/*!
* recogCreate()
*
* Input: scalew (scale all widths to this; use 0 for no scaling)
* scaleh (scale all heights to this; use 0 for no scaling)
* templ_type (L_USE_AVERAGE or L_USE_ALL)
* threshold (for binarization; typically ~128)
* maxyshift (from nominal centroid alignment; typically 0 or 1)
* Return: recog, or null on error
*
* Notes:
* (1) For a set trained on one font, such as numbers in a book,
* it is sensible to set scalew = scaleh = 0.
* (2) For a mixed training set, scaling to a fixed height,
* such as 32 pixels, but leaving the width unscaled, is effective.
* (3) The storage for most of the arrays is allocated when training
* is finished.
*/
L_RECOG *
recogCreate(l_int32 scalew,
l_int32 scaleh,
l_int32 templ_type,
l_int32 threshold,
l_int32 maxyshift)
{
L_RECOG *recog;
PIXA *pixa;
PIXAA *paa;
PROCNAME("recogCreate");
if (scalew < 0 || scaleh < 0)
return (L_RECOG *)ERROR_PTR("invalid scalew or scaleh", procName, NULL);
if (templ_type != L_USE_AVERAGE && templ_type != L_USE_ALL)
return (L_RECOG *)ERROR_PTR("invalid templ_type flag", procName, NULL);
if (threshold < 1 || threshold > 255)
return (L_RECOG *)ERROR_PTR("invalid threshold", procName, NULL);
if ((recog = (L_RECOG *)CALLOC(1, sizeof(L_RECOG))) == NULL)
return (L_RECOG *)ERROR_PTR("rec not made", procName, NULL);
recog->templ_type = templ_type;
recog->threshold = threshold;
recog->scalew = scalew;
recog->scaleh = scaleh;
recog->maxyshift = maxyshift;
recog->asperity_fr = DEFAULT_ASPERITY_FRACT;
recogSetPadParams(recog, NULL, NULL, NULL, -1, -1, -1);
recog->bmf = bmfCreate(NULL, 6);
recog->bmf_size = 6;
recog->maxarraysize = MAX_EXAMPLES_IN_CLASS;
recog->index = -1;
/* Generate the LUTs */
recog->centtab = makePixelCentroidTab8();
recog->sumtab = makePixelSumTab8();
recog->sa_text = sarrayCreate(0);
recog->dna_tochar = l_dnaCreate(0);
/* Input default values for min component size for splitting.
* These are overwritten when pixTrainingFinished() is called. */
recog->min_splitw = 6;
recog->min_splith = 6;
recog->max_splith = 60;
/* Generate the storage for the unscaled training bitmaps */
paa = pixaaCreate(recog->maxarraysize);
pixa = pixaCreate(1);
pixaaInitFull(paa, pixa);
pixaDestroy(&pixa);
recog->pixaa_u = paa;
/* Generate the storage for debugging */
recog->pixadb_boot = pixaCreate(2);
recog->pixadb_split = pixaCreate(2);
return recog;
}
/*!
* \brief recogCreate()
*
* \param[in] scalew scale all widths to this; use 0 otherwise
* \param[in] scaleh scale all heights to this; use 0 otherwise
* \param[in] linew width of normalized strokes; use 0 to skip
* \param[in] threshold for binarization; typically ~128; 0 for default
* \param[in] maxyshift from nominal centroid alignment; default is 1
* \return recog, or NULL on error
*
* <pre>
* Notes:
* (1) If %scalew == 0 and %scaleh == 0, no scaling is done.
* If one of these is 0 and the other is > 0, scaling is isotropic
* to the requested size. We typically do not set both > 0.
* (2) Use linew > 0 to convert the templates to images with fixed
* width strokes. linew == 0 skips the conversion.
* (3) The only valid values for %maxyshift are 0, 1 and 2.
* It is recommended to use %maxyshift == 1 (default value).
* Using %maxyshift == 0 is much faster than %maxyshift == 1, but
* it is much less likely to find the template with the best
* correlation. Use of anything but 1 results in a warning.
* (4) Scaling is used for finding outliers and for training a
* book-adapted recognizer (BAR) from a bootstrap recognizer (BSR).
* Scaling the height to a fixed value and scaling the width
* accordingly (e.g., %scaleh = 40, %scalew = 0) is recommended.
* (5) The storage for most of the arrays is allocated when training
* is finished.
* </pre>
*/
L_RECOG *
recogCreate(l_int32 scalew,
l_int32 scaleh,
l_int32 linew,
l_int32 threshold,
l_int32 maxyshift)
{
L_RECOG *recog;
PROCNAME("recogCreate");
if (scalew < 0 || scaleh < 0)
return (L_RECOG *)ERROR_PTR("invalid scalew or scaleh", procName, NULL);
if (linew > 10)
return (L_RECOG *)ERROR_PTR("invalid linew > 10", procName, NULL);
if (threshold == 0) threshold = DEFAULT_THRESHOLD;
if (threshold < 0 || threshold > 255) {
L_WARNING("invalid threshold; using default\n", procName);
threshold = DEFAULT_THRESHOLD;
}
if (maxyshift < 0 || maxyshift > 2) {
L_WARNING("invalid maxyshift; using default value\n", procName);
maxyshift = DEFAULT_MAXYSHIFT;
} else if (maxyshift == 0) {
L_WARNING("Using maxyshift = 0; faster, worse correlation results\n",
procName);
} else if (maxyshift == 2) {
L_WARNING("Using maxyshift = 2; slower\n", procName);
}
if ((recog = (L_RECOG *)LEPT_CALLOC(1, sizeof(L_RECOG))) == NULL)
return (L_RECOG *)ERROR_PTR("rec not made", procName, NULL);
recog->templ_use = L_USE_ALL_TEMPLATES; /* default */
recog->threshold = threshold;
recog->scalew = scalew;
recog->scaleh = scaleh;
recog->linew = linew;
recog->maxyshift = maxyshift;
recogSetParams(recog, 1, -1, -1.0, -1.0);
recog->bmf = bmfCreate(NULL, 6);
recog->bmf_size = 6;
recog->maxarraysize = MAX_EXAMPLES_IN_CLASS;
/* Generate the LUTs */
recog->centtab = makePixelCentroidTab8();
recog->sumtab = makePixelSumTab8();
recog->sa_text = sarrayCreate(0);
recog->dna_tochar = l_dnaCreate(0);
/* Input default values for min component size for splitting.
* These are overwritten when pixTrainingFinished() is called. */
recog->min_splitw = 6;
recog->max_splith = 60;
/* Allocate the paa for the unscaled training bitmaps */
recog->pixaa_u = pixaaCreate(recog->maxarraysize);
/* Generate the storage for debugging */
recog->pixadb_boot = pixaCreate(2);
recog->pixadb_split = pixaCreate(2);
return recog;
}
l_int32 main(int argc,
char **argv)
{
char buf[512];
l_int32 delx, dely, etransx, etransy, w, h, area1, area2;
l_int32 *stab, *ctab;
l_float32 cx1, cy1, cx2, cy2, score;
PIX *pix0, *pix1, *pix2;
L_REGPARAMS *rp;
if (regTestSetup(argc, argv, &rp))
return 1;
/* ------------ Test of pixBestCorrelation() --------------- */
pix0 = pixRead("harmoniam100-11.png");
pix1 = pixConvertTo1(pix0, 160);
pixGetDimensions(pix1, &w, &h, NULL);
/* Now make a smaller image, translated by (-32, -12)
* Except for the resizing, this is equivalent to
* pix2 = pixTranslate(NULL, pix1, -32, -12, L_BRING_IN_WHITE); */
pix2 = pixCreate(w - 10, h, 1);
pixRasterop(pix2, 0, 0, w, h, PIX_SRC, pix1, 32, 12);
/* Get the number of FG pixels and the centroid locations */
stab = makePixelSumTab8();
ctab = makePixelCentroidTab8();
pixCountPixels(pix1, &area1, stab);
pixCountPixels(pix2, &area2, stab);
pixCentroid(pix1, ctab, stab, &cx1, &cy1);
pixCentroid(pix2, ctab, stab, &cx2, &cy2);
etransx = lept_roundftoi(cx1 - cx2);
etransy = lept_roundftoi(cy1 - cy2);
fprintf(stderr, "delta cx = %d, delta cy = %d\n",
etransx, etransy);
/* Get the best correlation, searching around the translation
* where the centroids coincide */
pixBestCorrelation(pix1, pix2, area1, area2, etransx, etransy,
4, stab, &delx, &dely, &score, 5);
fprintf(stderr, "delx = %d, dely = %d, score = %7.4f\n",
delx, dely, score);
regTestCompareValues(rp, 32, delx, 0); /* 0 */
regTestCompareValues(rp, 12, dely, 0); /* 1 */
regTestCheckFile(rp, "/tmp/junkcorrel_5.png"); /* 2 */
lept_rm(NULL, "junkcorrel_5.png");
FREE(stab);
FREE(ctab);
pixDestroy(&pix0);
pixDestroy(&pix1);
pixDestroy(&pix2);
/* ------------ Test of pixCompareWithTranslation() ------------ */
/* Now use the pyramid to get the result. Do a translation
* to remove pixels at the bottom from pix2, so that the
* centroids are initially far apart. */
pix1 = pixRead("harmoniam-11.tif");
pix2 = pixTranslate(NULL, pix1, -45, 25, L_BRING_IN_WHITE);
l_pdfSetDateAndVersion(0);
pixCompareWithTranslation(pix1, pix2, 160, &delx, &dely, &score, 1);
pixDestroy(&pix1);
pixDestroy(&pix2);
fprintf(stderr, "delx = %d, dely = %d\n", delx, dely);
regTestCompareValues(rp, 45, delx, 0); /* 3 */
regTestCompareValues(rp, -25, dely, 0); /* 4 */
regTestCheckFile(rp, "/tmp/junkcmp.pdf"); /* 5 */
regTestCheckFile(rp, "/tmp/junkcorrel.pdf"); /* 6 */
return regTestCleanup(rp);
}
l_int32 main(int argc,
char **argv)
{
l_int32 delx, dely, etransx, etransy, w, h, area1, area2;
l_int32 *stab, *ctab;
l_float32 cx1, cy1, cx2, cy2, score, fract;
PIX *pix0, *pix1, *pix2, *pix3, *pix4, *pix5;
L_REGPARAMS *rp;
if (regTestSetup(argc, argv, &rp))
return 1;
/* ------------ Test of pixBestCorrelation() --------------- */
pix0 = pixRead("harmoniam100-11.png");
pix1 = pixConvertTo1(pix0, 160);
pixGetDimensions(pix1, &w, &h, NULL);
/* Now make a smaller image, translated by (-32, -12)
* Except for the resizing, this is equivalent to
* pix2 = pixTranslate(NULL, pix1, -32, -12, L_BRING_IN_WHITE); */
pix2 = pixCreate(w - 10, h, 1);
pixRasterop(pix2, 0, 0, w, h, PIX_SRC, pix1, 32, 12);
/* Get the number of FG pixels and the centroid locations */
stab = makePixelSumTab8();
ctab = makePixelCentroidTab8();
pixCountPixels(pix1, &area1, stab);
pixCountPixels(pix2, &area2, stab);
pixCentroid(pix1, ctab, stab, &cx1, &cy1);
pixCentroid(pix2, ctab, stab, &cx2, &cy2);
etransx = lept_roundftoi(cx1 - cx2);
etransy = lept_roundftoi(cy1 - cy2);
fprintf(stderr, "delta cx = %d, delta cy = %d\n",
etransx, etransy);
/* Get the best correlation, searching around the translation
* where the centroids coincide */
pixBestCorrelation(pix1, pix2, area1, area2, etransx, etransy,
4, stab, &delx, &dely, &score, 5);
fprintf(stderr, "delx = %d, dely = %d, score = %7.4f\n",
delx, dely, score);
regTestCompareValues(rp, 32, delx, 0); /* 0 */
regTestCompareValues(rp, 12, dely, 0); /* 1 */
lept_mv("/tmp/lept/correl_5.png", "regout", NULL, NULL);
regTestCheckFile(rp, "/tmp/regout/correl_5.png"); /* 2 */
lept_free(stab);
lept_free(ctab);
pixDestroy(&pix0);
pixDestroy(&pix1);
pixDestroy(&pix2);
/* ------------ Test of pixCompareWithTranslation() ------------ */
/* Now use the pyramid to get the result. Do a translation
* to remove pixels at the bottom from pix2, so that the
* centroids are initially far apart. */
pix1 = pixRead("harmoniam-11.tif");
pix2 = pixTranslate(NULL, pix1, -45, 25, L_BRING_IN_WHITE);
l_pdfSetDateAndVersion(0);
pixCompareWithTranslation(pix1, pix2, 160, &delx, &dely, &score, 1);
pixDestroy(&pix1);
pixDestroy(&pix2);
fprintf(stderr, "delx = %d, dely = %d\n", delx, dely);
regTestCompareValues(rp, 45, delx, 0); /* 3 */
regTestCompareValues(rp, -25, dely, 0); /* 4 */
lept_mv("/tmp/lept/correl.pdf", "regout", NULL, NULL);
lept_mv("/tmp/lept/compare.pdf", "regout", NULL, NULL);
regTestCheckFile(rp, "/tmp/regout/compare.pdf"); /* 5 */
regTestCheckFile(rp, "/tmp/regout/correl.pdf"); /* 6 */
/* ------------ Test of pixGetPerceptualDiff() --------------- */
pix0 = pixRead("greencover.jpg");
pix1 = pixRead("redcover.jpg"); /* pre-scaled to the same size */
/* Apply directly to the color images */
pixGetPerceptualDiff(pix0, pix1, 1, 3, 20, &fract, &pix2, &pix3);
fprintf(stderr, "Fraction of color pixels = %f\n", fract);
regTestCompareValues(rp, 0.061252, fract, 0.01); /* 7 */
regTestWritePixAndCheck(rp, pix2, IFF_JFIF_JPEG); /* 8 */
regTestWritePixAndCheck(rp, pix3, IFF_TIFF_G4); /* 9 */
pixDestroy(&pix2);
pixDestroy(&pix3);
/* Apply to grayscale images */
pix2 = pixConvertTo8(pix0, 0);
pix3 = pixConvertTo8(pix1, 0);
pixGetPerceptualDiff(pix2, pix3, 1, 3, 20, &fract, &pix4, &pix5);
fprintf(stderr, "Fraction of grayscale pixels = %f\n", fract);
regTestCompareValues(rp, 0.046928, fract, 0.0002); /* 10 */
regTestWritePixAndCheck(rp, pix4, IFF_JFIF_JPEG); /* 11 */
regTestWritePixAndCheck(rp, pix5, IFF_TIFF_G4); /* 12 */
pixDestroy(&pix0);
pixDestroy(&pix1);
pixDestroy(&pix2);
pixDestroy(&pix3);
pixDestroy(&pix4);
pixDestroy(&pix5);
return regTestCleanup(rp);
}
