function.
* (5) No masks are returned unless light color pixels are found.
* If colorfract > 0.0 and %pcolormask1 is defined, this returns
* a 1 bpp mask with fg pixels over the color background.
* This mask may have some holes in it.
* (6) If colorfract > 0.0 and %pcolormask2 is defined, this returns
* a filtered version of colormask1. The two changes are
* (a) small holes have been filled
* (b) components near the border have been removed.
* The latter insures that dark pixels near the edge of the
* image are not included.
* (7) To generate a boxa of rectangular regions from the overlap
* of components in the filtered mask:
* boxa1 = pixConnCompBB(colormask2, 8);
* boxa2 = boxaCombineOverlaps(boxa1);
* This is done here in debug mode.
* </pre>
*/
static l_int32
pixFindColorRegionsLight(PIX *pixs,
PIX *pixm,
l_int32 factor,
l_int32 darkthresh,
l_int32 lightthresh,
l_int32 mindiff,
l_int32 colordiff,
l_float32 *pcolorfract,
PIX **pcolormask1,
PIX **pcolormask2,
PIXA *pixadb)
{
l_int32 lightbg, w, h, count;
l_float32 ratio, val95, rank;
BOXA *boxa1, *boxa2;
NUMA *nah;
PIX *pix1, *pix2, *pix3, *pix4, *pix5, *pixm1, *pixm2, *pixm3;
PROCNAME("pixFindColorRegionsLight");
if (pcolormask1) *pcolormask1 = NULL;
if (pcolormask2) *pcolormask2 = NULL;
if (!pcolorfract)
return ERROR_INT("&colorfract not defined", procName, 1);
*pcolorfract = 0.0;
if (!pixs || pixGetDepth(pixs) != 32)
return ERROR_INT("pixs not defined or not 32 bpp", procName, 1);
if (factor < 1) factor = 1;
if (darkthresh < 0) darkthresh = 70; /* defaults */
if (lightthresh < 0) lightthresh = 220;
if (mindiff < 0) mindiff = 50;
if (colordiff < 0) colordiff = 40;
/* Check if pixm covers most of the image. If so, just return. */
pixGetDimensions(pixs, &w, &h, NULL);
if (pixm) {
pixCountPixels(pixm, &count, NULL);
ratio = (l_float32)count / ((l_float32)(w) * h);
if (ratio > 0.7) {
if (pixadb) L_INFO("pixm has big fg: %f5.2\n", procName, ratio);
return 0;
}
}
/* Make a mask pixm1 over the dark pixels in the image:
* convert to gray using the average of the components;
* threshold using %darkthresh; do a small dilation;
* combine with pixm. */
pix1 = pixConvertRGBToGray(pixs, 0.33, 0.34, 0.33);
if (pixadb) pixaAddPix(pixadb, pixs, L_COPY);
if (pixadb) pixaAddPix(pixadb, pix1, L_COPY);
pixm1 = pixThresholdToBinary(pix1, darkthresh);
pixDilateBrick(pixm1, pixm1, 7, 7);
if (pixadb) pixaAddPix(pixadb, pixm1, L_COPY);
if (pixm) {
pixOr(pixm1, pixm1, pixm);
if (pixadb) pixaAddPix(pixadb, pixm1, L_COPY);
}
pixDestroy(&pix1);
/* Convert to gray using the minimum component value and
* find the gray value at rank 0.95, that represents the light
* pixels in the image. If it is too dark, quit. */
pix1 = pixConvertRGBToGrayMinMax(pixs, L_SELECT_MIN);
pix2 = pixInvert(NULL, pixm1); /* pixels that are not dark */
pixGetRankValueMasked(pix1, pix2, 0, 0, factor, 0.95, &val95, &nah);
pixDestroy(&pix2);
if (pixadb) {
L_INFO("val at 0.95 rank = %5.1f\n", procName, val95);
gplotSimple1(nah, GPLOT_PNG, "/tmp/lept/histo1", "gray histo");
pix3 = pixRead("/tmp/lept/histo1.png");
pix4 = pixExpandReplicate(pix3, 2);
pixaAddPix(pixadb, pix4, L_INSERT);
pixDestroy(&pix3);
}
lightbg = (l_int32)val95 >= lightthresh;
numaDestroy(&nah);
if (!lightbg) {
pixDestroy(&pix1);
pixDestroy(&pixm1);
return 0;
}
/* Make mask pixm2 over pixels that are darker than val95 - mindiff. */
pixm2 = pixThresholdToBinary(pix1, val95 - mindiff);
if (pixadb) pixaAddPix(pixadb, pixm2, L_COPY);
pixDestroy(&pix1);
/* Make a mask pixm3 over pixels that have some color saturation,
* with a (max - min) component difference >= %colordiff,
* and combine using AND with pixm2. */
pix2 = pixConvertRGBToGrayMinMax(pixs, L_CHOOSE_MAXDIFF);
pixm3 = pixThresholdToBinary(pix2, colordiff);
pixDestroy(&pix2);
pixInvert(pixm3, pixm3); /* need pixels above threshold */
if (pixadb) pixaAddPix(pixadb, pixm3, L_COPY);
pixAnd(pixm2, pixm2, pixm3);
if (pixadb) pixaAddPix(pixadb, pixm2, L_COPY);
pixDestroy(&pixm3);
/* Subtract the dark pixels represented by pixm1.
* pixm2 now holds all the color pixels of interest */
pixSubtract(pixm2, pixm2, pixm1);
pixDestroy(&pixm1);
if (pixadb) pixaAddPix(pixadb, pixm2, L_COPY);
/* But we're not quite finished. Remove pixels from any component
* that is touching the image border. False color pixels can
* sometimes be found there if the image is much darker near
* the border, due to oxidation or reduced illumination. */
pixm3 = pixRemoveBorderConnComps(pixm2, 8);
pixDestroy(&pixm2);
if (pixadb) pixaAddPix(pixadb, pixm3, L_COPY);
/* Get the fraction of light color pixels */
pixCountPixels(pixm3, &count, NULL);
*pcolorfract = (l_float32)count / (w * h);
if (pixadb) {
if (count == 0)
L_INFO("no light color pixels found\n", procName);
else
L_INFO("fraction of light color pixels = %5.3f\n", procName,
*pcolorfract);
}
/* Debug: extract the color pixels from pixs */
if (pixadb && count > 0) {
/* Use pixm3 to extract the color pixels */
pix3 = pixCreateTemplate(pixs);
pixSetAll(pix3);
pixCombineMasked(pix3, pixs, pixm3);
pixaAddPix(pixadb, pix3, L_INSERT);
/* Use additional filtering to extract the color pixels */
pix3 = pixCloseSafeBrick(NULL, pixm3, 15, 15);
pixaAddPix(pixadb, pix3, L_INSERT);
pix5 = pixCreateTemplate(pixs);
pixSetAll(pix5);
pixCombineMasked(pix5, pixs, pix3);
pixaAddPix(pixadb, pix5, L_INSERT);
/* Get the combined bounding boxes of the mask components
* in pix3, and extract those pixels from pixs. */
boxa1 = pixConnCompBB(pix3, 8);
boxa2 = boxaCombineOverlaps(boxa1, NULL);
pix4 = pixCreateTemplate(pix3);
pixMaskBoxa(pix4, pix4, boxa2, L_SET_PIXELS);
pixaAddPix(pixadb, pix4, L_INSERT);
pix5 = pixCreateTemplate(pixs);
pixSetAll(pix5);
pixCombineMasked(pix5, pixs, pix4);
pixaAddPix(pixadb, pix5, L_INSERT);
boxaDestroy(&boxa1);
boxaDestroy(&boxa2);
pixaAddPix(pixadb, pixs, L_COPY);
}
/* Optional colormask returns */
if (pcolormask2 && count > 0)
*pcolormask2 = pixCloseSafeBrick(NULL, pixm3, 15, 15);
if (pcolormask1 && count > 0)
*pcolormask1 = pixm3;
else
pixDestroy(&pixm3);
return 0;
}
/* ----------------------------------------------------- */
void ProcessDigits(l_int32 index)
{
char rootname[8] = "digit5";
char buf[64];
l_int32 i, nc, ns, same;
NUMA *na1;
PIX *pix1, *pix2, *pix3, *pix4, *pix5, *pix6;
PIXA *pixa1, *pixa2, *pixa3;
/* Read the unfiltered, unscaled pixa of twenty-five 5s */
snprintf(buf, sizeof(buf), "digits/%s.orig-25.pa", rootname);
pixa1 = pixaRead(buf);
/* Number and show the input images */
snprintf(buf, sizeof(buf), "/tmp/lept/digit/%s.orig-num", rootname);
PixaDisplayNumbered(pixa1, buf);
/* Remove some of them */
na1 = numaCreateFromString(removeset);
pixaRemoveSelected(pixa1, na1);
numaDestroy(&na1);
snprintf(buf, sizeof(buf), "/tmp/lept/digit/%s.filt.pa", rootname);
pixaWrite(buf, pixa1);
/* Number and show the filtered images */
snprintf(buf, sizeof(buf), "/tmp/lept/digit/%s.filt-num", rootname);
PixaDisplayNumbered(pixa1, buf);
/* Extract the largest c.c., clip to the foreground,
* and scale the result to a fixed size. */
nc = pixaGetCount(pixa1);
pixa2 = pixaCreate(nc);
for (i = 0; i < nc; i++) {
pix1 = pixaGetPix(pixa1, i, L_CLONE);
/* A threshold of 140 gives reasonable results */
pix2 = pixThresholdToBinary(pix1, 140);
/* Join nearly touching pieces */
pix3 = pixCloseSafeBrick(NULL, pix2, 5, 5);
/* Take the largest (by area) connected component */
pix4 = pixFilterComponentBySize(pix3, 0, L_SELECT_BY_AREA, 8, NULL);
/* Extract the original 1 bpp pixels that have been
* covered by the closing operation */
pixAnd(pix4, pix4, pix2);
/* Grab the result as an image with no surrounding whitespace */
pixClipToForeground(pix4, &pix5, NULL);
/* Rescale the result to the canonical size */
pix6 = pixScaleToSize(pix5, 20, 30);
pixaAddPix(pixa2, pix6, L_INSERT);
pixDestroy(&pix1);
pixDestroy(&pix2);
pixDestroy(&pix3);
pixDestroy(&pix4);
pixDestroy(&pix5);
}
/* Add the index (a "5") in the text field of each pix; save pixa2 */
snprintf(buf, sizeof(buf), "%d", index);
for (i = 0; i < nc; i++) {
pix1 = pixaGetPix(pixa2, i, L_CLONE);
pixSetText(pix1, buf);
pixDestroy(&pix1);
}
snprintf(buf, sizeof(buf), "/tmp/lept/digit/%s.comp.pa", rootname);
pixaWrite(buf, pixa2);
/* Number and show the resulting binary templates */
snprintf(buf, sizeof(buf), "/tmp/lept/digit/%s.comp-num", rootname);
PixaDisplayNumbered(pixa2, buf);
/* Save the binary templates as a packed tiling (tiff g4).
* This is the most efficient way to represent the templates. */
pix1 = pixaDisplayOnLattice(pixa2, 20, 30, NULL, NULL);
pixDisplay(pix1, 1000, 500);
snprintf(buf, sizeof(buf), "/tmp/lept/digit/%s.comp.tif", rootname);
pixWrite(buf, pix1, IFF_TIFF_G4);
/* The number of templates is in the pix text string; check it. */
pix2 = pixRead(buf);
if (sscanf(pixGetText(pix2), "n = %d", &ns) != 1)
fprintf(stderr, "Failed to read the number of templates!\n");
if (ns != nc)
fprintf(stderr, "(stored = %d) != (actual number = %d)\n", ns, nc);
/* Reconstruct the pixa of templates from the tiled compressed
* image, and verify that the resulting pixa is the same. */
pixa3 = pixaMakeFromTiledPix(pix1, 20, 30, 0, 0, NULL);
pixaEqual(pixa2, pixa3, 0, NULL, &same);
if (!same)
fprintf(stderr, "Pixa are not the same!\n");
pixDestroy(&pix1);
pixDestroy(&pix2);
pixaDestroy(&pixa1);
pixaDestroy(&pixa2);
pixaDestroy(&pixa3);
}
/*!
* pixMirrorDetectDwa()
*
* Input: pixs (1 bpp, deskewed, English text)
* &conf (<return> confidence that text is not LR mirror reversed)
* mincount (min number of left + right; use 0 for default)
* debug (1 for debug output; 0 otherwise)
* Return: 0 if OK, 1 on error
*
* Notes:
* (1) We assume the text is horizontally oriented, with
* ascenders going up.
* (2) See notes in pixMirrorDetect().
*/
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 *pixt0, *pixt1, *pixt2, *pixt3;
PROCNAME("pixMirrorDetectDwa");
if (!pconf)
return ERROR_INT("&conf not defined", procName, 1);
*pconf = 0.0;
if (!pixs)
return ERROR_INT("pixs not defined", procName, 1);
if (mincount == 0)
mincount = DEFAULT_MIN_MIRROR_FLIP_COUNT;
/* Fill x-height characters but not space between them, sort of. */
pixt3 = pixMorphSequenceDwa(pixs, "d1.30", 0);
pixXor(pixt3, pixt3, pixs);
pixt0 = pixMorphSequenceDwa(pixs, "c15.1", 0);
pixXor(pixt0, pixt0, pixs);
pixAnd(pixt0, pixt0, pixt3);
pixOr(pixt3, pixt0, pixs);
pixDestroy(&pixt0);
pixt0 = pixAddBorderGeneral(pixt3, ADDED_BORDER, ADDED_BORDER,
ADDED_BORDER, ADDED_BORDER, 0);
pixDestroy(&pixt3);
/* Filter the right-facing characters. */
pixt1 = pixFlipFHMTGen(NULL, pixt0, flipsel1);
pixt3 = pixReduceRankBinaryCascade(pixt1, 1, 1, 0, 0);
pixCountPixels(pixt3, &count1, NULL);
pixDestroy(&pixt1);
pixDestroy(&pixt3);
/* Filter the left-facing characters. */
pixt2 = pixFlipFHMTGen(NULL, pixt0, flipsel2);
pixt3 = pixReduceRankBinaryCascade(pixt2, 1, 1, 0, 0);
pixCountPixels(pixt3, &count2, NULL);
pixDestroy(&pixt2);
pixDestroy(&pixt3);
pixDestroy(&pixt0);
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;
}
// Finds image regions within the source pix (page image) and returns
// the image regions as a Boxa, Pixa pair, analgous to pixConnComp.
// The returned boxa, pixa may be NULL, meaning no images found.
// If not NULL, they must be destroyed by the caller.
void ImageFinder::FindImages(Pix* pix, Boxa** boxa, Pixa** pixa) {
*boxa = NULL;
*pixa = NULL;
#ifdef HAVE_LIBLEPT
if (pixGetWidth(pix) < kMinImageFindSize ||
pixGetHeight(pix) < kMinImageFindSize)
return; // Not worth looking at small images.
// Reduce by factor 2.
Pix *pixr = pixReduceRankBinaryCascade(pix, 1, 0, 0, 0);
pixDisplayWrite(pixr, textord_tabfind_show_images);
// Get the halftone mask directly from Leptonica.
Pix *pixht2 = pixGenHalftoneMask(pixr, NULL, NULL,
textord_tabfind_show_images);
pixDestroy(&pixr);
if (pixht2 == NULL)
return;
// Expand back up again.
Pix *pixht = pixExpandReplicate(pixht2, 2);
pixDisplayWrite(pixht, textord_tabfind_show_images);
pixDestroy(&pixht2);
// Fill to capture pixels near the mask edges that were missed
Pix *pixt = pixSeedfillBinary(NULL, pixht, pix, 8);
pixOr(pixht, pixht, pixt);
pixDestroy(&pixt);
// Eliminate lines and bars that may be joined to images.
Pix* pixfinemask = pixReduceRankBinaryCascade(pixht, 1, 1, 3, 3);
pixDilateBrick(pixfinemask, pixfinemask, 5, 5);
pixDisplayWrite(pixfinemask, textord_tabfind_show_images);
Pix* pixreduced = pixReduceRankBinaryCascade(pixht, 1, 1, 1, 1);
Pix* pixreduced2 = pixReduceRankBinaryCascade(pixreduced, 3, 3, 3, 0);
pixDestroy(&pixreduced);
pixDilateBrick(pixreduced2, pixreduced2, 5, 5);
Pix* pixcoarsemask = pixExpandReplicate(pixreduced2, 8);
pixDestroy(&pixreduced2);
pixDisplayWrite(pixcoarsemask, textord_tabfind_show_images);
// Combine the coarse and fine image masks.
pixAnd(pixcoarsemask, pixcoarsemask, pixfinemask);
pixDestroy(&pixfinemask);
// Dilate a bit to make sure we get everything.
pixDilateBrick(pixcoarsemask, pixcoarsemask, 3, 3);
Pix* pixmask = pixExpandReplicate(pixcoarsemask, 16);
pixDestroy(&pixcoarsemask);
pixDisplayWrite(pixmask, textord_tabfind_show_images);
// And the image mask with the line and bar remover.
pixAnd(pixht, pixht, pixmask);
pixDestroy(&pixmask);
pixDisplayWrite(pixht, textord_tabfind_show_images);
// Find the individual image regions in the mask image.
*boxa = pixConnComp(pixht, pixa, 8);
pixDestroy(&pixht);
// Rectangularize the individual images. If a sharp edge in vertical and/or
// horizontal occupancy can be found, it indicates a probably rectangular
// image with unwanted bits merged on, so clip to the approximate rectangle.
int npixes = pixaGetCount(*pixa);
for (int i = 0; i < npixes; ++i) {
int x_start, x_end, y_start, y_end;
Pix* img_pix = pixaGetPix(*pixa, i, L_CLONE);
pixDisplayWrite(img_pix, textord_tabfind_show_images);
if (pixNearlyRectangular(img_pix, kMinRectangularFraction,
kMaxRectangularFraction,
kMaxRectangularGradient,
&x_start, &y_start, &x_end, &y_end)) {
// Add 1 to the size as a kludgy flag to indicate to the later stages
// of processing that it is a clipped rectangular image .
Pix* simple_pix = pixCreate(pixGetWidth(img_pix) + 1,
pixGetHeight(img_pix), 1);
pixDestroy(&img_pix);
pixRasterop(simple_pix, x_start, y_start, x_end - x_start,
y_end - y_start, PIX_SET, NULL, 0, 0);
// pixaReplacePix takes ownership of the simple_pix.
pixaReplacePix(*pixa, i, simple_pix, NULL);
img_pix = pixaGetPix(*pixa, i, L_CLONE);
}
// Subtract the pix from the correct location in the master image.
l_int32 x, y, width, height;
pixDisplayWrite(img_pix, textord_tabfind_show_images);
boxaGetBoxGeometry(*boxa, i, &x, &y, &width, &height);
pixRasterop(pix, x, y, width, height, PIX_NOT(PIX_SRC) & PIX_DST,
img_pix, 0, 0);
pixDestroy(&img_pix);
}
#endif
}
/*!
* pixMirrorDetect()
*
* Input: pixs (1 bpp, deskewed, English text)
* &conf (<return> confidence that text is not LR mirror reversed)
* mincount (min number of left + right; use 0 for default)
* debug (1 for debug output; 0 otherwise)
* Return: 0 if OK, 1 on error
*
* 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!
*/
l_int32
pixMirrorDetect(PIX *pixs,
l_float32 *pconf,
l_int32 mincount,
l_int32 debug)
{
l_int32 count1, count2, nmax;
l_float32 nleft, nright;
PIX *pixt0, *pixt1, *pixt2, *pixt3;
SEL *sel1, *sel2;
PROCNAME("pixMirrorDetect");
if (!pconf)
return ERROR_INT("&conf not defined", procName, 1);
*pconf = 0.0;
if (!pixs)
return ERROR_INT("pixs not defined", 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. */
pixt3 = pixMorphCompSequence(pixs, "d1.30", 0);
pixXor(pixt3, pixt3, pixs);
pixt0 = pixMorphCompSequence(pixs, "c15.1", 0);
pixXor(pixt0, pixt0, pixs);
pixAnd(pixt0, pixt0, pixt3);
pixOr(pixt0, pixt0, pixs);
pixDestroy(&pixt3);
/* pixDisplayWrite(pixt0, 1); */
/* Filter the right-facing characters. */
pixt1 = pixHMT(NULL, pixt0, sel1);
pixt3 = pixReduceRankBinaryCascade(pixt1, 1, 1, 0, 0);
pixCountPixels(pixt3, &count1, NULL);
pixDebugFlipDetect("junkpixright", pixs, pixt1, debug);
pixDestroy(&pixt1);
pixDestroy(&pixt3);
/* Filter the left-facing characters. */
pixt2 = pixHMT(NULL, pixt0, sel2);
pixt3 = pixReduceRankBinaryCascade(pixt2, 1, 1, 0, 0);
pixCountPixels(pixt3, &count2, NULL);
pixDebugFlipDetect("junkpixleft", pixs, pixt2, debug);
pixDestroy(&pixt2);
pixDestroy(&pixt3);
nright = (l_float32)count1;
nleft = (l_float32)count2;
nmax = L_MAX(count1, count2);
pixDestroy(&pixt0);
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;
}
/*!
* pixUpDownDetectGeneralDwa()
*
* Input: pixs (1 bpp, deskewed, English text)
* &conf (<return> confidence that text is rightside-up)
* mincount (min number of up + down; use 0 for default)
* npixels (number of pixels removed from each side of word box)
* debug (1 for debug output; 0 otherwise)
* Return: 0 if OK, 1 on error
*
* Notes:
* (1) See the notes in pixUpDownDetectGeneral() for usage.
*/
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, *pixt0, *pixt1, *pixt2, *pixt3, *pixm;
PROCNAME("pixUpDownDetectGeneralDwa");
if (!pconf)
return ERROR_INT("&conf not defined", procName, 1);
*pconf = 0.0;
if (!pixs)
return ERROR_INT("pixs not defined", procName, 1);
if (mincount == 0)
mincount = DEFAULT_MIN_UP_DOWN_COUNT;
if (npixels < 0)
npixels = 0;
/* 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! */
pixt0 = 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;
pixt1 = pixMorphSequenceDwa(pixt0, "o10.1", 0);
boxa = pixConnComp(pixt1, NULL, 8);
pixm = pixCreateTemplate(pixt1);
pixDestroy(&pixt1);
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(). */
pixt1 = pixFlipFHMTGen(NULL, pixt0, flipsel1);
pixt2 = pixFlipFHMTGen(NULL, pixt0, flipsel2);
pixOr(pixt1, pixt1, pixt2);
if (pixm)
pixAnd(pixt1, pixt1, pixm);
pixt3 = pixReduceRankBinaryCascade(pixt1, 1, 1, 0, 0);
pixCountPixels(pixt3, &countup, NULL);
pixDestroy(&pixt1);
pixDestroy(&pixt2);
pixDestroy(&pixt3);
/* Find the ascenders and optionally filter with pixm. */
pixt1 = pixFlipFHMTGen(NULL, pixt0, flipsel3);
pixt2 = pixFlipFHMTGen(NULL, pixt0, flipsel4);
pixOr(pixt1, pixt1, pixt2);
if (pixm)
pixAnd(pixt1, pixt1, pixm);
pixt3 = pixReduceRankBinaryCascade(pixt1, 1, 1, 0, 0);
pixCountPixels(pixt3, &countdown, NULL);
pixDestroy(&pixt1);
pixDestroy(&pixt2);
pixDestroy(&pixt3);
/* 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) pixWrite("junkpixm2", 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(&pixt0);
pixDestroy(&pixm);
return 0;
}
int main(int argc,
char **argv)
{
PIX *pixs, *pix1, *pix2, *pix3, *pix4;
L_REGPARAMS *rp;
if (regTestSetup(argc, argv, &rp))
return 1;
pixs = pixRead("test1.png");
/* pixInvert */
pix1 = pixInvert(NULL, pixs);
pix2 = pixCreateTemplate(pixs); /* into pixd of same size */
pixInvert(pix2, pixs);
regTestWritePixAndCheck(rp, pix1, IFF_PNG); /* 0 */
regTestComparePix(rp, pix1, pix2); /* 1 */
pix3 = pixRead("marge.jpg"); /* into pixd of different size */
pixInvert(pix3, pixs);
regTestComparePix(rp, pix1, pix3); /* 2 */
pixDestroy(&pix1);
pixDestroy(&pix2);
pixDestroy(&pix3);
pix1 = pixOpenBrick(NULL, pixs, 1, 9);
pix2 = pixDilateBrick(NULL, pixs, 1, 9);
/* pixOr */
pix3 = pixCreateTemplate(pixs);
pixOr(pix3, pixs, pix1); /* existing */
pix4 = pixOr(NULL, pixs, pix1); /* new */
regTestWritePixAndCheck(rp, pix3, IFF_PNG); /* 3 */
regTestComparePix(rp, pix3, pix4); /* 4 */
pixCopy(pix4, pix1);
pixOr(pix4, pix4, pixs); /* in-place */
regTestComparePix(rp, pix3, pix4); /* 5 */
pixDestroy(&pix3);
pixDestroy(&pix4);
pix3 = pixCreateTemplate(pixs);
pixOr(pix3, pixs, pix2); /* existing */
pix4 = pixOr(NULL, pixs, pix2); /* new */
regTestWritePixAndCheck(rp, pix3, IFF_PNG); /* 6 */
regTestComparePix(rp, pix3, pix4); /* 7 */
pixCopy(pix4, pix2);
pixOr(pix4, pix4, pixs); /* in-place */
regTestComparePix(rp, pix3, pix4); /* 8 */
pixDestroy(&pix3);
pixDestroy(&pix4);
/* pixAnd */
pix3 = pixCreateTemplate(pixs);
pixAnd(pix3, pixs, pix1); /* existing */
pix4 = pixAnd(NULL, pixs, pix1); /* new */
regTestWritePixAndCheck(rp, pix3, IFF_PNG); /* 9 */
regTestComparePix(rp, pix3, pix4); /* 10 */
pixCopy(pix4, pix1);
pixAnd(pix4, pix4, pixs); /* in-place */
regTestComparePix(rp, pix3, pix4); /* 11 */
pixDestroy(&pix3);
pixDestroy(&pix4);
pix3 = pixCreateTemplate(pixs);
pixAnd(pix3, pixs, pix2); /* existing */
pix4 = pixAnd(NULL, pixs, pix2); /* new */
regTestWritePixAndCheck(rp, pix3, IFF_PNG); /* 12 */
regTestComparePix(rp, pix3, pix4); /* 13 */
pixCopy(pix4, pix2);
pixAnd(pix4, pix4, pixs); /* in-place */
regTestComparePix(rp, pix3, pix4); /* 14 */
pixDestroy(&pix3);
pixDestroy(&pix4);
/* pixXor */
pix3 = pixCreateTemplate(pixs);
pixXor(pix3, pixs, pix1); /* existing */
pix4 = pixXor(NULL, pixs, pix1); /* new */
regTestWritePixAndCheck(rp, pix3, IFF_PNG); /* 15 */
regTestComparePix(rp, pix3, pix4); /* 16 */
pixCopy(pix4, pix1);
pixXor(pix4, pix4, pixs); /* in-place */
regTestComparePix(rp, pix3, pix4); /* 17 */
pixDestroy(&pix3);
pixDestroy(&pix4);
pix3 = pixCreateTemplate(pixs);
pixXor(pix3, pixs, pix2); /* existing */
pix4 = pixXor(NULL, pixs, pix2); /* new */
regTestWritePixAndCheck(rp, pix3, IFF_PNG); /* 18 */
regTestComparePix(rp, pix3, pix4); /* 19 */
pixCopy(pix4, pix2);
pixXor(pix4, pix4, pixs); /* in-place */
regTestComparePix(rp, pix3, pix4); /* 20 */
pixDestroy(&pix3);
pixDestroy(&pix4);
/* pixSubtract */
pix3 = pixCreateTemplate(pixs);
pixSubtract(pix3, pixs, pix1); /* existing */
pix4 = pixSubtract(NULL, pixs, pix1); /* new */
regTestWritePixAndCheck(rp, pix3, IFF_PNG); /* 21 */
regTestComparePix(rp, pix3, pix4); /* 22 */
pixCopy(pix4, pix1);
pixSubtract(pix4, pixs, pix4); /* in-place */
regTestComparePix(rp, pix3, pix4); /* 23 */
pixDestroy(&pix3);
pixDestroy(&pix4);
pix3 = pixCreateTemplate(pixs);
pixSubtract(pix3, pixs, pix2); /* existing */
pix4 = pixSubtract(NULL, pixs, pix2); /* new */
regTestWritePixAndCheck(rp, pix3, IFF_PNG); /* 24 */
regTestComparePix(rp, pix3, pix4); /* 25 */
pixCopy(pix4, pix2);
pixSubtract(pix4, pixs, pix4); /* in-place */
regTestComparePix(rp, pix3, pix4); /* 26 */
pixDestroy(&pix3);
pixDestroy(&pix4);
pix4 = pixRead("marge.jpg");
pixSubtract(pix4, pixs, pixs); /* subtract from itself; should be empty */
pix3 = pixCreateTemplate(pixs);
regTestComparePix(rp, pix3, pix4); /* 27*/
pixDestroy(&pix3);
pixDestroy(&pix4);
pixSubtract(pixs, pixs, pixs); /* subtract from itself; should be empty */
pix3 = pixCreateTemplate(pixs);
regTestComparePix(rp, pix3, pixs); /* 28*/
pixDestroy(&pix3);
pixDestroy(&pixs);
pixDestroy(&pix1);
pixDestroy(&pix2);
return regTestCleanup(rp);
}
int
main(int argc, char **argv) {
if (argc < 3)
return usage(argv[0]);
char highlight = 0;
char ignore_scrollbars = 1;
/* Default output filename; can be overridden by command line. */
const char *output_filename = "highlight.png";
int argi = 1;
for (; argi < argc; ++argi) {
if (strcmp("--highlight", argv[argi]) == 0) {
highlight = 1;
} else if (strcmp("--no-ignore-scrollbars", argv[argi]) == 0) {
ignore_scrollbars = 0;
} else if (strcmp("--output", argv[argi]) == 0) {
if (argi + 1 >= argc) {
fprintf(stderr, "missing argument to --output\n");
return 1;
}
output_filename = argv[++argi];
} else {
break;
}
}
if (argc - argi < 2)
return usage(argv[0]);
PIX *a = pixRead(argv[argi]);
PIX *b = pixRead(argv[argi + 1]);
if (!a) {
fprintf(stderr, "Failed to open %s\n", argv[argi]);
return 1;
}
if (!b) {
fprintf(stderr, "Failed to open %s\n", argv[argi + 1]);
return 1;
}
if (pixGetWidth(a) != pixGetWidth(b) ||
pixGetHeight(a) != pixGetHeight(b)) {
fprintf(stderr, "Inputs are difference sizes\n");
return 1;
}
PIX *delta = pixAbsDifference(a, b);
pixInvert(delta, delta);
if (!highlight)
pixDestroy(&a);
pixDestroy(&b);
PIX *deltagray = pixConvertRGBToGray(delta, 0, 0, 0);
pixDestroy(&delta);
PIX *deltabinary = pixThresholdToBinary(deltagray, 254);
PIX *deltabinaryclipped;
const int clipwidth = pixGetWidth(deltabinary) - 15;
const int clipheight = pixGetHeight(deltabinary) - 15;
if (ignore_scrollbars && clipwidth > 0 && clipheight > 0) {
BOX *clip = boxCreate(0, 0, clipwidth, clipheight);
deltabinaryclipped = pixClipRectangle(deltabinary, clip, NULL);
boxDestroy(&clip);
pixDestroy(&deltabinary);
} else {
deltabinaryclipped = deltabinary;
deltabinary = NULL;
}
PIX *hopened = pixOpenBrick(NULL, deltabinaryclipped, 3, 1);
PIX *vopened = pixOpenBrick(NULL, deltabinaryclipped, 1, 3);
pixDestroy(&deltabinaryclipped);
PIX *opened = pixOr(NULL, hopened, vopened);
pixDestroy(&hopened);
pixDestroy(&vopened);
l_int32 count;
pixCountPixels(opened, &count, NULL);
fprintf(stderr, "%d\n", count);
if (count && highlight) {
PIX *d1 = pixDilateBrick(NULL, opened, 7, 7);
PIX *d2 = pixDilateBrick(NULL, opened, 3, 3);
pixInvert(d2, d2);
pixAnd(d1, d1, d2);
pixPaintThroughMask(a, d1, 0, 0, 0xff << 24);
pixWrite(output_filename, a, IFF_PNG);
}
return count > 0;
}
/*!
* \brief pixUpDownDetectGeneral()
*
* \param[in] pixs 1 bpp, deskewed, English text, 150 - 300 ppi
* \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 pixOrientDetect() for other details.
* (2) %conf is the normalized difference between the number of
* detected up and down ascenders, assuming that the text
* is either rightside-up or upside-down and not rotated
* at a 90 degree angle.
* (3) The typical mode of operation is %npixels == 0.
* If %npixels > 0, this removes HMT matches at the
* beginning and ending of "words." This is useful for
* pages that may have mostly digits, because if npixels == 0,
* leading "1" and "3" digits can register as having
* ascenders or descenders, and "7" digits can match descenders.
* Consequently, a page image of only digits may register
* as being upside-down.
* (4) We want to count the number of instances found using the HMT.
* An expensive way to do this would be to count the
* number of connected components. A cheap way is to do a rank
* reduction cascade that reduces each component to a single
* pixel, and results (after two or three 2x reductions)
* in one pixel for each of the original components.
* After the reduction, you have a much smaller pix over
* which to count pixels. We do only 2 reductions, because
* this function is designed to work for input pix between
* 150 and 300 ppi, and an 8x reduction on a 150 ppi image
* is going too far -- components will get merged.
* </pre>
*/
l_int32
pixUpDownDetectGeneral(PIX *pixs,
l_float32 *pconf,
l_int32 mincount,
l_int32 npixels,
l_int32 debug)
{
l_int32 countup, countdown, nmax;
l_float32 nup, ndown;
PIX *pix0, *pix1, *pix2, *pix3, *pixm;
SEL *sel1, *sel2, *sel3, *sel4;
PROCNAME("pixUpDownDetectGeneral");
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");
sel1 = selCreateFromString(textsel1, 5, 6, NULL);
sel2 = selCreateFromString(textsel2, 5, 6, NULL);
sel3 = selCreateFromString(textsel3, 5, 6, NULL);
sel4 = selCreateFromString(textsel4, 5, 6, NULL);
/* 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. */
pix0 = pixMorphCompSequence(pixs, "c1.8 + c30.1", 0);
/* 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 = pixMorphSequence(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 at the beginning of this function. */
pix1 = pixHMT(NULL, pix0, sel1);
pix2 = pixHMT(NULL, pix0, sel2);
pixOr(pix1, pix1, pix2);
if (pixm)
pixAnd(pix1, pix1, pixm);
pix3 = pixReduceRankBinaryCascade(pix1, 1, 1, 0, 0);
pixCountPixels(pix3, &countup, NULL);
pixDebugFlipDetect("/tmp/lept/orient/up.png", pixs, pix1, debug);
pixDestroy(&pix1);
pixDestroy(&pix2);
pixDestroy(&pix3);
/* Find the ascenders and optionally filter with pixm. */
pix1 = pixHMT(NULL, pix0, sel3);
pix2 = pixHMT(NULL, pix0, sel4);
pixOr(pix1, pix1, pix2);
if (pixm)
pixAnd(pix1, pix1, pixm);
pix3 = pixReduceRankBinaryCascade(pix1, 1, 1, 0, 0);
pixCountPixels(pix3, &countdown, NULL);
pixDebugFlipDetect("/tmp/lept/orient/down.png", pixs, pix1, debug);
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/pixm1.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);
selDestroy(&sel1);
selDestroy(&sel2);
selDestroy(&sel3);
selDestroy(&sel4);
return 0;
}
