/*! * \brief pixGetWordsInTextlines() * * \param[in] pixs 1 bpp, typ. 75 - 150 ppi * \param[in] minwidth, minheight of saved components; smaller are discarded * \param[in] maxwidth, maxheight of saved components; larger are discarded * \param[out] pboxad word boxes sorted in textline line order * \param[out] ppixad word images sorted in textline line order * \param[out] pnai index of textline for each word * \return 0 if OK, 1 on error * * <pre> * Notes: * (1) The input should be at a resolution of between 75 and 150 ppi. * (2) The four size constraints on saved components are all * scaled by %reduction. * (3) The result are word images (and their b.b.), extracted in * textline order, at either full res or 2x reduction, * and with a numa giving the textline index for each word. * (4) The pixa and boxa interfaces should make this type of * application simple to put together. The steps are: * ~ generate first estimate of word masks * ~ get b.b. of these, and remove the small and big ones * ~ extract pixa of the word images, using the b.b. * ~ sort actual word images in textline order (2d) * ~ flatten them to a pixa (1d), saving the textline index * for each pix * (5) In an actual application, it may be desirable to pre-filter * the input image to remove large components, to extract * single columns of text, and to deskew them. For example, * to remove both large components and small noisy components * that can interfere with the statistics used to estimate * parameters for segmenting by words, but still retain text lines, * the following image preprocessing can be done: * Pix *pixt = pixMorphSequence(pixs, "c40.1", 0); * Pix *pixf = pixSelectBySize(pixt, 0, 60, 8, * L_SELECT_HEIGHT, L_SELECT_IF_LT, NULL); * pixAnd(pixf, pixf, pixs); // the filtered image * The closing turns text lines into long blobs, but does not * significantly increase their height. But if there are many * small connected components in a dense texture, this is likely * to generate tall components that will be eliminated in pixf. * </pre> */ l_int32 pixGetWordsInTextlines(PIX *pixs, l_int32 minwidth, l_int32 minheight, l_int32 maxwidth, l_int32 maxheight, BOXA **pboxad, PIXA **ppixad, NUMA **pnai) { BOXA *boxa1, *boxad; BOXAA *baa; NUMA *nai; NUMAA *naa; PIXA *pixa1, *pixad; PIXAA *paa; PROCNAME("pixGetWordsInTextlines"); if (!pboxad || !ppixad || !pnai) return ERROR_INT("&boxad, &pixad, &nai not all defined", procName, 1); *pboxad = NULL; *ppixad = NULL; *pnai = NULL; if (!pixs) return ERROR_INT("pixs not defined", procName, 1); /* Get the bounding boxes of the words from the word mask. */ pixWordBoxesByDilation(pixs, minwidth, minheight, maxwidth, maxheight, &boxa1, NULL, NULL); /* Generate a pixa of the word images */ pixa1 = pixaCreateFromBoxa(pixs, boxa1, NULL); /* mask over each word */ /* Sort the bounding boxes of these words by line. We use the * index mapping to allow identical sorting of the pixa. */ baa = boxaSort2d(boxa1, &naa, -1, -1, 4); paa = pixaSort2dByIndex(pixa1, naa, L_CLONE); /* Flatten the word paa */ pixad = pixaaFlattenToPixa(paa, &nai, L_CLONE); boxad = pixaGetBoxa(pixad, L_COPY); *pnai = nai; *pboxad = boxad; *ppixad = pixad; pixaDestroy(&pixa1); boxaDestroy(&boxa1); boxaaDestroy(&baa); pixaaDestroy(&paa); numaaDestroy(&naa); return 0; }
l_int32 main(int argc, char **argv) { char *boxatxt; l_int32 i; BOXA *boxa1, *boxa2, *boxa3; BOXAA *baa, *baa1; NUMAA *naa1; PIX *pixdb, *pix1, *pix2, *pix3, *pix4; PIXA *pixa1, *pixa2, *pixa3, *pixat; L_RECOG *recog; L_RECOGA *recoga; SARRAY *sa1; /* ----- Example identifying samples using training data ----- */ #if 1 /* Read the training data */ pixat = pixaRead("recog/sets/train06.pa"); recog = recogCreateFromPixa(pixat, 0, 0, L_USE_ALL, 128, 1); recoga = recogaCreateFromRecog(recog); pixaDestroy(&pixat); /* Read the data from all samples */ pix1 = pixRead("recog/sets/samples06.png"); boxatxt = pixGetText(pix1); boxa1 = boxaReadMem((l_uint8 *)boxatxt, strlen(boxatxt)); pixa1 = pixaCreateFromBoxa(pix1, boxa1, NULL); pixDestroy(&pix1); /* destroys boxa1 */ /* Identify components in the sample data */ pixa2 = pixaCreate(0); pixa3 = pixaCreate(0); for (i = 0; i < 9; i++) { /* if (i != 4) continue; */ /* dots form separate boxa */ /* if (i != 8) continue; */ /* broken 2 in '24' */ pix1 = pixaGetPix(pixa1, i, L_CLONE); /* Show the 2d box data in the sample */ boxa2 = pixConnComp(pix1, NULL, 8); baa = boxaSort2d(boxa2, NULL, 6, 6, 5); pix2 = boxaaDisplay(baa, 3, 1, 0xff000000, 0x00ff0000, 0, 0); pixaAddPix(pixa3, pix2, L_INSERT); boxaaDestroy(&baa); boxaDestroy(&boxa2); /* Get the numbers in the sample */ recogaIdentifyMultiple(recoga, pix1, 0, 5, 3, &boxa3, NULL, &pixdb, 0); sa1 = recogaExtractNumbers(recoga, boxa3, 0.7, -1, &baa1, &naa1); sarrayWriteStream(stderr, sa1); boxaaWriteStream(stderr, baa1); numaaWriteStream(stderr, naa1); pixaAddPix(pixa2, pixdb, L_INSERT); /* pixaWrite("/tmp/pixa.pa", pixa2); */ pixDestroy(&pix1); boxaDestroy(&boxa3); boxaaDestroy(&baa1); numaaDestroy(&naa1); sarrayDestroy(&sa1); } pix3 = pixaDisplayLinearly(pixa2, L_VERT, 1.0, 0, 20, 1, NULL); pixWrite("/tmp/pix3.png", pix3, IFF_PNG); pix4 = pixaDisplayTiledInRows(pixa3, 32, 1500, 1.0, 0, 20, 2); pixDisplay(pix4, 500, 0); pixWrite("/tmp/pix4.png", pix4, IFF_PNG); pixaDestroy(&pixa2); pixaDestroy(&pixa3); pixDestroy(&pix1); pixDestroy(&pix3); pixDestroy(&pix4); pixaDestroy(&pixa1); boxaDestroy(&boxa1); recogaDestroy(&recoga); #endif return 0; }
/*! * \brief pixGetWordsInTextlines() * * \param[in] pixs 1 bpp, typ. 300 ppi * \param[in] reduction 1 for input res; 2 for 2x reduction of input res * \param[in] minwidth, minheight of saved components; smaller are discarded * \param[in] maxwidth, maxheight of saved components; larger are discarded * \param[out] pboxad word boxes sorted in textline line order * \param[out] ppixad word images sorted in textline line order * \param[out] pnai index of textline for each word * \return 0 if OK, 1 on error * * <pre> * Notes: * (1) The input should be at a resolution of about 300 ppi. * The word masks and word images can be computed at either * 150 ppi or 300 ppi. For the former, set reduction = 2. * (2) The four size constraints on saved components are all * scaled by %reduction. * (3) The result are word images (and their b.b.), extracted in * textline order, at either full res or 2x reduction, * and with a numa giving the textline index for each word. * (4) The pixa and boxa interfaces should make this type of * application simple to put together. The steps are: * ~ optionally reduce by 2x * ~ generate first estimate of word masks * ~ get b.b. of these, and remove the small and big ones * ~ extract pixa of the word images, using the b.b. * ~ sort actual word images in textline order (2d) * ~ flatten them to a pixa (1d), saving the textline index * for each pix * (5) In an actual application, it may be desirable to pre-filter * the input image to remove large components, to extract * single columns of text, and to deskew them. For example, * to remove both large components and small noisy components * that can interfere with the statistics used to estimate * parameters for segmenting by words, but still retain text lines, * the following image preprocessing can be done: * Pix *pixt = pixMorphSequence(pixs, "c40.1", 0); * Pix *pixf = pixSelectBySize(pixt, 0, 60, 8, * L_SELECT_HEIGHT, L_SELECT_IF_LT, NULL); * pixAnd(pixf, pixf, pixs); // the filtered image * The closing turns text lines into long blobs, but does not * significantly increase their height. But if there are many * small connected components in a dense texture, this is likely * to generate tall components that will be eliminated in pixf. * </pre> */ l_int32 pixGetWordsInTextlines(PIX *pixs, l_int32 reduction, l_int32 minwidth, l_int32 minheight, l_int32 maxwidth, l_int32 maxheight, BOXA **pboxad, PIXA **ppixad, NUMA **pnai) { l_int32 maxdil; BOXA *boxa1, *boxad; BOXAA *baa; NUMA *nai; NUMAA *naa; PIXA *pixa1, *pixad; PIX *pix1; PIXAA *paa; PROCNAME("pixGetWordsInTextlines"); if (!pboxad || !ppixad || !pnai) return ERROR_INT("&boxad, &pixad, &nai not all defined", procName, 1); *pboxad = NULL; *ppixad = NULL; *pnai = NULL; if (!pixs) return ERROR_INT("pixs not defined", procName, 1); if (reduction != 1 && reduction != 2) return ERROR_INT("reduction not in {1,2}", procName, 1); if (reduction == 1) { pix1 = pixClone(pixs); maxdil = 18; } else { /* reduction == 2 */ pix1 = pixReduceRankBinaryCascade(pixs, 1, 0, 0, 0); maxdil = 9; } /* Get the bounding boxes of the words from the word mask. */ pixWordBoxesByDilation(pix1, maxdil, minwidth, minheight, maxwidth, maxheight, &boxa1, NULL); /* Generate a pixa of the word images */ pixa1 = pixaCreateFromBoxa(pix1, boxa1, NULL); /* mask over each word */ /* Sort the bounding boxes of these words by line. We use the * index mapping to allow identical sorting of the pixa. */ baa = boxaSort2d(boxa1, &naa, -1, -1, 4); paa = pixaSort2dByIndex(pixa1, naa, L_CLONE); /* Flatten the word paa */ pixad = pixaaFlattenToPixa(paa, &nai, L_CLONE); boxad = pixaGetBoxa(pixad, L_COPY); *pnai = nai; *pboxad = boxad; *ppixad = pixad; pixDestroy(&pix1); pixaDestroy(&pixa1); boxaDestroy(&boxa1); boxaaDestroy(&baa); pixaaDestroy(&paa); numaaDestroy(&naa); return 0; }