#endif /* HAVE_CONFIG_H */
#include "allheaders.h"
/* Maximum allowed iterations in Phase 1. */
static const l_int32 MAX_ALLOWED_ITERATIONS = 20;
/* Factor by which max dist is increased on each iteration */
static const l_float32 DIST_EXPAND_FACT = 1.3f;
/* Octcube division level for computing nearest colormap color using LUT.
* Using 4 should suffice for up to 50 - 100 colors, and it is
* very fast. Using 5 takes 8 times as long to set up the LUT
* for little perceptual gain, even with 100 colors. */
static const l_int32 LEVEL_IN_OCTCUBE = 4;
static l_int32 pixColorSegmentTryCluster(PIX *pixd, PIX *pixs,
l_int32 maxdist, l_int32 maxcolors,
l_int32 debugflag);
/*------------------------------------------------------------------*
* Unsupervised color segmentation *
*------------------------------------------------------------------*/
/*!
* \brief pixColorSegment()
*
* \param[in] pixs 32 bpp; 24-bit color
* \param[in] maxdist max euclidean dist to existing cluster
* \param[in] maxcolors max number of colors allowed in first pass
* \param[in] selsize linear size of sel for closing to remove noise
* \param[in] finalcolors max number of final colors allowed after 4th pass
* \param[in] debugflag 1 for debug output; 0 otherwise
* \return pixd 8 bit with colormap, or NULL on error
*
*
* Color segmentation proceeds in four phases:
*
* Phase 1: pixColorSegmentCluster()
* The image is traversed in raster order. Each pixel either
* becomes the representative for a new cluster or is assigned to an
* existing cluster. Assignment is greedy. The data is stored in
* a colormapped image. Three auxiliary arrays are used to hold
* the colors of the representative pixels, for fast lookup.
* The average color in each cluster is computed.
*
* Phase 2. pixAssignToNearestColor()
* A second non-greedy clustering pass is performed, where each pixel
* is assigned to the nearest cluster average. We also keep track
* of how many pixels are assigned to each cluster.
*
* Phase 3. pixColorSegmentClean()
* For each cluster, starting with the largest, do a morphological
* closing to eliminate small components within larger ones.
*
* Phase 4. pixColorSegmentRemoveColors()
* Eliminate all colors except the most populated 'finalcolors'.
* Then remove unused colors from the colormap, and reassign those
* pixels to the nearest remaining cluster, using the original pixel values.
*
* Notes:
* (1) The goal is to generate a small number of colors.
* Typically this would be specified by 'finalcolors',
* a number that would be somewhere between 3 and 6.
* The parameter 'maxcolors' specifies the maximum number of
* colors generated in the first phase. This should be
* larger than finalcolors, perhaps twice as large.
* If more than 'maxcolors' are generated in the first phase
* using the input 'maxdist', the distance is repeatedly
* increased by a multiplicative factor until the condition
* is satisfied. The implicit relation between 'maxdist'
* and 'maxcolors' is thus adjusted programmatically.
* (2) As a very rough guideline, given a target value of 'finalcolors',
* here are approximate values of 'maxdist' and 'maxcolors'
* to start with:
*
* finalcolors maxcolors maxdist
* ----------- --------- -------
* 3 6 100
* 4 8 90
* 5 10 75
* 6 12 60
*
* For a given number of finalcolors, if you use too many
* maxcolors, the result will be noisy. If you use too few,
* the result will be a relatively poor assignment of colors.
*
*/
PIX *
pixColorSegment(PIX *pixs,
l_int32 maxdist,
l_int32 maxcolors,
l_int32 selsize,
l_int32 finalcolors,
l_int32 debugflag)
{
l_int32 *countarray;
PIX *pixd;
PROCNAME("pixColorSegment");
if (!pixs)
return (PIX *)ERROR_PTR("pixs not defined", procName, NULL);
if (pixGetDepth(pixs) != 32)
return (PIX *)ERROR_PTR("must be rgb color", procName, NULL);
/* Phase 1; original segmentation */
pixd = pixColorSegmentCluster(pixs, maxdist, maxcolors, debugflag);
if (!pixd)
return (PIX *)ERROR_PTR("pixd not made", procName, NULL);
if (debugflag) {
lept_mkdir("lept/segment");
pixWriteDebug("/tmp/lept/segment/colorseg1.png", pixd, IFF_PNG);
}
/* Phase 2; refinement in pixel assignment */
if ((countarray = (l_int32 *)LEPT_CALLOC(256, sizeof(l_int32))) == NULL) {
pixDestroy(&pixd);
return (PIX *)ERROR_PTR("countarray not made", procName, NULL);
}
pixAssignToNearestColor(pixd, pixs, NULL, LEVEL_IN_OCTCUBE, countarray);
if (debugflag)
pixWriteDebug("/tmp/lept/segment/colorseg2.png", pixd, IFF_PNG);
/* Phase 3: noise removal by separately closing each color */
pixColorSegmentClean(pixd, selsize, countarray);
LEPT_FREE(countarray);
if (debugflag)
pixWriteDebug("/tmp/lept/segment/colorseg3.png", pixd, IFF_PNG);
/* Phase 4: removal of colors with small population and
* reassignment of pixels to remaining colors */
pixColorSegmentRemoveColors(pixd, pixs, finalcolors);
return pixd;
}
/*!
* \brief pixColorSegmentCluster()
*
* \param[in] pixs 32 bpp; 24-bit color
* \param[in] maxdist max euclidean dist to existing cluster
* \param[in] maxcolors max number of colors allowed in first pass
* \param[in] debugflag 1 for debug output; 0 otherwise
* \return pixd 8 bit with colormap, or NULL on error
*
*
* Notes:
* (1) This is phase 1. See description in pixColorSegment().
* (2) Greedy unsupervised classification. If the limit 'maxcolors'
* is exceeded, the computation is repeated with a larger
* allowed cluster size.
* (3) On each successive iteration, 'maxdist' is increased by a
* constant factor. See comments in pixColorSegment() for
* a guideline on parameter selection.
* Note that the diagonal of the 8-bit rgb color cube is about
* 440, so for 'maxdist' = 440, you are guaranteed to get 1 color!
*
*/
PIX *
pixColorSegmentCluster(PIX *pixs,
l_int32 maxdist,
l_int32 maxcolors,
l_int32 debugflag)
{
l_int32 w, h, newmaxdist, ret, niters, ncolors, success;
PIX *pixd;
PIXCMAP *cmap;
PROCNAME("pixColorSegmentCluster");
if (!pixs)
return (PIX *)ERROR_PTR("pixs not defined", procName, NULL);
if (pixGetDepth(pixs) != 32)
return (PIX *)ERROR_PTR("must be rgb color", procName, NULL);
pixGetDimensions(pixs, &w, &h, NULL);
if ((pixd = pixCreate(w, h, 8)) == NULL)
return (PIX *)ERROR_PTR("pixd not made", procName, NULL);
cmap = pixcmapCreate(8);
pixSetColormap(pixd, cmap);
pixCopyResolution(pixd, pixs);
newmaxdist = maxdist;
niters = 0;
success = TRUE;
while (1) {
ret = pixColorSegmentTryCluster(pixd, pixs, newmaxdist,
maxcolors, debugflag);
niters++;
if (!ret) {
ncolors = pixcmapGetCount(cmap);
if (debugflag)
L_INFO("Success with %d colors after %d iters\n", procName,
ncolors, niters);
break;
}
if (niters == MAX_ALLOWED_ITERATIONS) {
L_WARNING("too many iters; newmaxdist = %d\n",
procName, newmaxdist);
success = FALSE;
break;
}
newmaxdist = (l_int32)(DIST_EXPAND_FACT * (l_float32)newmaxdist);
}
if (!success) {
pixDestroy(&pixd);
return (PIX *)ERROR_PTR("failure in phase 1", procName, NULL);
}
return pixd;
}
/*!
* \brief pixColorSegmentTryCluster()
*
* \param[in] pixd
* \param[in] pixs
* \param[in] maxdist
* \param[in] maxcolors
* \param[in] debugflag 1 for debug output; 0 otherwise
* \return 0 if OK, 1 on error
*
*
* Notes:
* This function should only be called from pixColorSegCluster()
*
*/
static l_int32
pixColorSegmentTryCluster(PIX *pixd,
PIX *pixs,
l_int32 maxdist,
l_int32 maxcolors,
l_int32 debugflag)
{
l_int32 rmap[256], gmap[256], bmap[256];
l_int32 w, h, wpls, wpld, i, j, k, found, ret, index, ncolors;
l_int32 rval, gval, bval, dist2, maxdist2;
l_int32 countarray[256];
l_int32 rsum[256], gsum[256], bsum[256];
l_uint32 *ppixel;
l_uint32 *datas, *datad, *lines, *lined;
PIXCMAP *cmap;
PROCNAME("pixColorSegmentTryCluster");
if (!pixs)
return ERROR_INT("pixs not defined", procName, 1);
if (!pixd)
return ERROR_INT("pixd not defined", procName, 1);
w = pixGetWidth(pixs);
h = pixGetHeight(pixs);
maxdist2 = maxdist * maxdist;
cmap = pixGetColormap(pixd);
pixcmapClear(cmap);
for (k = 0; k < 256; k++) {
rsum[k] = gsum[k] = bsum[k] = 0;
rmap[k] = gmap[k] = bmap[k] = 0;
}
datas = pixGetData(pixs);
datad = pixGetData(pixd);
wpls = pixGetWpl(pixs);
wpld = pixGetWpl(pixd);
ncolors = 0;
for (i = 0; i < h; i++) {
lines = datas + i * wpls;
lined = datad + i * wpld;
for (j = 0; j < w; j++) {
ppixel = lines + j;
rval = GET_DATA_BYTE(ppixel, COLOR_RED);
gval = GET_DATA_BYTE(ppixel, COLOR_GREEN);
bval = GET_DATA_BYTE(ppixel, COLOR_BLUE);
ncolors = pixcmapGetCount(cmap);
found = FALSE;
for (k = 0; k < ncolors; k++) {
dist2 = (rval - rmap[k]) * (rval - rmap[k]) +
(gval - gmap[k]) * (gval - gmap[k]) +
(bval - bmap[k]) * (bval - bmap[k]);
if (dist2 <= maxdist2) { /* take it; greedy */
found = TRUE;
SET_DATA_BYTE(lined, j, k);
countarray[k]++;
rsum[k] += rval;
gsum[k] += gval;
bsum[k] += bval;
break;
}
}
if (!found) { /* Add a new color */
ret = pixcmapAddNewColor(cmap, rval, gval, bval, &index);
/* lept_stderr(
"index = %d, (i,j) = (%d,%d), rgb = (%d, %d, %d)\n",
index, i, j, rval, gval, bval); */
if (ret == 0 && index < maxcolors) {
countarray[index] = 1;
SET_DATA_BYTE(lined, j, index);
rmap[index] = rval;
gmap[index] = gval;
bmap[index] = bval;
rsum[index] = rval;
gsum[index] = gval;
bsum[index] = bval;
} else {
if (debugflag) {
L_INFO("maxcolors exceeded for maxdist = %d\n",
procName, maxdist);
}
return 1;
}
}
}
}
/* Replace the colors in the colormap by the averages */
for (k = 0; k < ncolors; k++) {
rval = rsum[k] / countarray[k];
gval = gsum[k] / countarray[k];
bval = bsum[k] / countarray[k];
pixcmapResetColor(cmap, k, rval, gval, bval);
}
return 0;
}
/*!
* \brief pixAssignToNearestColor()
*
* \param[in] pixd 8 bpp, colormapped
* \param[in] pixs 32 bpp; 24-bit color
* \param[in] pixm [optional] 1 bpp
* \param[in] level of octcube used for finding nearest color in cmap
* \param[in] countarray [optional] ptr to array, in which we can store
* the number of pixels found in each color in
* the colormap in pixd
* \return 0 if OK, 1 on error
*
*
* Notes:
* (1) This is used in phase 2 of color segmentation, where pixs
* is the original input image to pixColorSegment(), and
* pixd is the colormapped image returned from
* pixColorSegmentCluster(). It is also used, with a mask,
* in phase 4.
* (2) This is an in-place operation.
* (3) The colormap in pixd is unchanged.
* (4) pixs and pixd must be the same size (w, h).
* (5) The selection mask pixm can be null. If it exists, it must
* be the same size as pixs and pixd, and only pixels
* corresponding to fg in pixm are assigned. Set to
* NULL if all pixels in pixd are to be assigned.
* (6) The countarray can be null. If it exists, it is pre-allocated
* and of a size at least equal to the size of the colormap in pixd.
* (7) This does a best-fit (non-greedy) assignment of pixels to
* existing clusters. Specifically, it assigns each pixel
* in pixd to the color index in the pixd colormap that has a
* color closest to the corresponding rgb pixel in pixs.
* (8) 'level' is the octcube level used to quickly find the nearest
* color in the colormap for each pixel. For color segmentation,
* this parameter is set to LEVEL_IN_OCTCUBE.
* (9) We build a mapping table from octcube to colormap index so
* that this function can run in a time (otherwise) independent
* of the number of colors in the colormap. This avoids a
* brute-force search for the closest colormap color to each
* pixel in the image.
*
*/
l_ok
pixAssignToNearestColor(PIX *pixd,
PIX *pixs,
PIX *pixm,
l_int32 level,
l_int32 *countarray)
{
l_int32 w, h, wpls, wpld, wplm, i, j, success;
l_int32 rval, gval, bval, index;
l_int32 *cmaptab;
l_uint32 octindex;
l_uint32 *rtab, *gtab, *btab;
l_uint32 *ppixel;
l_uint32 *datas, *datad, *datam, *lines, *lined, *linem;
PIXCMAP *cmap;
PROCNAME("pixAssignToNearestColor");
if (!pixd)
return ERROR_INT("pixd not defined", procName, 1);
if ((cmap = pixGetColormap(pixd)) == NULL)
return ERROR_INT("cmap not found", procName, 1);
if (!pixs)
return ERROR_INT("pixs not defined", procName, 1);
if (pixGetDepth(pixs) != 32)
return ERROR_INT("pixs not 32 bpp", procName, 1);
if (level < 1 || level > 6)
return ERROR_INT("level not in [1 ... 6]", procName, 1);
/* Set up the tables to map rgb to the nearest colormap index */
success = TRUE;
makeRGBToIndexTables(level, &rtab, >ab, &btab);
cmaptab = pixcmapToOctcubeLUT(cmap, level, L_MANHATTAN_DISTANCE);
if (!rtab || !gtab || !btab || !cmaptab) {
L_ERROR("failure to make a table\n", procName);
success = FALSE;
goto cleanup_arrays;
}
pixGetDimensions(pixs, &w, &h, NULL);
datas = pixGetData(pixs);
datad = pixGetData(pixd);
wpls = pixGetWpl(pixs);
wpld = pixGetWpl(pixd);
if (pixm) {
datam = pixGetData(pixm);
wplm = pixGetWpl(pixm);
}
for (i = 0; i < h; i++) {
lines = datas + i * wpls;
lined = datad + i * wpld;
if (pixm)
linem = datam + i * wplm;
for (j = 0; j < w; j++) {
if (pixm) {
if (!GET_DATA_BIT(linem, j))
continue;
}
ppixel = lines + j;
rval = GET_DATA_BYTE(ppixel, COLOR_RED);
gval = GET_DATA_BYTE(ppixel, COLOR_GREEN);
bval = GET_DATA_BYTE(ppixel, COLOR_BLUE);
/* Map from rgb to octcube index */
getOctcubeIndexFromRGB(rval, gval, bval, rtab, gtab, btab,
&octindex);
/* Map from octcube index to nearest colormap index */
index = cmaptab[octindex];
if (countarray)
countarray[index]++;
SET_DATA_BYTE(lined, j, index);
}
}
cleanup_arrays:
LEPT_FREE(cmaptab);
LEPT_FREE(rtab);
LEPT_FREE(gtab);
LEPT_FREE(btab);
return (success) ? 0 : 1;
}
/*!
* \brief pixColorSegmentClean()
*
* \param[in] pixs 8 bpp, colormapped
* \param[in] selsize for closing
* \param[in] countarray ptr to array containing the number of pixels
* found in each color in the colormap
* \return 0 if OK, 1 on error
*
*
* Notes:
* (1) This operation is in-place.
* (2) This is phase 3 of color segmentation. It is the first
* part of a two-step noise removal process. Colors with a
* large population are closed first; this operation absorbs
* small sets of intercolated pixels of a different color.
*
*/
l_ok
pixColorSegmentClean(PIX *pixs,
l_int32 selsize,
l_int32 *countarray)
{
l_int32 i, ncolors, val;
l_uint32 val32;
NUMA *na, *nasi;
PIX *pixt1, *pixt2;
PIXCMAP *cmap;
PROCNAME("pixColorSegmentClean");
if (!pixs)
return ERROR_INT("pixs not defined", procName, 1);
if (pixGetDepth(pixs) != 8)
return ERROR_INT("pixs not 8 bpp", procName, 1);
if ((cmap = pixGetColormap(pixs)) == NULL)
return ERROR_INT("cmap not found", procName, 1);
if (!countarray)
return ERROR_INT("countarray not defined", procName, 1);
if (selsize <= 1)
return 0; /* nothing to do */
/* Sort colormap indices in decreasing order of pixel population */
ncolors = pixcmapGetCount(cmap);
na = numaCreate(ncolors);
for (i = 0; i < ncolors; i++)
numaAddNumber(na, countarray[i]);
nasi = numaGetSortIndex(na, L_SORT_DECREASING);
numaDestroy(&na);
if (!nasi)
return ERROR_INT("nasi not made", procName, 1);
/* For each color, in order of decreasing population,
* do a closing and absorb the added pixels. Note that
* if the closing removes pixels at the border, they'll
* still appear in the xor and will be properly (re)set. */
for (i = 0; i < ncolors; i++) {
numaGetIValue(nasi, i, &val);
pixt1 = pixGenerateMaskByValue(pixs, val, 1);
pixt2 = pixCloseSafeCompBrick(NULL, pixt1, selsize, selsize);
pixXor(pixt2, pixt2, pixt1); /* pixels to be added to type 'val' */
pixcmapGetColor32(cmap, val, &val32);
pixSetMasked(pixs, pixt2, val32); /* add them */
pixDestroy(&pixt1);
pixDestroy(&pixt2);
}
numaDestroy(&nasi);
return 0;
}
/*!
* \brief pixColorSegmentRemoveColors()
*
* \param[in] pixd 8 bpp, colormapped
* \param[in] pixs 32 bpp rgb, with initial pixel values
* \param[in] finalcolors max number of colors to retain
* \return 0 if OK, 1 on error
*
*
* Notes:
* (1) This operation is in-place.
* (2) This is phase 4 of color segmentation, and the second part
* of the 2-step noise removal. Only 'finalcolors' different
* colors are retained, with colors with smaller populations
* being replaced by the nearest color of the remaining colors.
* For highest accuracy, for pixels that are being replaced,
* we find the nearest colormap color to the original rgb color.
*
*/
l_ok
pixColorSegmentRemoveColors(PIX *pixd,
PIX *pixs,
l_int32 finalcolors)
{
l_int32 i, ncolors, index, tempindex;
l_int32 *tab;
l_uint32 tempcolor;
NUMA *na, *nasi;
PIX *pixm;
PIXCMAP *cmap;
PROCNAME("pixColorSegmentRemoveColors");
if (!pixd)
return ERROR_INT("pixd not defined", procName, 1);
if (pixGetDepth(pixd) != 8)
return ERROR_INT("pixd not 8 bpp", procName, 1);
if ((cmap = pixGetColormap(pixd)) == NULL)
return ERROR_INT("cmap not found", procName, 1);
if (!pixs)
return ERROR_INT("pixs not defined", procName, 1);
ncolors = pixcmapGetCount(cmap);
if (finalcolors >= ncolors) /* few enough colors already; nothing to do */
return 0;
/* Generate a mask over all pixels that are not in the
* 'finalcolors' most populated colors. Save the colormap
* index of any one of the retained colors in 'tempindex'.
* The LUT has values 0 for the 'finalcolors' most populated colors,
* which will be retained; and 1 for the rest, which are marked
* by fg pixels in pixm and will be removed. */
na = pixGetCmapHistogram(pixd, 1);
if ((nasi = numaGetSortIndex(na, L_SORT_DECREASING)) == NULL) {
numaDestroy(&na);
return ERROR_INT("nasi not made", procName, 1);
}
numaGetIValue(nasi, finalcolors - 1, &tempindex); /* retain down to this */
pixcmapGetColor32(cmap, tempindex, &tempcolor); /* use this color */
tab = (l_int32 *)LEPT_CALLOC(256, sizeof(l_int32));
for (i = finalcolors; i < ncolors; i++) {
numaGetIValue(nasi, i, &index);
tab[index] = 1;
}
pixm = pixMakeMaskFromLUT(pixd, tab);
LEPT_FREE(tab);
/* Reassign the masked pixels temporarily to the saved index
* (tempindex). This guarantees that no pixels are labeled by
* a colormap index of any colors that will be removed.
* The actual value doesn't matter, as long as it's one
* of the retained colors, because these pixels will later
* be reassigned based on the full set of colors retained
* in the colormap. */
pixSetMasked(pixd, pixm, tempcolor);
/* Now remove unused colors from the colormap. This reassigns
* image pixels as required. */
pixRemoveUnusedColors(pixd);
/* Finally, reassign the pixels under the mask (those that were
* given a 'tempindex' value) to the nearest color in the colormap.
* This is the function used in phase 2 on all image pixels; here
* it is only used on the masked pixels given by pixm. */
pixAssignToNearestColor(pixd, pixs, pixm, LEVEL_IN_OCTCUBE, NULL);
pixDestroy(&pixm);
numaDestroy(&na);
numaDestroy(&nasi);
return 0;
}