/**
* Auto page segmentation. Divide the page image into blocks of uniform
* text linespacing and images.
*
* Resolution (in ppi) is derived from the input image.
*
* The output goes in the blocks list with corresponding TO_BLOCKs in the
* to_blocks list.
*
* If single_column is true, then no attempt is made to divide the image
* into columns, but multiple blocks are still made if the text is of
* non-uniform linespacing.
*
* If osd (orientation and script detection) is true then that is performed
* as well. If only_osd is true, then only orientation and script detection is
* performed. If osd is desired, (osd or only_osd) then osr_tess must be
* another Tesseract that was initialized especially for osd, and the results
* will be output into osr (orientation and script result).
*/
int Tesseract::AutoPageSeg(PageSegMode pageseg_mode,
BLOCK_LIST* blocks, TO_BLOCK_LIST* to_blocks,
Tesseract* osd_tess, OSResults* osr) {
if (textord_debug_images) {
WriteDebugBackgroundImage(textord_debug_printable, pix_binary_);
}
Pix* photomask_pix = NULL;
Pix* musicmask_pix = NULL;
// The blocks made by the ColumnFinder. Moved to blocks before return.
BLOCK_LIST found_blocks;
TO_BLOCK_LIST temp_blocks;
bool single_column = !PSM_COL_FIND_ENABLED(pageseg_mode);
bool osd_enabled = PSM_OSD_ENABLED(pageseg_mode);
bool osd_only = pageseg_mode == PSM_OSD_ONLY;
ColumnFinder* finder = SetupPageSegAndDetectOrientation(
single_column, osd_enabled, osd_only, blocks, osd_tess, osr,
&temp_blocks, &photomask_pix, &musicmask_pix);
int result = 0;
if (finder != NULL) {
TO_BLOCK_IT to_block_it(&temp_blocks);
TO_BLOCK* to_block = to_block_it.data();
if (musicmask_pix != NULL) {
// TODO(rays) pass the musicmask_pix into FindBlocks and mark music
// blocks separately. For now combine with photomask_pix.
pixOr(photomask_pix, photomask_pix, musicmask_pix);
}
if (equ_detect_) {
finder->SetEquationDetect(equ_detect_);
}
result = finder->FindBlocks(pageseg_mode, scaled_color_, scaled_factor_,
to_block, photomask_pix,
pix_thresholds_, pix_grey_,
&found_blocks, to_blocks);
if (result >= 0)
finder->GetDeskewVectors(&deskew_, &reskew_);
delete finder;
}
pixDestroy(&photomask_pix);
pixDestroy(&musicmask_pix);
if (result < 0) return result;
blocks->clear();
BLOCK_IT block_it(blocks);
// Move the found blocks to the input/output blocks.
block_it.add_list_after(&found_blocks);
if (textord_debug_images) {
// The debug image is no longer needed so delete it.
unlink(AlignedBlob::textord_debug_pix().string());
}
return result;
}
/**
* Auto page segmentation. Divide the page image into blocks of uniform
* text linespacing and images.
*
* Resolution (in ppi) is derived from the input image.
*
* The output goes in the blocks list with corresponding TO_BLOCKs in the
* to_blocks list.
*
* If !PSM_COL_FIND_ENABLED(pageseg_mode), then no attempt is made to divide
* the image into columns, but multiple blocks are still made if the text is
* of non-uniform linespacing.
*
* If diacritic_blobs is non-null, then diacritics/noise blobs, that would
* confuse layout anaylsis by causing textline overlap, are placed there,
* with the expectation that they will be reassigned to words later and
* noise/diacriticness determined via classification.
*
* If osd (orientation and script detection) is true then that is performed
* as well. If only_osd is true, then only orientation and script detection is
* performed. If osd is desired, (osd or only_osd) then osr_tess must be
* another Tesseract that was initialized especially for osd, and the results
* will be output into osr (orientation and script result).
*/
int Tesseract::AutoPageSeg(PageSegMode pageseg_mode, BLOCK_LIST* blocks,
TO_BLOCK_LIST* to_blocks,
BLOBNBOX_LIST* diacritic_blobs, Tesseract* osd_tess,
OSResults* osr) {
Pix* photomask_pix = NULL;
Pix* musicmask_pix = NULL;
// The blocks made by the ColumnFinder. Moved to blocks before return.
BLOCK_LIST found_blocks;
TO_BLOCK_LIST temp_blocks;
ColumnFinder* finder = SetupPageSegAndDetectOrientation(
pageseg_mode, blocks, osd_tess, osr, &temp_blocks, &photomask_pix,
&musicmask_pix);
int result = 0;
if (finder != NULL) {
TO_BLOCK_IT to_block_it(&temp_blocks);
TO_BLOCK* to_block = to_block_it.data();
if (musicmask_pix != NULL) {
// TODO(rays) pass the musicmask_pix into FindBlocks and mark music
// blocks separately. For now combine with photomask_pix.
pixOr(photomask_pix, photomask_pix, musicmask_pix);
}
if (equ_detect_) {
finder->SetEquationDetect(equ_detect_);
}
result = finder->FindBlocks(pageseg_mode, scaled_color_, scaled_factor_,
to_block, photomask_pix, pix_thresholds_,
pix_grey_, &pixa_debug_, &found_blocks,
diacritic_blobs, to_blocks);
if (result >= 0)
finder->GetDeskewVectors(&deskew_, &reskew_);
delete finder;
}
pixDestroy(&photomask_pix);
pixDestroy(&musicmask_pix);
if (result < 0) return result;
blocks->clear();
BLOCK_IT block_it(blocks);
// Move the found blocks to the input/output blocks.
block_it.add_list_after(&found_blocks);
return result;
}
l_int32 main(int argc,
char **argv)
{
l_int32 bx, by, bw, bh;
l_uint32 pixval;
BOX *box1, *box2;
BOXA *boxa;
PIX *pixs, *pixm, *pixd;
PIX *pix0, *pix1, *pix2, *pix3, *pix4, *pix5, *pix6;
L_REGPARAMS *rp;
if (regTestSetup(argc, argv, &rp))
return 1;
/* Find a mask for repainting pixels */
pixs = pixRead("amoris.2.150.jpg");
pix1 = MakeReplacementMask(pixs);
boxa = pixConnCompBB(pix1, 8);
box1 = boxaGetBox(boxa, 0, L_COPY);
boxaDestroy(&boxa);
/*--------------------------------------------------------*
* Show the individual steps *
*--------------------------------------------------------*/
/* Locate a good tile to use */
pixFindRepCloseTile(pixs, box1, L_VERT, 20, 30, 7, &box2, 1);
pix0 = pixCopy(NULL, pix1);
pixRenderBox(pix0, box2, 2, L_SET_PIXELS);
/* Make a patch using this tile */
boxGetGeometry(box1, &bx, &by, &bw, &bh);
pix2 = pixClipRectangle(pixs, box2, NULL);
regTestWritePixAndCheck(rp, pix2, IFF_PNG); /* 0 */
pixDisplayWithTitle(pix2, 400, 100, NULL, rp->display);
pix3 = pixMirroredTiling(pix2, bw, bh);
regTestWritePixAndCheck(rp, pix3, IFF_PNG); /* 1 */
pixDisplayWithTitle(pix3, 1000, 0, NULL, rp->display);
/* Paint the patch through the mask */
pixd = pixCopy(NULL, pixs);
pixm = pixClipRectangle(pix1, box1, NULL);
pixCombineMaskedGeneral(pixd, pix3, pixm, bx, by);
regTestWritePixAndCheck(rp, pixd, IFF_PNG); /* 2 */
pixDisplayWithTitle(pixd, 0, 0, NULL, rp->display);
boxDestroy(&box2);
pixDestroy(&pixm);
pixDestroy(&pixd);
pixDestroy(&pix2);
/* Blend two patches and then overlay. Use the previous
* tile found vertically and a new one found horizontally. */
pixFindRepCloseTile(pixs, box1, L_HORIZ, 20, 30, 7, &box2, 1);
pixRenderBox(pix0, box2, 2, L_SET_PIXELS);
regTestWritePixAndCheck(rp, pix0, IFF_TIFF_G4); /* 3 */
pixDisplayWithTitle(pix0, 100, 100, NULL, rp->display);
pix2 = pixClipRectangle(pixs, box2, NULL);
pix4 = pixMirroredTiling(pix2, bw, bh);
regTestWritePixAndCheck(rp, pix4, IFF_PNG); /* 4 */
pixDisplayWithTitle(pix4, 1100, 0, NULL, rp->display);
pix5 = pixBlend(pix3, pix4, 0, 0, 0.5);
regTestWritePixAndCheck(rp, pix5, IFF_PNG); /* 5 */
pixDisplayWithTitle(pix5, 1200, 0, NULL, rp->display);
pix6 = pixClipRectangle(pix1, box1, NULL);
pixd = pixCopy(NULL, pixs);
pixCombineMaskedGeneral(pixd, pix5, pix6, bx, by);
regTestWritePixAndCheck(rp, pixd, IFF_PNG); /* 6 */
pixDisplayWithTitle(pixd, 700, 200, NULL, rp->display);
boxDestroy(&box2);
pixDestroy(&pixd);
pixDestroy(&pix0);
pixDestroy(&pix2);
pixDestroy(&pix3);
pixDestroy(&pix4);
pixDestroy(&pix5);
pixDestroy(&pix6);
/*--------------------------------------------------------*
* Show painting from a color near region *
*--------------------------------------------------------*/
pix2 = pixCopy(NULL, pixs);
pixGetColorNearMaskBoundary(pix2, pix1, box1, 20, &pixval, 0);
pix3 = pixClipRectangle(pix1, box1, NULL);
boxGetGeometry(box1, &bx, &by, NULL, NULL);
pixSetMaskedGeneral(pix2, pix3, pixval, bx, by);
regTestWritePixAndCheck(rp, pix2, IFF_PNG); /* 7 */
pixDisplayWithTitle(pix2, 0, 0, NULL, rp->display);
boxDestroy(&box1);
pixDestroy(&pix2);
pixDestroy(&pix3);
/*--------------------------------------------------------*
* Use the higher-level function *
*--------------------------------------------------------*/
/* Use various tile selections and tile blending with one component */
pix2 = pixCopy(NULL, pixs);
pix3 = pixCopy(NULL, pixs);
pix4 = pixCopy(NULL, pixs);
pixPaintSelfThroughMask(pix2, pix1, 0, 0, L_HORIZ, 30, 50, 5, 10);
pixPaintSelfThroughMask(pix3, pix1, 0, 0, L_VERT, 30, 50, 5, 0);
pixPaintSelfThroughMask(pixs, pix1, 0, 0, L_BOTH_DIRECTIONS, 30, 50, 5, 20);
regTestWritePixAndCheck(rp, pix2, IFF_PNG); /* 8 */
regTestWritePixAndCheck(rp, pix3, IFF_PNG); /* 9 */
regTestWritePixAndCheck(rp, pixs, IFF_PNG); /* 10 */
pixDisplayWithTitle(pix2, 300, 0, NULL, rp->display);
pixDisplayWithTitle(pix3, 500, 0, NULL, rp->display);
pixDisplayWithTitle(pixs, 700, 0, NULL, rp->display);
/* Test with two components; */
pix5 = pixFlipLR(NULL, pix1);
pixOr(pix5, pix5, pix1);
pixPaintSelfThroughMask(pix4, pix5, 0, 0, L_BOTH_DIRECTIONS, 50, 100, 5, 9);
regTestWritePixAndCheck(rp, pix4, IFF_PNG); /* 11 */
pixDisplayWithTitle(pix4, 900, 0, NULL, rp->display);
pixDestroy(&pixs);
pixDestroy(&pix1);
pixDestroy(&pix2);
pixDestroy(&pix3);
pixDestroy(&pix4);
pixDestroy(&pix5);
return regTestCleanup(rp);
}
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;
}
/*!
* pixGetRegionsBinary()
*
* Input: pixs (1 bpp, assumed to be 300 to 400 ppi)
* &pixhm (<optional return> halftone mask)
* &pixtm (<optional return> textline mask)
* &pixtb (<optional return> textblock mask)
* debug (flag: set to 1 for debug output)
* Return: 0 if OK, 1 on error
*
* Notes:
* (1) It is best to deskew the image before segmenting.
* (2) The debug flag enables a number of outputs. These
* are included to show how to generate and save/display
* these results.
*/
l_int32
pixGetRegionsBinary(PIX *pixs,
PIX **ppixhm,
PIX **ppixtm,
PIX **ppixtb,
l_int32 debug)
{
char *tempname;
l_int32 htfound, tlfound;
PIX *pixr, *pixt1, *pixt2;
PIX *pixtext; /* text pixels only */
PIX *pixhm2; /* halftone mask; 2x reduction */
PIX *pixhm; /* halftone mask; */
PIX *pixtm2; /* textline mask; 2x reduction */
PIX *pixtm; /* textline mask */
PIX *pixvws; /* vertical white space mask */
PIX *pixtb2; /* textblock mask; 2x reduction */
PIX *pixtbf2; /* textblock mask; 2x reduction; small comps filtered */
PIX *pixtb; /* textblock mask */
PROCNAME("pixGetRegionsBinary");
if (ppixhm) *ppixhm = NULL;
if (ppixtm) *ppixtm = NULL;
if (ppixtb) *ppixtb = NULL;
if (!pixs)
return ERROR_INT("pixs not defined", procName, 1);
if (pixGetDepth(pixs) != 1)
return ERROR_INT("pixs not 1 bpp", procName, 1);
/* 2x reduce, to 150 -200 ppi */
pixr = pixReduceRankBinaryCascade(pixs, 1, 0, 0, 0);
pixDisplayWrite(pixr, debug);
/* Get the halftone mask */
pixhm2 = pixGenHalftoneMask(pixr, &pixtext, &htfound, debug);
/* Get the textline mask from the text pixels */
pixtm2 = pixGenTextlineMask(pixtext, &pixvws, &tlfound, debug);
/* Get the textblock mask from the textline mask */
pixtb2 = pixGenTextblockMask(pixtm2, pixvws, debug);
pixDestroy(&pixr);
pixDestroy(&pixtext);
pixDestroy(&pixvws);
/* Remove small components from the mask, where a small
* component is defined as one with both width and height < 60 */
pixtbf2 = pixSelectBySize(pixtb2, 60, 60, 4, L_SELECT_IF_EITHER,
L_SELECT_IF_GTE, NULL);
pixDestroy(&pixtb2);
pixDisplayWriteFormat(pixtbf2, debug, IFF_PNG);
/* Expand all masks to full resolution, and do filling or
* small dilations for better coverage. */
pixhm = pixExpandReplicate(pixhm2, 2);
pixt1 = pixSeedfillBinary(NULL, pixhm, pixs, 8);
pixOr(pixhm, pixhm, pixt1);
pixDestroy(&pixt1);
pixDisplayWriteFormat(pixhm, debug, IFF_PNG);
pixt1 = pixExpandReplicate(pixtm2, 2);
pixtm = pixDilateBrick(NULL, pixt1, 3, 3);
pixDestroy(&pixt1);
pixDisplayWriteFormat(pixtm, debug, IFF_PNG);
pixt1 = pixExpandReplicate(pixtbf2, 2);
pixtb = pixDilateBrick(NULL, pixt1, 3, 3);
pixDestroy(&pixt1);
pixDisplayWriteFormat(pixtb, debug, IFF_PNG);
pixDestroy(&pixhm2);
pixDestroy(&pixtm2);
pixDestroy(&pixtbf2);
/* Debug: identify objects that are neither text nor halftone image */
if (debug) {
pixt1 = pixSubtract(NULL, pixs, pixtm); /* remove text pixels */
pixt2 = pixSubtract(NULL, pixt1, pixhm); /* remove halftone pixels */
pixDisplayWriteFormat(pixt2, 1, IFF_PNG);
pixDestroy(&pixt1);
pixDestroy(&pixt2);
}
/* Debug: display textline components with random colors */
if (debug) {
l_int32 w, h;
BOXA *boxa;
PIXA *pixa;
boxa = pixConnComp(pixtm, &pixa, 8);
pixGetDimensions(pixtm, &w, &h, NULL);
pixt1 = pixaDisplayRandomCmap(pixa, w, h);
pixcmapResetColor(pixGetColormap(pixt1), 0, 255, 255, 255);
pixDisplay(pixt1, 100, 100);
pixDisplayWriteFormat(pixt1, 1, IFF_PNG);
pixaDestroy(&pixa);
boxaDestroy(&boxa);
pixDestroy(&pixt1);
}
/* Debug: identify the outlines of each textblock */
if (debug) {
PIXCMAP *cmap;
PTAA *ptaa;
ptaa = pixGetOuterBordersPtaa(pixtb);
tempname = genTempFilename("/tmp", "tb_outlines.ptaa", 0, 0);
ptaaWrite(tempname, ptaa, 1);
FREE(tempname);
pixt1 = pixRenderRandomCmapPtaa(pixtb, ptaa, 1, 16, 1);
cmap = pixGetColormap(pixt1);
pixcmapResetColor(cmap, 0, 130, 130, 130);
pixDisplay(pixt1, 500, 100);
pixDisplayWriteFormat(pixt1, 1, IFF_PNG);
pixDestroy(&pixt1);
ptaaDestroy(&ptaa);
}
/* Debug: get b.b. for all mask components */
if (debug) {
BOXA *bahm, *batm, *batb;
bahm = pixConnComp(pixhm, NULL, 4);
batm = pixConnComp(pixtm, NULL, 4);
batb = pixConnComp(pixtb, NULL, 4);
tempname = genTempFilename("/tmp", "htmask.boxa", 0, 0);
boxaWrite(tempname, bahm);
FREE(tempname);
tempname = genTempFilename("/tmp", "textmask.boxa", 0, 0);
boxaWrite(tempname, batm);
FREE(tempname);
tempname = genTempFilename("/tmp", "textblock.boxa", 0, 0);
boxaWrite(tempname, batb);
FREE(tempname);
boxaDestroy(&bahm);
boxaDestroy(&batm);
boxaDestroy(&batb);
}
if (ppixhm)
*ppixhm = pixhm;
else
pixDestroy(&pixhm);
if (ppixtm)
*ppixtm = pixtm;
else
pixDestroy(&pixtm);
if (ppixtb)
*ppixtb = pixtb;
else
pixDestroy(&pixtb);
return 0;
}
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;
}
// Finds the sample for each font, class pair that has least maximum
// distance to all the other samples of the same font, class.
// OrganizeByFontAndClass must have been already called.
void TrainingSampleSet::ComputeCanonicalSamples(const IntFeatureMap &map,
bool debug) {
ASSERT_HOST(font_class_array_ != NULL);
IntFeatureDist f_table;
if (debug) tprintf("feature table size %d\n", map.sparse_size());
f_table.Init(&map);
int worst_s1 = 0;
int worst_s2 = 0;
double global_worst_dist = 0.0;
// Compute distances independently for each font and char index.
int font_size = font_id_map_.CompactSize();
for (int font_index = 0; font_index < font_size; ++font_index) {
int font_id = font_id_map_.CompactToSparse(font_index);
for (int c = 0; c < unicharset_size_; ++c) {
int samples_found = 0;
FontClassInfo &fcinfo = (*font_class_array_)(font_index, c);
if (fcinfo.samples.size() == 0 ||
(kTestChar >= 0 && c != kTestChar)) {
fcinfo.canonical_sample = -1;
fcinfo.canonical_dist = 0.0f;
if (debug) tprintf("Skipping class %d\n", c);
continue;
}
// The canonical sample will be the one with the min_max_dist, which
// is the sample with the lowest maximum distance to all other samples.
double min_max_dist = 2.0;
// We keep track of the farthest apart pair (max_s1, max_s2) which
// are max_max_dist apart, so we can see how bad the variability is.
double max_max_dist = 0.0;
int max_s1 = 0;
int max_s2 = 0;
fcinfo.canonical_sample = fcinfo.samples[0];
fcinfo.canonical_dist = 0.0f;
for (int i = 0; i < fcinfo.samples.size(); ++i) {
int s1 = fcinfo.samples[i];
const GenericVector <int> &features1 = samples_[s1]->indexed_features();
f_table.Set(features1, features1.size(), true);
double max_dist = 0.0;
// Run the full squared-order search for similar samples. It is still
// reasonably fast because f_table.FeatureDistance is fast, but we
// may have to reconsider if we start playing with too many samples
// of a single char/font.
for (int j = 0; j < fcinfo.samples.size(); ++j) {
int s2 = fcinfo.samples[j];
if (samples_[s2]->class_id() != c ||
samples_[s2]->font_id() != font_id ||
s2 == s1)
continue;
GenericVector <int> features2 = samples_[s2]->indexed_features();
double dist = f_table.FeatureDistance(features2);
if (dist > max_dist) {
max_dist = dist;
if (dist > max_max_dist) {
max_s1 = s1;
max_s2 = s2;
}
}
}
// Using Set(..., false) is far faster than re initializing, due to
// the sparseness of the feature space.
f_table.Set(features1, features1.size(), false);
samples_[s1]->set_max_dist(max_dist);
++samples_found;
if (max_dist < min_max_dist) {
fcinfo.canonical_sample = s1;
fcinfo.canonical_dist = max_dist;
}
UpdateRange(max_dist, &min_max_dist, &max_max_dist);
}
if (max_max_dist > global_worst_dist) {
// Keep a record of the worst pair over all characters/fonts too.
global_worst_dist = max_max_dist;
worst_s1 = max_s1;
worst_s2 = max_s2;
}
if (debug) {
tprintf("Found %d samples of class %d=%s, font %d, "
"dist range [%g, %g], worst pair= %s, %s\n",
samples_found, c, unicharset_.debug_str(c).string(),
font_index, min_max_dist, max_max_dist,
SampleToString(*samples_[max_s1]).string(),
SampleToString(*samples_[max_s2]).string());
}
}
}
if (debug) {
tprintf("Global worst dist = %g, between sample %d and %d\n",
global_worst_dist, worst_s1, worst_s2);
Pix *pix1 = DebugSample(unicharset_, samples_[worst_s1]);
Pix *pix2 = DebugSample(unicharset_, samples_[worst_s2]);
pixOr(pix1, pix1, pix2);
pixWrite("worstpair.png", pix1, IFF_PNG);
pixDestroy(&pix1);
pixDestroy(&pix2);
}
}
int main(int argc,
char **argv)
{
l_int32 w, h, n, i, sum, sumi, empty;
BOX *box1, *box2, *box3, *box4;
BOXA *boxa, *boxat;
NUMA *na1, *na2, *na3, *na4, *na5;
NUMA *na2i, *na3i, *na4i, *nat, *naw, *nah;
PIX *pixs, *pixc, *pixt, *pixt2, *pixd, *pixcount;
PIXA *pixas, *pixad, *pixac;
pixDisplayWrite(NULL, -1);
/* Draw 4 filled boxes of different sizes */
pixs = pixCreate(200, 200, 1);
box1 = boxCreate(10, 10, 20, 30);
box2 = boxCreate(50, 10, 40, 20);
box3 = boxCreate(110, 10, 35, 5);
box4 = boxCreate(160, 10, 5, 15);
boxa = boxaCreate(4);
boxaAddBox(boxa, box1, L_INSERT);
boxaAddBox(boxa, box2, L_INSERT);
boxaAddBox(boxa, box3, L_INSERT);
boxaAddBox(boxa, box4, L_INSERT);
pixRenderBox(pixs, box1, 1, L_SET_PIXELS);
pixRenderBox(pixs, box2, 1, L_SET_PIXELS);
pixRenderBox(pixs, box3, 1, L_SET_PIXELS);
pixRenderBox(pixs, box4, 1, L_SET_PIXELS);
pixt = pixFillClosedBorders(pixs, 4);
pixDisplayWrite(pixt, 1);
pixt2 = pixCreateTemplate(pixs);
pixRenderHashBox(pixt2, box1, 6, 4, L_POS_SLOPE_LINE, 1, L_SET_PIXELS);
pixRenderHashBox(pixt2, box2, 7, 2, L_POS_SLOPE_LINE, 1, L_SET_PIXELS);
pixRenderHashBox(pixt2, box3, 4, 2, L_VERTICAL_LINE, 1, L_SET_PIXELS);
pixRenderHashBox(pixt2, box4, 3, 1, L_HORIZONTAL_LINE, 1, L_SET_PIXELS);
pixDisplayWrite(pixt2, 1);
/* Exercise the parameters */
pixd = pixSelectBySize(pixt, 0, 22, 8, L_SELECT_HEIGHT,
L_SELECT_IF_GT, NULL);
count_pieces(pixd, 1);
pixd = pixSelectBySize(pixt, 0, 30, 8, L_SELECT_HEIGHT,
L_SELECT_IF_LT, NULL);
count_pieces(pixd, 3);
pixd = pixSelectBySize(pixt, 0, 5, 8, L_SELECT_HEIGHT,
L_SELECT_IF_GT, NULL);
count_pieces(pixd, 3);
pixd = pixSelectBySize(pixt, 0, 6, 8, L_SELECT_HEIGHT,
L_SELECT_IF_LT, NULL);
count_pieces(pixd, 1);
pixd = pixSelectBySize(pixt, 20, 0, 8, L_SELECT_WIDTH,
L_SELECT_IF_GT, NULL);
count_pieces(pixd, 2);
pixd = pixSelectBySize(pixt, 31, 0, 8, L_SELECT_WIDTH,
L_SELECT_IF_LT, NULL);
count_pieces(pixd, 2);
pixd = pixSelectBySize(pixt, 21, 10, 8, L_SELECT_IF_EITHER,
L_SELECT_IF_LT, NULL);
count_pieces(pixd, 3);
pixd = pixSelectBySize(pixt, 20, 30, 8, L_SELECT_IF_EITHER,
L_SELECT_IF_GT, NULL);
count_pieces(pixd, 2);
pixd = pixSelectBySize(pixt, 22, 32, 8, L_SELECT_IF_BOTH,
L_SELECT_IF_LT, NULL);
count_pieces(pixd, 2);
pixd = pixSelectBySize(pixt, 6, 32, 8, L_SELECT_IF_BOTH,
L_SELECT_IF_LT, NULL);
count_pieces(pixd, 1);
pixd = pixSelectBySize(pixt, 5, 25, 8, L_SELECT_IF_BOTH,
L_SELECT_IF_GT, NULL);
count_pieces(pixd, 1);
pixd = pixSelectBySize(pixt, 25, 5, 8, L_SELECT_IF_BOTH,
L_SELECT_IF_GT, NULL);
count_pieces(pixd, 1);
pixd = pixSelectByPerimToAreaRatio(pixt, 0.3, 8, L_SELECT_IF_GT, NULL);
count_pieces(pixd, 2);
pixd = pixSelectByPerimToAreaRatio(pixt, 0.15, 8, L_SELECT_IF_GT, NULL);
count_pieces(pixd, 3);
pixd = pixSelectByPerimToAreaRatio(pixt, 0.4, 8, L_SELECT_IF_LTE, NULL);
count_pieces(pixd, 2);
pixd = pixSelectByPerimToAreaRatio(pixt, 0.45, 8, L_SELECT_IF_LT, NULL);
count_pieces(pixd, 3);
pixd = pixSelectByPerimSizeRatio(pixt2, 2.3, 8, L_SELECT_IF_GT, NULL);
count_pieces(pixd, 2);
pixd = pixSelectByPerimSizeRatio(pixt2, 1.2, 8, L_SELECT_IF_GT, NULL);
count_pieces(pixd, 3);
pixd = pixSelectByPerimSizeRatio(pixt2, 1.7, 8, L_SELECT_IF_LTE, NULL);
count_pieces(pixd, 1);
pixd = pixSelectByPerimSizeRatio(pixt2, 2.9, 8, L_SELECT_IF_LT, NULL);
count_pieces(pixd, 3);
pixd = pixSelectByAreaFraction(pixt2, 0.3, 8, L_SELECT_IF_LT, NULL);
count_pieces(pixd, 0);
pixd = pixSelectByAreaFraction(pixt2, 0.9, 8, L_SELECT_IF_LT, NULL);
count_pieces(pixd, 4);
pixd = pixSelectByAreaFraction(pixt2, 0.5, 8, L_SELECT_IF_GTE, NULL);
count_pieces(pixd, 3);
pixd = pixSelectByAreaFraction(pixt2, 0.7, 8, L_SELECT_IF_GT, NULL);
count_pieces(pixd, 2);
boxat = boxaSelectBySize(boxa, 21, 10, L_SELECT_IF_EITHER,
L_SELECT_IF_LT, NULL);
count_pieces2(boxat, 3);
boxat = boxaSelectBySize(boxa, 22, 32, L_SELECT_IF_BOTH,
L_SELECT_IF_LT, NULL);
count_pieces2(boxat, 2);
boxaDestroy(&boxa);
pixDestroy(&pixt);
pixDestroy(&pixt2);
pixDestroy(&pixs);
/* Here's the most general method for selecting components.
* We do it for area fraction, but any combination of
* size, area/perimeter ratio and area fraction can be used. */
pixs = pixRead("feyn.tif");
/* pixs = pixRead("rabi.png"); */
pixc = pixCopy(NULL, pixs); /* subtract bands from this */
pixt = pixCreateTemplate(pixs); /* add bands to this */
pixGetDimensions(pixs, &w, &h, NULL);
boxa = pixConnComp(pixs, &pixas, 8);
n = boxaGetCount(boxa);
fprintf(stderr, "total: %d\n", n);
na1 = pixaFindAreaFraction(pixas);
nat = numaCreate(0);
numaSetCount(nat, n); /* initialize to all 0 */
sum = sumi = 0;
pixac = pixaCreate(0);
for (i = 0; i < 12; i++) {
/* Compute within the intervals using an intersection. */
na2 = numaMakeThresholdIndicator(na1, edges[i], L_SELECT_IF_GTE);
if (i != 11)
na3 = numaMakeThresholdIndicator(na1, edges[i + 1], L_SELECT_IF_LT);
else
na3 = numaMakeThresholdIndicator(na1, edges[i + 1],
L_SELECT_IF_LTE);
na4 = numaLogicalOp(NULL, na2, na3, L_INTERSECTION);
sum += count_ones(na4, 0, 0, NULL);
/* Compute outside the intervals using a union, and invert */
na2i = numaMakeThresholdIndicator(na1, edges[i], L_SELECT_IF_LT);
if (i != 11)
na3i = numaMakeThresholdIndicator(na1, edges[i + 1],
L_SELECT_IF_GTE);
else
na3i = numaMakeThresholdIndicator(na1, edges[i + 1],
L_SELECT_IF_GT);
na4i = numaLogicalOp(NULL, na3i, na2i, L_UNION);
numaInvert(na4i, na4i);
sumi += count_ones(na4i, 0, 0, NULL);
/* Compare the two methods */
if (sum == sumi)
fprintf(stderr, "\nCorrect: sum = sumi = %d\n", sum);
else
fprintf(stderr, "\nWRONG: sum = %d, sumi = %d\n", sum, sumi);
/* Reconstruct the image, band by band. */
numaLogicalOp(nat, nat, na4, L_UNION);
pixad = pixaSelectWithIndicator(pixas, na4, NULL);
pixd = pixaDisplay(pixad, w, h);
pixOr(pixt, pixt, pixd); /* add them in */
pixcount = pixCopy(NULL, pixt); /* destroyed by count_pieces */
count_ones(na4, band[i], i, "band");
count_pieces(pixd, band[i]);
count_ones(nat, total[i], i, "total");
count_pieces(pixcount, total[i]);
pixaDestroy(&pixad);
/* Remove band successively from full image */
pixRemoveWithIndicator(pixc, pixas, na4);
pixSaveTiled(pixc, pixac, 0.25, 1 - i % 2, 25, 8);
numaDestroy(&na2);
numaDestroy(&na3);
numaDestroy(&na4);
numaDestroy(&na2i);
numaDestroy(&na3i);
numaDestroy(&na4i);
}
/* Did we remove all components from pixc? */
pixZero(pixc, &empty);
if (!empty)
fprintf(stderr, "\nWRONG: not all pixels removed from pixc\n");
pixDestroy(&pixs);
pixDestroy(&pixc);
pixDestroy(&pixt);
boxaDestroy(&boxa);
pixaDestroy(&pixas);
numaDestroy(&na1);
numaDestroy(&nat);
/* One last extraction. Get all components that have either
* a height of at least 50 or a width of between 30 and 35,
* and also have a relatively large perimeter/area ratio. */
pixs = pixRead("feyn.tif");
boxa = pixConnComp(pixs, &pixas, 8);
n = boxaGetCount(boxa);
pixaFindDimensions(pixas, &naw, &nah);
na1 = pixaFindPerimToAreaRatio(pixas);
na2 = numaMakeThresholdIndicator(nah, 50, L_SELECT_IF_GTE);
na3 = numaMakeThresholdIndicator(naw, 30, L_SELECT_IF_GTE);
na4 = numaMakeThresholdIndicator(naw, 35, L_SELECT_IF_LTE);
na5 = numaMakeThresholdIndicator(na1, 0.4, L_SELECT_IF_GTE);
numaLogicalOp(na3, na3, na4, L_INTERSECTION);
numaLogicalOp(na2, na2, na3, L_UNION);
numaLogicalOp(na2, na2, na5, L_INTERSECTION);
numaInvert(na2, na2); /* get components to be removed */
pixRemoveWithIndicator(pixs, pixas, na2);
pixSaveTiled(pixs, pixac, 0.25, 1, 25, 8);
pixDestroy(&pixs);
boxaDestroy(&boxa);
pixaDestroy(&pixas);
numaDestroy(&naw);
numaDestroy(&nah);
numaDestroy(&na1);
numaDestroy(&na2);
numaDestroy(&na3);
numaDestroy(&na4);
numaDestroy(&na5);
pixDisplayMultiple("/tmp/display/file*");
pixd = pixaDisplay(pixac, 0, 0);
pixDisplay(pixd, 100, 100);
pixWrite("/tmp/comp.jpg", pixd, IFF_JFIF_JPEG);
pixDestroy(&pixd);
pixaDestroy(&pixac);
return 0;
}
int main(int argc,
char **argv) {
char *filein, *fileout;
l_int32 i;
l_uint32 val;
l_float32 size;
PIX *pixs, *pixd, *pixm, *pixmi, *pixt1, *pixt2, *pixt3;
static char mainName[] = "seedfilltest";
if (argc != 3)
return ERROR_INT(" Syntax: seedfilltest filein fileout", mainName, 1);
filein = argv[1];
fileout = argv[2];
pixd = NULL;
if ((pixm = pixRead(filein)) == NULL)
return ERROR_INT("pixm not made", mainName, 1);
pixmi = pixInvert(NULL, pixm);
size = pixGetWidth(pixm) * pixGetHeight(pixm);
pixs = pixCreateTemplate(pixm);
for (i = 0; i < 100; i++) {
pixGetPixel(pixm, XS + 5 * i, YS + 5 * i, &val);
if (val == 0) break;
}
if (i == 100)
return ERROR_INT("no seed pixel found", mainName, 1);
pixSetPixel(pixs, XS + 5 * i, YS + 5 * i, 1);
#if 0
/* hole filling; use "hole-filler.png" */
pixt1 = pixHDome(pixmi, 100, 4);
pixt2 = pixThresholdToBinary(pixt1, 10);
/* pixInvert(pixt1, pixt1); */
pixDisplay(pixt1, 100, 500);
pixDisplay(pixt2, 600, 500);
pixt3 = pixHolesByFilling(pixt2, 4);
pixDilateBrick(pixt3, pixt3, 7, 7);
pixd = pixConvertTo8(pixt3, FALSE);
pixDisplay(pixd, 0, 100);
pixSeedfillGray(pixd, pixmi, CONNECTIVITY);
pixInvert(pixd, pixd);
pixDisplay(pixmi, 500, 100);
pixDisplay(pixd, 1000, 100);
pixWrite("/tmp/junkpixm.png", pixmi, IFF_PNG);
pixWrite("/tmp/junkpixd.png", pixd, IFF_PNG);
#endif
#if 0
/* hole filling; use "hole-filler.png" */
pixt1 = pixThresholdToBinary(pixm, 110);
pixInvert(pixt1, pixt1);
pixDisplay(pixt1, 100, 500);
pixt2 = pixHolesByFilling(pixt1, 4);
pixd = pixConvertTo8(pixt2, FALSE);
pixDisplay(pixd, 0, 100);
pixSeedfillGray(pixd, pixmi, CONNECTIVITY);
pixInvert(pixd, pixd);
pixDisplay(pixmi, 500, 100);
pixDisplay(pixd, 1000, 100);
pixWrite("/tmp/junkpixm.png", pixmi, IFF_PNG);
pixWrite("/tmp/junkpixd.png", pixd, IFF_PNG);
#endif
#if 0
/* hole filling; use "hole-filler.png" */
pixd = pixInvert(NULL, pixm);
pixAddConstantGray(pixd, -50);
pixDisplay(pixd, 0, 100);
/* pixt1 = pixThresholdToBinary(pixd, 20);
pixDisplayWithTitle(pixt1, 600, 600, "pixt1", DFLAG); */
pixSeedfillGray(pixd, pixmi, CONNECTIVITY);
/* pixInvert(pixd, pixd); */
pixDisplay(pixmi, 500, 100);
pixDisplay(pixd, 1000, 100);
pixWrite("/tmp/junkpixm.png", pixmi, IFF_PNG);
pixWrite("/tmp/junkpixd.png", pixd, IFF_PNG);
#endif
#if 0
/* test in-place seedfill for speed */
pixd = pixClone(pixs);
startTimer();
pixSeedfillBinary(pixs, pixs, pixmi, CONNECTIVITY);
fprintf(stderr, "Filling rate: %7.4f Mpix/sec\n",
(size/1000000.) / stopTimer());
pixWrite(fileout, pixd, IFF_PNG);
pixOr(pixd, pixd, pixm);
pixWrite("/tmp/junkout1.png", pixd, IFF_PNG);
#endif
#if 0
/* test seedfill to dest for speed */
pixd = pixCreateTemplate(pixm);
startTimer();
for (i = 0; i < NTIMES; i++) {
pixSeedfillBinary(pixd, pixs, pixmi, CONNECTIVITY);
}
fprintf(stderr, "Filling rate: %7.4f Mpix/sec\n",
(size/1000000.) * NTIMES / stopTimer());
pixWrite(fileout, pixd, IFF_PNG);
pixOr(pixd, pixd, pixm);
pixWrite("/tmp/junkout1.png", pixd, IFF_PNG);
#endif
/* use same connectivity to compare with the result of the
* slow parallel operation */
#if 1
pixDestroy(&pixd);
pixd = pixSeedfillMorph(pixs, pixmi, 100, CONNECTIVITY);
pixOr(pixd, pixd, pixm);
pixWrite("/tmp/junkout2.png", pixd, IFF_PNG);
#endif
pixDestroy(&pixs);
pixDestroy(&pixm);
pixDestroy(&pixmi);
pixDestroy(&pixd);
return 0;
}
/*!
* 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);
}
* 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 */
/*!
* 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;
}
/*!
* pixThinGeneral()
*
* Input: pixs (1 bpp)
* type (L_THIN_FG, L_THIN_BG)
* sela (of Sels for parallel composite HMTs)
* maxiters (max number of iters allowed; use 0 to iterate
* until completion)
* Return: pixd, or null on error
*
* Notes:
* (1) See notes in pixThin(). That function chooses among
* the best of the Sels for thinning.
* (2) This is a general function that takes a Sela of HMTs
* that are used in parallel for thinning from each
* of four directions. One iteration consists of four
* such parallel thins.
*/
PIX *
pixThinGeneral(PIX *pixs,
l_int32 type,
SELA *sela,
l_int32 maxiters)
{
l_int32 i, j, r, nsels, same;
PIXA *pixahmt;
PIX **pixhmt; /* array owned by pixahmt; do not destroy! */
PIX *pixd, *pixt;
SEL *sel, *selr;
PROCNAME("pixThinGeneral");
if (!pixs)
return (PIX *)ERROR_PTR("pixs not defined", procName, NULL);
if (pixGetDepth(pixs) != 1)
return (PIX *)ERROR_PTR("pixs not 1 bpp", procName, NULL);
if (type != L_THIN_FG && type != L_THIN_BG)
return (PIX *)ERROR_PTR("invalid fg/bg type", procName, NULL);
if (!sela)
return (PIX *)ERROR_PTR("sela not defined", procName, NULL);
if (maxiters == 0) maxiters = 10000;
/* Set up array of temp pix to hold hmts */
nsels = selaGetCount(sela);
pixahmt = pixaCreate(nsels);
for (i = 0; i < nsels; i++) {
pixt = pixCreateTemplate(pixs);
pixaAddPix(pixahmt, pixt, L_INSERT);
}
pixhmt = pixaGetPixArray(pixahmt);
if (!pixhmt)
return (PIX *)ERROR_PTR("pixhmt array not made", procName, NULL);
#if DEBUG_SELS
pixt = selaDisplayInPix(sela, 35, 3, 15, 4);
pixDisplayWithTitle(pixt, 100, 100, "allsels", 1);
pixDestroy(&pixt);
#endif /* DEBUG_SELS */
/* Set up initial image for fg thinning */
if (type == L_THIN_FG)
pixd = pixCopy(NULL, pixs);
else /* bg thinning */
pixd = pixInvert(NULL, pixs);
/* Thin the fg, with up to maxiters iterations */
for (i = 0; i < maxiters; i++) {
pixt = pixCopy(NULL, pixd); /* test for completion */
for (r = 0; r < 4; r++) { /* over 90 degree rotations of Sels */
for (j = 0; j < nsels; j++) { /* over individual sels in sela */
sel = selaGetSel(sela, j); /* not a copy */
selr = selRotateOrth(sel, r);
pixHMT(pixhmt[j], pixd, selr);
selDestroy(&selr);
if (j > 0)
pixOr(pixhmt[0], pixhmt[0], pixhmt[j]); /* accum result */
}
pixSubtract(pixd, pixd, pixhmt[0]); /* remove result */
}
pixEqual(pixd, pixt, &same);
pixDestroy(&pixt);
if (same) {
L_INFO("%d iterations to completion\n", procName, i);
break;
}
}
if (type == L_THIN_BG)
pixInvert(pixd, pixd);
pixaDestroy(&pixahmt);
return pixd;
}
/*!
* \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;
}
