/*!
* \brief boxaMakeWHRatioIndicator()
*
* \param[in] boxa
* \param[in] ratio width/height threshold value
* \param[in] relation L_SELECT_IF_LT, L_SELECT_IF_GT,
* L_SELECT_IF_LTE, L_SELECT_IF_GTE
* \return na indicator array, or NULL on error
*
* <pre>
* Notes:
* (1) To keep narrow components, use relation = L_SELECT_IF_LT or
* L_SELECT_IF_LTE.
* To keep wide components, use relation = L_SELECT_IF_GT or
* L_SELECT_IF_GTE.
* </pre>
*/
NUMA *
boxaMakeWHRatioIndicator(BOXA *boxa,
l_float32 ratio,
l_int32 relation)
{
l_int32 i, n, w, h, ival;
l_float32 whratio;
NUMA *na;
PROCNAME("boxaMakeWHRatioIndicator");
if (!boxa)
return (NUMA *)ERROR_PTR("boxa not defined", procName, NULL);
if ((n = boxaGetCount(boxa)) == 0)
return (NUMA *)ERROR_PTR("boxa is empty", procName, NULL);
if (relation != L_SELECT_IF_LT && relation != L_SELECT_IF_GT &&
relation != L_SELECT_IF_LTE && relation != L_SELECT_IF_GTE)
return (NUMA *)ERROR_PTR("invalid relation", procName, NULL);
na = numaCreate(n);
for (i = 0; i < n; i++) {
ival = 0;
boxaGetBoxGeometry(boxa, i, NULL, NULL, &w, &h);
whratio = (l_float32)w / (l_float32)h;
if ((relation == L_SELECT_IF_LT && whratio < ratio) ||
(relation == L_SELECT_IF_GT && whratio > ratio) ||
(relation == L_SELECT_IF_LTE && whratio <= ratio) ||
(relation == L_SELECT_IF_GTE && whratio >= ratio))
ival = 1;
numaAddNumber(na, ival);
}
return na;
}
/*!
* \brief boxaMakeAreaIndicator()
*
* \param[in] boxa
* \param[in] area threshold value of width * height
* \param[in] relation L_SELECT_IF_LT, L_SELECT_IF_GT,
* L_SELECT_IF_LTE, L_SELECT_IF_GTE
* \return na indicator array, or NULL on error
*
* <pre>
* Notes:
* (1) To keep small components, use relation = L_SELECT_IF_LT or
* L_SELECT_IF_LTE.
* To keep large components, use relation = L_SELECT_IF_GT or
* L_SELECT_IF_GTE.
* </pre>
*/
NUMA *
boxaMakeAreaIndicator(BOXA *boxa,
l_int32 area,
l_int32 relation)
{
l_int32 i, n, w, h, ival;
NUMA *na;
PROCNAME("boxaMakeAreaIndicator");
if (!boxa)
return (NUMA *)ERROR_PTR("boxa not defined", procName, NULL);
if ((n = boxaGetCount(boxa)) == 0)
return (NUMA *)ERROR_PTR("boxa is empty", procName, NULL);
if (relation != L_SELECT_IF_LT && relation != L_SELECT_IF_GT &&
relation != L_SELECT_IF_LTE && relation != L_SELECT_IF_GTE)
return (NUMA *)ERROR_PTR("invalid relation", procName, NULL);
na = numaCreate(n);
for (i = 0; i < n; i++) {
ival = 0;
boxaGetBoxGeometry(boxa, i, NULL, NULL, &w, &h);
if ((relation == L_SELECT_IF_LT && w * h < area) ||
(relation == L_SELECT_IF_GT && w * h > area) ||
(relation == L_SELECT_IF_LTE && w * h <= area) ||
(relation == L_SELECT_IF_GTE && w * h >= area))
ival = 1;
numaAddNumber(na, ival);
}
return na;
}
// Converts the Boxa array to a list of C_BLOB, getting rid of severely
// overlapping outlines and those that are children of a bigger one.
// The output is a list of C_BLOBs that are owned by the list.
// The C_OUTLINEs in the C_BLOBs contain no outline data - just empty
// bounding boxes. The Boxa is consumed and destroyed.
void LineFinder::ConvertBoxaToBlobs(int image_width, int image_height,
Boxa** boxes, C_BLOB_LIST* blobs) {
#ifdef HAVE_LIBLEPT
C_OUTLINE_LIST outlines;
C_OUTLINE_IT ol_it = &outlines;
// Iterate the boxes to convert to outlines.
int nboxes = boxaGetCount(*boxes);
for (int i = 0; i < nboxes; ++i) {
l_int32 x, y, width, height;
boxaGetBoxGeometry(*boxes, i, &x, &y, &width, &height);
// Make a C_OUTLINE from the leptonica box. This is a bit of a hack,
// as there is no outline, just a bounding box, but with some very
// small changes to coutln.cpp, it works nicely.
ICOORD top_left(x, image_height - y);
ICOORD bot_right(x + width, image_height - (y + height));
CRACKEDGE startpt;
startpt.pos = top_left;
C_OUTLINE* outline = new C_OUTLINE(&startpt, top_left, bot_right, 0);
ol_it.add_after_then_move(outline);
}
// Use outlines_to_blobs to convert the outlines to blobs and find
// overlapping and contained objects. The output list of blobs in the block
// has all the bad ones filtered out and deleted.
BLOCK block;
ICOORD page_tl(0, 0);
ICOORD page_br(image_width, image_height);
outlines_to_blobs(&block, page_tl, page_br, &outlines);
// Transfer the created blobs to the output list.
C_BLOB_IT blob_it(blobs);
blob_it.add_list_after(block.blob_list());
// The boxes aren't needed any more.
boxaDestroy(boxes);
#endif
}
/*!
* boxaConvertToPta()
*
* Input: boxa
* ncorners (2 or 4 for the representation of each box)
* Return: pta (with @ncorners points for each box in the boxa),
* or null on error
*
* Notes:
* (1) If ncorners == 2, we select the UL and LR corners.
* Otherwise we save all 4 corners in this order: UL, UR, LL, LR.
*/
PTA *
boxaConvertToPta(BOXA *boxa,
l_int32 ncorners)
{
l_int32 i, n, x, y, w, h;
PTA *pta;
PROCNAME("boxaConvertToPta");
if (!boxa)
return (PTA *)ERROR_PTR("boxa not defined", procName, NULL);
if (ncorners != 2 && ncorners != 4)
return (PTA *)ERROR_PTR("ncorners not 2 or 4", procName, NULL);
n = boxaGetCount(boxa);
if ((pta = ptaCreate(n)) == NULL)
return (PTA *)ERROR_PTR("pta not made", procName, NULL);
for (i = 0; i < n; i++) {
boxaGetBoxGeometry(boxa, i, &x, &y, &w, &h);
ptaAddPt(pta, x, y);
if (ncorners == 2)
ptaAddPt(pta, x + w - 1, y + h - 1);
else {
ptaAddPt(pta, x + w - 1, y);
ptaAddPt(pta, x, y + h - 1);
ptaAddPt(pta, x + w - 1, y + h - 1);
}
}
return pta;
}
/*!
* boxaaGetExtent()
*
* Input: boxaa
* &w (<optional return> width)
* &h (<optional return> height)
* &box (<optional return>, minimum box containing all boxa
* in boxaa)
* Return: 0 if OK, 1 on error
*
* Notes:
* (1) The returned w and h are the minimum size image
* that would contain all boxes untranslated.
*/
l_int32
boxaaGetExtent(BOXAA *boxaa,
l_int32 *pw,
l_int32 *ph,
BOX **pbox)
{
l_int32 i, j, n, m, x, y, w, h, xmax, ymax, xmin, ymin, found;
BOXA *boxa;
PROCNAME("boxaaGetExtent");
if (!pw && !ph && !pbox)
return ERROR_INT("no ptrs defined", procName, 1);
if (pbox) *pbox = NULL;
if (pw) *pw = 0;
if (ph) *ph = 0;
if (!boxaa)
return ERROR_INT("boxaa not defined", procName, 1);
n = boxaaGetCount(boxaa);
if (n == 0)
return ERROR_INT("no boxa in boxaa", procName, 1);
xmax = ymax = 0;
xmin = ymin = 100000000;
found = FALSE;
for (i = 0; i < n; i++) {
boxa = boxaaGetBoxa(boxaa, i, L_CLONE);
m = boxaGetCount(boxa);
for (j = 0; j < m; j++) {
boxaGetBoxGeometry(boxa, j, &x, &y, &w, &h);
if (w <= 0 || h <= 0)
continue;
found = TRUE;
xmin = L_MIN(xmin, x);
ymin = L_MIN(ymin, y);
xmax = L_MAX(xmax, x + w);
ymax = L_MAX(ymax, y + h);
}
}
if (!found)
return ERROR_INT("no valid boxes in boxaa", procName, 1);
if (pw) *pw = xmax;
if (ph) *ph = ymax;
if (pbox)
*pbox = boxCreate(xmin, ymin, xmax - xmin, ymax - ymin);
return 0;
}
/*!
* boxaaAlignBox()
*
* Input: boxaa
* box (to be aligned with the last of one of the boxa
* in boxaa, if possible)
* delta (amount by which consecutive components can miss
* in overlap and still be included in the array)
* &index (of boxa with best overlap, or if none match,
* this is the index of the next boxa to be generated)
* Return: 0 if OK, 1 on error
*
* Notes:
* (1) This is not greedy; it finds the boxa whose last box has
* the biggest overlap with the input box.
*/
l_int32
boxaaAlignBox(BOXAA *baa,
BOX *box,
l_int32 delta,
l_int32 *pindex)
{
l_int32 i, n, m, y, yt, h, ht, ovlp, maxovlp, maxindex;
BOXA *boxa;
PROCNAME("boxaaAlignBox");
if (!baa)
return ERROR_INT("baa not defined", procName, 1);
if (!box)
return ERROR_INT("box not defined", procName, 1);
if (!pindex)
return ERROR_INT("&index not defined", procName, 1);
n = boxaaGetCount(baa);
boxGetGeometry(box, NULL, &y, NULL, &h);
maxovlp = -10000000;
for (i = 0; i < n; i++) {
boxa = boxaaGetBoxa(baa, i, L_CLONE);
if ((m = boxaGetCount(boxa)) == 0) {
L_WARNING("no boxes in boxa", procName);
continue;
}
boxaGetBoxGeometry(boxa, m - 1, NULL, &yt, NULL, &ht); /* last one */
boxaDestroy(&boxa);
/* Overlap < 0 means the components do not overlap vertically */
if (yt >= y)
ovlp = y + h - 1 - yt;
else
ovlp = yt + ht - 1 - y;
if (ovlp > maxovlp) {
maxovlp = ovlp;
maxindex = i;
}
}
if (maxovlp + delta >= 0)
*pindex = maxindex;
else
*pindex = n;
return 0;
}
/*!
* boxaGetValidCount()
*
* Input: boxa
* Return: count (of valid boxes); 0 if no valid boxes or on error
*/
l_int32
boxaGetValidCount(BOXA *boxa)
{
l_int32 n, i, w, h, count;
PROCNAME("boxaGetValidCount");
if (!boxa)
return ERROR_INT("boxa not defined", procName, 0);
n = boxaGetCount(boxa);
for (i = 0, count = 0; i < n; i++) {
boxaGetBoxGeometry(boxa, i, NULL, NULL, &w, &h);
if (w > 0 && h > 0)
count++;
}
return count;
}
// Get a set of bounding boxes of possible horizontal lines in the image.
// The input resolution overrides any resolution set in src_pix.
// The output line_pix contains just all the detected lines.
// The output boxes undergo the transformation (x,y)->(height-y,x) so the
// lines can be found with a vertical line finder afterwards.
// This transformation allows a simple x/y flip to reverse it in tesseract
// coordinates and it is faster to flip the lines than rotate the image.
Boxa* LineFinder::GetHLineBoxes(int resolution, Pix* src_pix, Pix** line_pix) {
#ifdef HAVE_LIBLEPT
// Remove any parts of 1 inch/kThinLineFraction high or more, by opening
// away the thin lines and subtracting what's left.
// This is very generous and will leave in even quite wide lines.
Pix* pixt1 = pixOpenBrick(NULL, src_pix, 1, resolution / kThinLineFraction);
pixSubtract(pixt1, src_pix, pixt1);
// Spread vertically to allow for some skew.
Pix* pixt2 = pixDilateBrick(NULL, pixt1, 1, 3);
// Now keep only wide stuff of width at least 1 inch/kMinLineLengthFraction.
pixOpenBrick(pixt1, pixt2, resolution / kMinLineLengthFraction, 1);
pixDestroy(&pixt2);
// Put a single pixel crack in every line at an arbitrary spacing,
// so they break up and the bounding boxes can be used to get the
// direction accurately enough without needing outlines.
int wpl = pixGetWpl(pixt1);
int width = pixGetWidth(pixt1);
int height = pixGetHeight(pixt1);
l_uint32* data = pixGetData(pixt1);
for (int y = 0; y < height; ++y, data += wpl) {
for (int x = kCrackSpacing; x < width; x += kCrackSpacing) {
CLEAR_DATA_BIT(data, x);
}
}
if (textord_tabfind_show_vlines)
pixWrite("hlines.png", pixt1, IFF_PNG);
Boxa* boxa = pixConnComp(pixt1, NULL, 8);
*line_pix = pixt1;
// Iterate the boxes to flip x and y.
int nboxes = boxaGetCount(boxa);
for (int i = 0; i < nboxes; ++i) {
l_int32 x, y, box_width, box_height;
boxaGetBoxGeometry(boxa, i, &x, &y, &box_width, &box_height);
Box* box = boxCreate(height - (y + box_height),
width - (x + box_width), box_height, box_width);
boxaReplaceBox(boxa, i, box);
}
return boxa;
#else
return NULL;
#endif
}
/*!
* \brief boxaLocationRange()
*
* \param[in] boxa
* \param[out] pminx [optional] min (UL corner) x value of all boxes
* \param[out] pminy [optional] min (UL corner) y value of all boxes
* \param[out] pmaxx [optional] max (UL corner) x value of all boxes
* \param[out] pmaxy [optional] max (UL corner) y value of all boxes
* \return 0 if OK, 1 on error
*/
l_ok
boxaLocationRange(BOXA *boxa,
l_int32 *pminx,
l_int32 *pminy,
l_int32 *pmaxx,
l_int32 *pmaxy)
{
l_int32 minx, miny, maxx, maxy, i, n, x, y;
PROCNAME("boxaLocationRange");
if (!pminx && !pminy && !pmaxx && !pmaxy)
return ERROR_INT("no data can be returned", procName, 1);
if (pminx) *pminx = 0;
if (pminy) *pminy = 0;
if (pmaxx) *pmaxx = 0;
if (pmaxy) *pmaxy = 0;
if (!boxa)
return ERROR_INT("boxa not defined", procName, 1);
minx = miny = 100000000;
maxx = maxy = 0;
n = boxaGetCount(boxa);
for (i = 0; i < n; i++) {
boxaGetBoxGeometry(boxa, i, &x, &y, NULL, NULL);
if (x < minx)
minx = x;
if (y < miny)
miny = y;
if (x > maxx)
maxx = x;
if (y > maxy)
maxy = y;
}
if (pminx) *pminx = minx;
if (pminy) *pminy = miny;
if (pmaxx) *pmaxx = maxx;
if (pmaxy) *pmaxy = maxy;
return 0;
}
/*!
* boxaExtractAsPta()
*
* Input: boxa
* &ptal (<optional return> array of left locations vs. index)
* &ptat (<optional return> array of top locations vs. index)
* &ptar (<optional return> array of right locations vs. index)
* &ptab (<optional return> array of bottom locations vs. index)
* keepinvalid (1 to keep invalid boxes; 0 to remove them)
* Return: 0 if OK, 1 on error
*/
l_int32
boxaExtractAsPta(BOXA *boxa,
PTA **pptal,
PTA **pptat,
PTA **pptar,
PTA **pptab,
l_int32 keepinvalid)
{
l_int32 i, n, left, top, right, bot, w, h;
PROCNAME("boxaExtractAsPta");
if (!pptal && !pptar && !pptat && !pptab)
return ERROR_INT("no output requested", procName, 1);
if (pptal) *pptal = NULL;
if (pptat) *pptat = NULL;
if (pptar) *pptar = NULL;
if (pptab) *pptab = NULL;
if (!boxa)
return ERROR_INT("boxa not defined", procName, 1);
if (!keepinvalid && boxaGetValidCount(boxa) == 0)
return ERROR_INT("no valid boxes", procName, 1);
n = boxaGetCount(boxa);
if (pptal) *pptal = ptaCreate(n);
if (pptat) *pptat = ptaCreate(n);
if (pptar) *pptar = ptaCreate(n);
if (pptab) *pptab = ptaCreate(n);
for (i = 0; i < n; i++) {
boxaGetBoxGeometry(boxa, i, &left, &top, &w, &h);
if (!keepinvalid && (w <= 0 || h <= 0))
continue;
right = left + w - 1;
bot = top + h - 1;
if (pptal) ptaAddPt(*pptal, i, left);
if (pptat) ptaAddPt(*pptat, i, top);
if (pptar) ptaAddPt(*pptar, i, right);
if (pptab) ptaAddPt(*pptab, i, bot);
}
return 0;
}
/*!
* \brief boxaGetExtent()
*
* \param[in] boxa
* \param[out] pw [optional] width
* \param[out] ph [optional] height
* \param[out] pbox [optional] minimum box containing all boxes in boxa
* \return 0 if OK, 1 on error
*
* <pre>
* Notes:
* (1) This computes the minimum rectangular bounding region
* that contains all valid boxes in a boxa.
* (2) The returned w and h are the minimum size image
* that would contain all boxes untranslated.
* (3) If there are no valid boxes, returned w and h are 0 and
* all parameters in the returned box are 0. This
* is not an error, because an empty boxa is valid and
* boxaGetExtent() is required for serialization.
* </pre>
*/
l_ok
boxaGetExtent(BOXA *boxa,
l_int32 *pw,
l_int32 *ph,
BOX **pbox)
{
l_int32 i, n, x, y, w, h, xmax, ymax, xmin, ymin, found;
PROCNAME("boxaGetExtent");
if (!pw && !ph && !pbox)
return ERROR_INT("no ptrs defined", procName, 1);
if (pw) *pw = 0;
if (ph) *ph = 0;
if (pbox) *pbox = NULL;
if (!boxa)
return ERROR_INT("boxa not defined", procName, 1);
n = boxaGetCount(boxa);
xmax = ymax = 0;
xmin = ymin = 100000000;
found = FALSE;
for (i = 0; i < n; i++) {
boxaGetBoxGeometry(boxa, i, &x, &y, &w, &h);
if (w <= 0 || h <= 0)
continue;
found = TRUE;
xmin = L_MIN(xmin, x);
ymin = L_MIN(ymin, y);
xmax = L_MAX(xmax, x + w);
ymax = L_MAX(ymax, y + h);
}
if (found == FALSE) /* no valid boxes in boxa */
xmin = ymin = 0;
if (pw) *pw = xmax;
if (ph) *ph = ymax;
if (pbox)
*pbox = boxCreate(xmin, ymin, xmax - xmin, ymax - ymin);
return 0;
}
/*!
* \brief boxaSizeRange()
*
* \param[in] boxa
* \param[out] pminw [optional] min width of all boxes
* \param[out] pmaxw [optional] max width of all boxes
* \param[out] pminh [optional] min height of all boxes
* \param[out] pmaxh [optional] max height of all boxes
* \return 0 if OK, 1 on error
*/
l_ok
boxaSizeRange(BOXA *boxa,
l_int32 *pminw,
l_int32 *pminh,
l_int32 *pmaxw,
l_int32 *pmaxh)
{
l_int32 minw, minh, maxw, maxh, i, n, w, h;
PROCNAME("boxaSizeRange");
if (!pminw && !pmaxw && !pminh && !pmaxh)
return ERROR_INT("no data can be returned", procName, 1);
if (pminw) *pminw = 0;
if (pminh) *pminh = 0;
if (pmaxw) *pmaxw = 0;
if (pmaxh) *pmaxh = 0;
if (!boxa)
return ERROR_INT("boxa not defined", procName, 1);
minw = minh = 100000000;
maxw = maxh = 0;
n = boxaGetCount(boxa);
for (i = 0; i < n; i++) {
boxaGetBoxGeometry(boxa, i, NULL, NULL, &w, &h);
if (w < minw)
minw = w;
if (h < minh)
minh = h;
if (w > maxw)
maxw = w;
if (h > maxh)
maxh = h;
}
if (pminw) *pminw = minw;
if (pminh) *pminh = minh;
if (pmaxw) *pmaxw = maxw;
if (pmaxh) *pmaxh = maxh;
return 0;
}
/*!
* \brief boxaGetArea()
*
* \param[in] boxa
* \param[out] parea total area of all boxes
* \return 0 if OK, 1 on error
*
* <pre>
* Notes:
* (1) Measures the total area of the boxes, without regard to overlaps.
* </pre>
*/
l_ok
boxaGetArea(BOXA *boxa,
l_int32 *parea)
{
l_int32 i, n, w, h;
PROCNAME("boxaGetArea");
if (!parea)
return ERROR_INT("&area not defined", procName, 1);
*parea = 0;
if (!boxa)
return ERROR_INT("boxa not defined", procName, 1);
n = boxaGetCount(boxa);
for (i = 0; i < n; i++) {
boxaGetBoxGeometry(boxa, i, NULL, NULL, &w, &h);
*parea += w * h;
}
return 0;
}
jboolean Java_com_googlecode_leptonica_android_Boxa_nativeGetGeometry(JNIEnv *env, jclass clazz,
jlong nativeBoxa,
jint index,
jintArray dimensions) {
BOXA *boxa = (BOXA *) nativeBoxa;
jint *dimensionArray = env->GetIntArrayElements(dimensions, NULL);
l_int32 x, y, w, h;
if (boxaGetBoxGeometry(boxa, index, &x, &y, &w, &h)) {
return JNI_FALSE;
}
dimensionArray[0] = x;
dimensionArray[1] = y;
dimensionArray[2] = w;
dimensionArray[3] = h;
env->ReleaseIntArrayElements(dimensions, dimensionArray, 0);
return JNI_TRUE;
}
/*!
* boxaExtractAsNuma()
*
* Input: boxa
* &nax (<optional return> array of x locations)
* &nay (<optional return> array of y locations)
* &naw (<optional return> array of w locations)
* &nah (<optional return> array of h locations)
* keepinvalid (1 to keep invalid boxes; 0 to remove them)
* Return: 0 if OK, 1 on error
*/
l_int32
boxaExtractAsNuma(BOXA *boxa,
NUMA **pnax,
NUMA **pnay,
NUMA **pnaw,
NUMA **pnah,
l_int32 keepinvalid)
{
l_int32 i, n, x, y, w, h;
PROCNAME("boxaExtractAsNuma");
if (!pnax && !pnay && !pnaw && !pnah)
return ERROR_INT("no output requested", procName, 1);
if (pnax) *pnax = NULL;
if (pnay) *pnay = NULL;
if (pnaw) *pnaw = NULL;
if (pnah) *pnah = NULL;
if (!boxa)
return ERROR_INT("boxa not defined", procName, 1);
if (!keepinvalid && boxaGetValidCount(boxa) == 0)
return ERROR_INT("no valid boxes", procName, 1);
n = boxaGetCount(boxa);
if (pnax) *pnax = numaCreate(n);
if (pnay) *pnay = numaCreate(n);
if (pnaw) *pnaw = numaCreate(n);
if (pnah) *pnah = numaCreate(n);
for (i = 0; i < n; i++) {
boxaGetBoxGeometry(boxa, i, &x, &y, &w, &h);
if (!keepinvalid && (w <= 0 || h <= 0))
continue;
if (pnax) numaAddNumber(*pnax, x);
if (pnay) numaAddNumber(*pnay, y);
if (pnaw) numaAddNumber(*pnaw, w);
if (pnah) numaAddNumber(*pnah, h);
}
return 0;
}
/*!
* \brief pixFindBaselines()
*
* \param[in] pixs 1 bpp, 300 ppi
* \param[out] ppta [optional] pairs of pts corresponding to
* approx. ends of each text line
* \param[in] pixadb for debug output; use NULL to skip
* \return na of baseline y values, or NULL on error
*
* <pre>
* Notes:
* (1) Input binary image must have text lines already aligned
* horizontally. This can be done by either rotating the
* image with pixDeskew(), or, if a projective transform
* is required, by doing pixDeskewLocal() first.
* (2) Input null for &pta if you don't want this returned.
* The pta will come in pairs of points (left and right end
* of each baseline).
* (3) Caution: this will not work properly on text with multiple
* columns, where the lines are not aligned between columns.
* If there are multiple columns, they should be extracted
* separately before finding the baselines.
* (4) This function constructs different types of output
* for baselines; namely, a set of raster line values and
* a set of end points of each baseline.
* (5) This function was designed to handle short and long text lines
* without using dangerous thresholds on the peak heights. It does
* this by combining the differential signal with a morphological
* analysis of the locations of the text lines. One can also
* combine this data to normalize the peak heights, by weighting
* the differential signal in the region of each baseline
* by the inverse of the width of the text line found there.
* </pre>
*/
NUMA *
pixFindBaselines(PIX *pixs,
PTA **ppta,
PIXA *pixadb)
{
l_int32 h, i, j, nbox, val1, val2, ndiff, bx, by, bw, bh;
l_int32 imaxloc, peakthresh, zerothresh, inpeak;
l_int32 mintosearch, max, maxloc, nloc, locval;
l_int32 *array;
l_float32 maxval;
BOXA *boxa1, *boxa2, *boxa3;
GPLOT *gplot;
NUMA *nasum, *nadiff, *naloc, *naval;
PIX *pix1, *pix2;
PTA *pta;
PROCNAME("pixFindBaselines");
if (ppta) *ppta = NULL;
if (!pixs || pixGetDepth(pixs) != 1)
return (NUMA *)ERROR_PTR("pixs undefined or not 1 bpp", procName, NULL);
/* Close up the text characters, removing noise */
pix1 = pixMorphSequence(pixs, "c25.1 + e15.1", 0);
/* Estimate the resolution */
if (pixadb) pixaAddPix(pixadb, pixScale(pix1, 0.25, 0.25), L_INSERT);
/* Save the difference of adjacent row sums.
* The high positive-going peaks are the baselines */
if ((nasum = pixCountPixelsByRow(pix1, NULL)) == NULL) {
pixDestroy(&pix1);
return (NUMA *)ERROR_PTR("nasum not made", procName, NULL);
}
h = pixGetHeight(pixs);
nadiff = numaCreate(h);
numaGetIValue(nasum, 0, &val2);
for (i = 0; i < h - 1; i++) {
val1 = val2;
numaGetIValue(nasum, i + 1, &val2);
numaAddNumber(nadiff, val1 - val2);
}
numaDestroy(&nasum);
if (pixadb) { /* show the difference signal */
lept_mkdir("lept/baseline");
gplotSimple1(nadiff, GPLOT_PNG, "/tmp/lept/baseline/diff", "Diff Sig");
pix2 = pixRead("/tmp/lept/baseline/diff.png");
pixaAddPix(pixadb, pix2, L_INSERT);
}
/* Use the zeroes of the profile to locate each baseline. */
array = numaGetIArray(nadiff);
ndiff = numaGetCount(nadiff);
numaGetMax(nadiff, &maxval, &imaxloc);
numaDestroy(&nadiff);
/* Use this to begin locating a new peak: */
peakthresh = (l_int32)maxval / PEAK_THRESHOLD_RATIO;
/* Use this to begin a region between peaks: */
zerothresh = (l_int32)maxval / ZERO_THRESHOLD_RATIO;
naloc = numaCreate(0);
naval = numaCreate(0);
inpeak = FALSE;
for (i = 0; i < ndiff; i++) {
if (inpeak == FALSE) {
if (array[i] > peakthresh) { /* transition to in-peak */
inpeak = TRUE;
mintosearch = i + MIN_DIST_IN_PEAK; /* accept no zeros
* between i and mintosearch */
max = array[i];
maxloc = i;
}
} else { /* inpeak == TRUE; look for max */
if (array[i] > max) {
max = array[i];
maxloc = i;
mintosearch = i + MIN_DIST_IN_PEAK;
} else if (i > mintosearch && array[i] <= zerothresh) { /* leave */
inpeak = FALSE;
numaAddNumber(naval, max);
numaAddNumber(naloc, maxloc);
}
}
}
LEPT_FREE(array);
/* If array[ndiff-1] is max, eg. no descenders, baseline at bottom */
if (inpeak) {
numaAddNumber(naval, max);
numaAddNumber(naloc, maxloc);
}
if (pixadb) { /* show the raster locations for the peaks */
gplot = gplotCreate("/tmp/lept/baseline/loc", GPLOT_PNG, "Peak locs",
"rasterline", "height");
gplotAddPlot(gplot, naloc, naval, GPLOT_POINTS, "locs");
gplotMakeOutput(gplot);
gplotDestroy(&gplot);
pix2 = pixRead("/tmp/lept/baseline/loc.png");
pixaAddPix(pixadb, pix2, L_INSERT);
}
numaDestroy(&naval);
/* Generate an approximate profile of text line width.
* First, filter the boxes of text, where there may be
* more than one box for a given textline. */
pix2 = pixMorphSequence(pix1, "r11 + c20.1 + o30.1 +c1.3", 0);
if (pixadb) pixaAddPix(pixadb, pix2, L_COPY);
boxa1 = pixConnComp(pix2, NULL, 4);
pixDestroy(&pix1);
pixDestroy(&pix2);
if (boxaGetCount(boxa1) == 0) {
numaDestroy(&naloc);
boxaDestroy(&boxa1);
L_INFO("no compnents after filtering\n", procName);
return NULL;
}
boxa2 = boxaTransform(boxa1, 0, 0, 4., 4.);
boxa3 = boxaSort(boxa2, L_SORT_BY_Y, L_SORT_INCREASING, NULL);
boxaDestroy(&boxa1);
boxaDestroy(&boxa2);
/* Optionally, find the baseline segments */
pta = NULL;
if (ppta) {
pta = ptaCreate(0);
*ppta = pta;
}
if (pta) {
nloc = numaGetCount(naloc);
nbox = boxaGetCount(boxa3);
for (i = 0; i < nbox; i++) {
boxaGetBoxGeometry(boxa3, i, &bx, &by, &bw, &bh);
for (j = 0; j < nloc; j++) {
numaGetIValue(naloc, j, &locval);
if (L_ABS(locval - (by + bh)) > 25)
continue;
ptaAddPt(pta, bx, locval);
ptaAddPt(pta, bx + bw, locval);
break;
}
}
}
boxaDestroy(&boxa3);
if (pixadb && pta) { /* display baselines */
l_int32 npts, x1, y1, x2, y2;
pix1 = pixConvertTo32(pixs);
npts = ptaGetCount(pta);
for (i = 0; i < npts; i += 2) {
ptaGetIPt(pta, i, &x1, &y1);
ptaGetIPt(pta, i + 1, &x2, &y2);
pixRenderLineArb(pix1, x1, y1, x2, y2, 2, 255, 0, 0);
}
pixWrite("/tmp/lept/baseline/baselines.png", pix1, IFF_PNG);
pixaAddPix(pixadb, pixScale(pix1, 0.25, 0.25), L_INSERT);
pixDestroy(&pix1);
}
return naloc;
}
/*!
* boxaSort()
*
* Input: boxa
* sorttype (L_SORT_BY_X, L_SORT_BY_Y, L_SORT_BY_WIDTH,
* L_SORT_BY_HEIGHT, L_SORT_BY_MIN_DIMENSION,
* L_SORT_BY_MAX_DIMENSION, L_SORT_BY_PERIMETER,
* L_SORT_BY_AREA, L_SORT_BY_ASPECT_RATIO)
* sortorder (L_SORT_INCREASING, L_SORT_DECREASING)
* &naindex (<optional return> index of sorted order into
* original array)
* Return: boxad (sorted version of boxas), or null on error
*/
BOXA *
boxaSort(BOXA *boxas,
l_int32 sorttype,
l_int32 sortorder,
NUMA **pnaindex)
{
l_int32 i, n, x, y, w, h, size;
BOXA *boxad;
NUMA *na, *naindex;
PROCNAME("boxaSort");
if (pnaindex) *pnaindex = NULL;
if (!boxas)
return (BOXA *)ERROR_PTR("boxas not defined", procName, NULL);
if (sorttype != L_SORT_BY_X && sorttype != L_SORT_BY_Y &&
sorttype != L_SORT_BY_WIDTH && sorttype != L_SORT_BY_HEIGHT &&
sorttype != L_SORT_BY_MIN_DIMENSION &&
sorttype != L_SORT_BY_MAX_DIMENSION &&
sorttype != L_SORT_BY_PERIMETER &&
sorttype != L_SORT_BY_AREA &&
sorttype != L_SORT_BY_ASPECT_RATIO)
return (BOXA *)ERROR_PTR("invalid sort type", procName, NULL);
if (sortorder != L_SORT_INCREASING && sortorder != L_SORT_DECREASING)
return (BOXA *)ERROR_PTR("invalid sort order", procName, NULL);
/* Use O(n) binsort if possible */
n = boxaGetCount(boxas);
if (n > MIN_COMPS_FOR_BIN_SORT &&
((sorttype == L_SORT_BY_X) || (sorttype == L_SORT_BY_Y) ||
(sorttype == L_SORT_BY_WIDTH) || (sorttype == L_SORT_BY_HEIGHT) ||
(sorttype == L_SORT_BY_PERIMETER)))
return boxaBinSort(boxas, sorttype, sortorder, pnaindex);
/* Build up numa of specific data */
if ((na = numaCreate(n)) == NULL)
return (BOXA *)ERROR_PTR("na not made", procName, NULL);
for (i = 0; i < n; i++) {
boxaGetBoxGeometry(boxas, i, &x, &y, &w, &h);
switch (sorttype)
{
case L_SORT_BY_X:
numaAddNumber(na, x);
break;
case L_SORT_BY_Y:
numaAddNumber(na, y);
break;
case L_SORT_BY_WIDTH:
numaAddNumber(na, w);
break;
case L_SORT_BY_HEIGHT:
numaAddNumber(na, h);
break;
case L_SORT_BY_MIN_DIMENSION:
size = L_MIN(w, h);
numaAddNumber(na, size);
break;
case L_SORT_BY_MAX_DIMENSION:
size = L_MAX(w, h);
numaAddNumber(na, size);
break;
case L_SORT_BY_PERIMETER:
size = w + h;
numaAddNumber(na, size);
break;
case L_SORT_BY_AREA:
size = w * h;
numaAddNumber(na, size);
break;
case L_SORT_BY_ASPECT_RATIO:
numaAddNumber(na, (l_float32)w / (l_float32)h);
break;
default:
L_WARNING("invalid sort type", procName);
}
}
/* Get the sort index for data array */
if ((naindex = numaGetSortIndex(na, sortorder)) == NULL)
return (BOXA *)ERROR_PTR("naindex not made", procName, NULL);
/* Build up sorted boxa using sort index */
boxad = boxaSortByIndex(boxas, naindex);
if (pnaindex)
*pnaindex = naindex;
else
numaDestroy(&naindex);
numaDestroy(&na);
return boxad;
}
/*!
* boxaBinSort()
*
* Input: boxa
* sorttype (L_SORT_BY_X, L_SORT_BY_Y, L_SORT_BY_WIDTH,
* L_SORT_BY_HEIGHT, L_SORT_BY_PERIMETER)
* sortorder (L_SORT_INCREASING, L_SORT_DECREASING)
* &naindex (<optional return> index of sorted order into
* original array)
* Return: boxad (sorted version of boxas), or null on error
*
* Notes:
* (1) For a large number of boxes (say, greater than 1000), this
* O(n) binsort is much faster than the O(nlogn) shellsort.
* For 5000 components, this is over 20x faster than boxaSort().
* (2) Consequently, boxaSort() calls this function if it will
* likely go much faster.
*/
BOXA *
boxaBinSort(BOXA *boxas,
l_int32 sorttype,
l_int32 sortorder,
NUMA **pnaindex)
{
l_int32 i, n, x, y, w, h;
BOXA *boxad;
NUMA *na, *naindex;
PROCNAME("boxaBinSort");
if (pnaindex) *pnaindex = NULL;
if (!boxas)
return (BOXA *)ERROR_PTR("boxas not defined", procName, NULL);
if (sorttype != L_SORT_BY_X && sorttype != L_SORT_BY_Y &&
sorttype != L_SORT_BY_WIDTH && sorttype != L_SORT_BY_HEIGHT &&
sorttype != L_SORT_BY_PERIMETER)
return (BOXA *)ERROR_PTR("invalid sort type", procName, NULL);
if (sortorder != L_SORT_INCREASING && sortorder != L_SORT_DECREASING)
return (BOXA *)ERROR_PTR("invalid sort order", procName, NULL);
/* Generate Numa of appropriate box dimensions */
n = boxaGetCount(boxas);
if ((na = numaCreate(n)) == NULL)
return (BOXA *)ERROR_PTR("na not made", procName, NULL);
for (i = 0; i < n; i++) {
boxaGetBoxGeometry(boxas, i, &x, &y, &w, &h);
switch (sorttype)
{
case L_SORT_BY_X:
numaAddNumber(na, x);
break;
case L_SORT_BY_Y:
numaAddNumber(na, y);
break;
case L_SORT_BY_WIDTH:
numaAddNumber(na, w);
break;
case L_SORT_BY_HEIGHT:
numaAddNumber(na, h);
break;
case L_SORT_BY_PERIMETER:
numaAddNumber(na, w + h);
break;
default:
L_WARNING("invalid sort type", procName);
}
}
/* Get the sort index for data array */
if ((naindex = numaGetBinSortIndex(na, sortorder)) == NULL)
return (BOXA *)ERROR_PTR("naindex not made", procName, NULL);
/* Build up sorted boxa using the sort index */
boxad = boxaSortByIndex(boxas, naindex);
if (pnaindex)
*pnaindex = naindex;
else
numaDestroy(&naindex);
numaDestroy(&na);
return boxad;
}
/*!
* pixaBinSort()
*
* Input: pixas
* sorttype (L_SORT_BY_X, L_SORT_BY_Y, L_SORT_BY_WIDTH,
* L_SORT_BY_HEIGHT, L_SORT_BY_PERIMETER)
* sortorder (L_SORT_INCREASING, L_SORT_DECREASING)
* &naindex (<optional return> index of sorted order into
* original array)
* copyflag (L_COPY, L_CLONE)
* Return: pixad (sorted version of pixas), or null on error
*
* Notes:
* (1) This sorts based on the data in the boxa. If the boxa
* count is not the same as the pixa count, this returns an error.
* (2) The copyflag refers to the pix and box copies that are
* inserted into the sorted pixa. These are either L_COPY
* or L_CLONE.
* (3) For a large number of boxes (say, greater than 1000), this
* O(n) binsort is much faster than the O(nlogn) shellsort.
* For 5000 components, this is over 20x faster than boxaSort().
* (4) Consequently, pixaSort() calls this function if it will
* likely go much faster.
*/
PIXA *
pixaBinSort(PIXA *pixas,
l_int32 sorttype,
l_int32 sortorder,
NUMA **pnaindex,
l_int32 copyflag)
{
l_int32 i, n, x, y, w, h;
BOXA *boxa;
NUMA *na, *naindex;
PIXA *pixad;
PROCNAME("pixaBinSort");
if (pnaindex) *pnaindex = NULL;
if (!pixas)
return (PIXA *)ERROR_PTR("pixas not defined", procName, NULL);
if (sorttype != L_SORT_BY_X && sorttype != L_SORT_BY_Y &&
sorttype != L_SORT_BY_WIDTH && sorttype != L_SORT_BY_HEIGHT &&
sorttype != L_SORT_BY_PERIMETER)
return (PIXA *)ERROR_PTR("invalid sort type", procName, NULL);
if (sortorder != L_SORT_INCREASING && sortorder != L_SORT_DECREASING)
return (PIXA *)ERROR_PTR("invalid sort order", procName, NULL);
if (copyflag != L_COPY && copyflag != L_CLONE)
return (PIXA *)ERROR_PTR("invalid copy flag", procName, NULL);
/* Verify that the pixa and its boxa have the same count */
if ((boxa = pixas->boxa) == NULL) /* not owned; do not destroy */
return (PIXA *)ERROR_PTR("boxa not found", procName, NULL);
n = pixaGetCount(pixas);
if (boxaGetCount(boxa) != n)
return (PIXA *)ERROR_PTR("boxa and pixa counts differ", procName, NULL);
/* Generate Numa of appropriate box dimensions */
if ((na = numaCreate(n)) == NULL)
return (PIXA *)ERROR_PTR("na not made", procName, NULL);
for (i = 0; i < n; i++) {
boxaGetBoxGeometry(boxa, i, &x, &y, &w, &h);
switch (sorttype)
{
case L_SORT_BY_X:
numaAddNumber(na, x);
break;
case L_SORT_BY_Y:
numaAddNumber(na, y);
break;
case L_SORT_BY_WIDTH:
numaAddNumber(na, w);
break;
case L_SORT_BY_HEIGHT:
numaAddNumber(na, h);
break;
case L_SORT_BY_PERIMETER:
numaAddNumber(na, w + h);
break;
default:
L_WARNING("invalid sort type", procName);
}
}
/* Get the sort index for data array */
if ((naindex = numaGetBinSortIndex(na, sortorder)) == NULL)
return (PIXA *)ERROR_PTR("naindex not made", procName, NULL);
/* Build up sorted pixa using sort index */
if ((pixad = pixaSortByIndex(pixas, naindex, copyflag)) == NULL)
return (PIXA *)ERROR_PTR("pixad not made", procName, NULL);
if (pnaindex)
*pnaindex = naindex;
else
numaDestroy(&naindex);
numaDestroy(&na);
return pixad;
}
/*!
* pixaSort()
*
* Input: pixas
* sorttype (L_SORT_BY_X, L_SORT_BY_Y, L_SORT_BY_WIDTH,
* L_SORT_BY_HEIGHT, L_SORT_BY_MIN_DIMENSION,
* L_SORT_BY_MAX_DIMENSION, L_SORT_BY_PERIMETER,
* L_SORT_BY_AREA, L_SORT_BY_ASPECT_RATIO)
* sortorder (L_SORT_INCREASING, L_SORT_DECREASING)
* &naindex (<optional return> index of sorted order into
* original array)
* copyflag (L_COPY, L_CLONE)
* Return: pixad (sorted version of pixas), or null on error
*
* Notes:
* (1) This sorts based on the data in the boxa. If the boxa
* count is not the same as the pixa count, this returns an error.
* (2) The copyflag refers to the pix and box copies that are
* inserted into the sorted pixa. These are either L_COPY
* or L_CLONE.
*/
PIXA *
pixaSort(PIXA *pixas,
l_int32 sorttype,
l_int32 sortorder,
NUMA **pnaindex,
l_int32 copyflag)
{
l_int32 i, n, x, y, w, h;
BOXA *boxa;
NUMA *na, *naindex;
PIXA *pixad;
PROCNAME("pixaSort");
if (pnaindex) *pnaindex = NULL;
if (!pixas)
return (PIXA *)ERROR_PTR("pixas not defined", procName, NULL);
if (sorttype != L_SORT_BY_X && sorttype != L_SORT_BY_Y &&
sorttype != L_SORT_BY_WIDTH && sorttype != L_SORT_BY_HEIGHT &&
sorttype != L_SORT_BY_MIN_DIMENSION &&
sorttype != L_SORT_BY_MAX_DIMENSION &&
sorttype != L_SORT_BY_PERIMETER &&
sorttype != L_SORT_BY_AREA &&
sorttype != L_SORT_BY_ASPECT_RATIO)
return (PIXA *)ERROR_PTR("invalid sort type", procName, NULL);
if (sortorder != L_SORT_INCREASING && sortorder != L_SORT_DECREASING)
return (PIXA *)ERROR_PTR("invalid sort order", procName, NULL);
if (copyflag != L_COPY && copyflag != L_CLONE)
return (PIXA *)ERROR_PTR("invalid copy flag", procName, NULL);
if ((boxa = pixas->boxa) == NULL) /* not owned; do not destroy */
return (PIXA *)ERROR_PTR("boxa not found", procName, NULL);
n = pixaGetCount(pixas);
if (boxaGetCount(boxa) != n)
return (PIXA *)ERROR_PTR("boxa and pixa counts differ", procName, NULL);
/* Use O(n) binsort if possible */
if (n > MIN_COMPS_FOR_BIN_SORT &&
((sorttype == L_SORT_BY_X) || (sorttype == L_SORT_BY_Y) ||
(sorttype == L_SORT_BY_WIDTH) || (sorttype == L_SORT_BY_HEIGHT) ||
(sorttype == L_SORT_BY_PERIMETER)))
return pixaBinSort(pixas, sorttype, sortorder, pnaindex, copyflag);
/* Build up numa of specific data */
if ((na = numaCreate(n)) == NULL)
return (PIXA *)ERROR_PTR("na not made", procName, NULL);
for (i = 0; i < n; i++) {
boxaGetBoxGeometry(boxa, i, &x, &y, &w, &h);
switch (sorttype)
{
case L_SORT_BY_X:
numaAddNumber(na, x);
break;
case L_SORT_BY_Y:
numaAddNumber(na, y);
break;
case L_SORT_BY_WIDTH:
numaAddNumber(na, w);
break;
case L_SORT_BY_HEIGHT:
numaAddNumber(na, h);
break;
case L_SORT_BY_MIN_DIMENSION:
numaAddNumber(na, L_MIN(w, h));
break;
case L_SORT_BY_MAX_DIMENSION:
numaAddNumber(na, L_MAX(w, h));
break;
case L_SORT_BY_PERIMETER:
numaAddNumber(na, w + h);
break;
case L_SORT_BY_AREA:
numaAddNumber(na, w * h);
break;
case L_SORT_BY_ASPECT_RATIO:
numaAddNumber(na, (l_float32)w / (l_float32)h);
break;
default:
L_WARNING("invalid sort type", procName);
}
}
/* Get the sort index for data array */
if ((naindex = numaGetSortIndex(na, sortorder)) == NULL)
return (PIXA *)ERROR_PTR("naindex not made", procName, NULL);
/* Build up sorted pixa using sort index */
if ((pixad = pixaSortByIndex(pixas, naindex, copyflag)) == NULL)
return (PIXA *)ERROR_PTR("pixad not made", procName, NULL);
if (pnaindex)
*pnaindex = naindex;
else
numaDestroy(&naindex);
numaDestroy(&na);
return pixad;
}
/*!
* pixaaDisplay()
*
* Input: pixaa
* w, h (if set to 0, determines the size from the
* b.b. of the components in pixaa)
* Return: pix, or null on error
*
* Notes:
* (1) Each pix of the pixaa is displayed at the location given by
* its box, translated by the box of the containing pixa
* if it exists.
*/
PIX *
pixaaDisplay(PIXAA *pixaa,
l_int32 w,
l_int32 h)
{
l_int32 i, j, n, nbox, na, d, wmax, hmax, x, y, xb, yb, wb, hb;
BOXA *boxa1; /* top-level boxa */
BOXA *boxa;
PIX *pixt, *pixd;
PIXA *pixa;
PROCNAME("pixaaDisplay");
if (!pixaa)
return (PIX *)ERROR_PTR("pixaa not defined", procName, NULL);
n = pixaaGetCount(pixaa);
if (n == 0)
return (PIX *)ERROR_PTR("no components", procName, NULL);
/* If w and h not input, determine the minimum size required
* to contain the origin and all c.c. */
boxa1 = pixaaGetBoxa(pixaa, L_CLONE);
nbox = boxaGetCount(boxa1);
if (w == 0 || h == 0) {
if (nbox == n)
boxaGetExtent(boxa1, &w, &h, NULL);
else { /* have to use the lower-level boxa for each pixa */
wmax = hmax = 0;
for (i = 0; i < n; i++) {
pixa = pixaaGetPixa(pixaa, i, L_CLONE);
boxa = pixaGetBoxa(pixa, L_CLONE);
boxaGetExtent(boxa, &w, &h, NULL);
wmax = L_MAX(wmax, w);
hmax = L_MAX(hmax, h);
pixaDestroy(&pixa);
boxaDestroy(&boxa);
}
w = wmax;
h = hmax;
}
}
/* Get depth from first pix */
pixa = pixaaGetPixa(pixaa, 0, L_CLONE);
pixt = pixaGetPix(pixa, 0, L_CLONE);
d = pixGetDepth(pixt);
pixaDestroy(&pixa);
pixDestroy(&pixt);
if ((pixd = pixCreate(w, h, d)) == NULL)
return (PIX *)ERROR_PTR("pixd not made", procName, NULL);
x = y = 0;
for (i = 0; i < n; i++) {
pixa = pixaaGetPixa(pixaa, i, L_CLONE);
if (nbox == n)
boxaGetBoxGeometry(boxa1, i, &x, &y, NULL, NULL);
na = pixaGetCount(pixa);
for (j = 0; j < na; j++) {
pixaGetBoxGeometry(pixa, j, &xb, &yb, &wb, &hb);
pixt = pixaGetPix(pixa, j, L_CLONE);
pixRasterop(pixd, x + xb, y + yb, wb, hb, PIX_PAINT, pixt, 0, 0);
pixDestroy(&pixt);
}
pixaDestroy(&pixa);
}
boxaDestroy(&boxa1);
return pixd;
}
main(int argc,
char **argv)
{
l_int32 i, n, w, h, success, display;
FILE *fp;
BOXA *boxa;
NUMA *naindex, *naw, *nah, *naw_med, *nah_med;
PIX *pixs, *pixt, *pixd;
if (regTestSetup(argc, argv, &fp, &display, &success, NULL))
return 1;
/* Generate arrays of word widths and heights */
pixs = pixRead("feyn.tif");
pixGetWordBoxesInTextlines(pixs, 1, 6, 6, 500, 50, &boxa, &naindex);
n = boxaGetCount(boxa);
naw = numaCreate(0);
nah = numaCreate(0);
for (i = 0; i < n; i++) {
boxaGetBoxGeometry(boxa, i, NULL, NULL, &w, &h);
numaAddNumber(naw, w);
numaAddNumber(nah, h);
}
boxaDestroy(&boxa);
numaDestroy(&naindex);
/* Make the rank bin arrays of median values, with 10 bins */
numaGetRankBinValues(naw, 10, NULL, &naw_med);
numaGetRankBinValues(nah, 10, NULL, &nah_med);
gplotSimple1(naw_med, GPLOT_PNG, "/tmp/w_10bin",
"width vs rank bins (10)");
gplotSimple1(nah_med, GPLOT_PNG, "/tmp/h_10bin",
"height vs rank bins (10)");
numaDestroy(&naw_med);
numaDestroy(&nah_med);
/* Make the rank bin arrays of median values, with 30 bins */
numaGetRankBinValues(naw, 30, NULL, &naw_med);
numaGetRankBinValues(nah, 30, NULL, &nah_med);
gplotSimple1(naw_med, GPLOT_PNG, "/tmp/w_30bin",
"width vs rank bins (30)");
gplotSimple1(nah_med, GPLOT_PNG, "/tmp/h_30bin",
"height vs rank bins (30)");
numaDestroy(&naw_med);
numaDestroy(&nah_med);
/* Give gnuplot time to write out the files */
#ifndef _WIN32
sleep(2);
#else
Sleep(2000);
#endif /* _WIN32 */
/* Save as golden files, or check against them */
regTestCheckFile(fp, argv, "/tmp/w_10bin.png", 0, &success);
regTestCheckFile(fp, argv, "/tmp/h_10bin.png", 1, &success);
regTestCheckFile(fp, argv, "/tmp/w_30bin.png", 2, &success);
regTestCheckFile(fp, argv, "/tmp/h_30bin.png", 3, &success);
/* Display results for debugging */
pixt = pixRead("/tmp/w_10bin.png");
pixDisplayWithTitle(pixt, 0, 0, NULL, display);
pixDestroy(&pixt);
pixt = pixRead("/tmp/h_10bin.png");
pixDisplayWithTitle(pixt, 650, 0, NULL, display);
pixDestroy(&pixt);
pixt = pixRead("/tmp/w_30bin.png");
pixDisplayWithTitle(pixt, 0, 550, NULL, display);
pixDestroy(&pixt);
pixt = pixRead("/tmp/h_30bin.png");
pixDisplayWithTitle(pixt, 650, 550, NULL, display);
pixDestroy(&pixt);
pixDestroy(&pixs);
numaDestroy(&naw);
numaDestroy(&nah);
regTestCleanup(argc, argv, fp, success, NULL);
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
}
/*!
* \brief boxaMakeSizeIndicator()
*
* \param[in] boxa
* \param[in] width, height threshold dimensions
* \param[in] type L_SELECT_WIDTH, L_SELECT_HEIGHT,
* L_SELECT_IF_EITHER, L_SELECT_IF_BOTH
* \param[in] relation L_SELECT_IF_LT, L_SELECT_IF_GT,
* L_SELECT_IF_LTE, L_SELECT_IF_GTE
* \return na indicator array, or NULL on error
*
* <pre>
* Notes:
* (1) The args specify constraints on the size of the
* components that are kept.
* (2) If the selection type is L_SELECT_WIDTH, the input
* height is ignored, and v.v.
* (3) To keep small components, use relation = L_SELECT_IF_LT or
* L_SELECT_IF_LTE.
* To keep large components, use relation = L_SELECT_IF_GT or
* L_SELECT_IF_GTE.
* </pre>
*/
NUMA *
boxaMakeSizeIndicator(BOXA *boxa,
l_int32 width,
l_int32 height,
l_int32 type,
l_int32 relation)
{
l_int32 i, n, w, h, ival;
NUMA *na;
PROCNAME("boxaMakeSizeIndicator");
if (!boxa)
return (NUMA *)ERROR_PTR("boxa not defined", procName, NULL);
if ((n = boxaGetCount(boxa)) == 0)
return (NUMA *)ERROR_PTR("boxa is empty", procName, NULL);
if (type != L_SELECT_WIDTH && type != L_SELECT_HEIGHT &&
type != L_SELECT_IF_EITHER && type != L_SELECT_IF_BOTH)
return (NUMA *)ERROR_PTR("invalid type", procName, NULL);
if (relation != L_SELECT_IF_LT && relation != L_SELECT_IF_GT &&
relation != L_SELECT_IF_LTE && relation != L_SELECT_IF_GTE)
return (NUMA *)ERROR_PTR("invalid relation", procName, NULL);
na = numaCreate(n);
for (i = 0; i < n; i++) {
ival = 0;
boxaGetBoxGeometry(boxa, i, NULL, NULL, &w, &h);
switch (type)
{
case L_SELECT_WIDTH:
if ((relation == L_SELECT_IF_LT && w < width) ||
(relation == L_SELECT_IF_GT && w > width) ||
(relation == L_SELECT_IF_LTE && w <= width) ||
(relation == L_SELECT_IF_GTE && w >= width))
ival = 1;
break;
case L_SELECT_HEIGHT:
if ((relation == L_SELECT_IF_LT && h < height) ||
(relation == L_SELECT_IF_GT && h > height) ||
(relation == L_SELECT_IF_LTE && h <= height) ||
(relation == L_SELECT_IF_GTE && h >= height))
ival = 1;
break;
case L_SELECT_IF_EITHER:
if (((relation == L_SELECT_IF_LT) && (w < width || h < height)) ||
((relation == L_SELECT_IF_GT) && (w > width || h > height)) ||
((relation == L_SELECT_IF_LTE) && (w <= width || h <= height)) ||
((relation == L_SELECT_IF_GTE) && (w >= width || h >= height)))
ival = 1;
break;
case L_SELECT_IF_BOTH:
if (((relation == L_SELECT_IF_LT) && (w < width && h < height)) ||
((relation == L_SELECT_IF_GT) && (w > width && h > height)) ||
((relation == L_SELECT_IF_LTE) && (w <= width && h <= height)) ||
((relation == L_SELECT_IF_GTE) && (w >= width && h >= height)))
ival = 1;
break;
default:
L_WARNING("can't get here!\n", procName);
break;
}
numaAddNumber(na, ival);
}
return na;
}
int main(int argc,
char **argv)
{
l_int32 i, n, w, h;
BOXA *boxa;
NUMA *naindex, *naw, *nah, *naw_med, *nah_med;
PIX *pixs, *pixt;
L_REGPARAMS *rp;
if (regTestSetup(argc, argv, &rp))
return 1;
/* Generate arrays of word widths and heights */
pixs = pixRead("feyn.tif");
pixGetWordBoxesInTextlines(pixs, 1, 6, 6, 500, 50, &boxa, &naindex);
n = boxaGetCount(boxa);
naw = numaCreate(0);
nah = numaCreate(0);
for (i = 0; i < n; i++) {
boxaGetBoxGeometry(boxa, i, NULL, NULL, &w, &h);
numaAddNumber(naw, w);
numaAddNumber(nah, h);
}
boxaDestroy(&boxa);
numaDestroy(&naindex);
/* Make the rank bin arrays of median values, with 10 bins */
lept_rmfile("/tmp/lept/regout/w_10bin.png"); /* remove existing ones */
lept_rmfile("/tmp/lept/regout/h_10bin.png");
lept_rmfile("/tmp/lept/regout/w_30bin.png");
lept_rmfile("/tmp/lept/regout/h_30bin.png");
numaGetRankBinValues(naw, 10, NULL, &naw_med);
numaGetRankBinValues(nah, 10, NULL, &nah_med);
gplotSimple1(naw_med, GPLOT_PNG, "/tmp/lept/regout/w_10bin",
"width vs rank bins (10)");
gplotSimple1(nah_med, GPLOT_PNG, "/tmp/lept/regout/h_10bin",
"height vs rank bins (10)");
numaDestroy(&naw_med);
numaDestroy(&nah_med);
/* Make the rank bin arrays of median values, with 30 bins */
numaGetRankBinValues(naw, 30, NULL, &naw_med);
numaGetRankBinValues(nah, 30, NULL, &nah_med);
gplotSimple1(naw_med, GPLOT_PNG, "/tmp/lept/regout/w_30bin",
"width vs rank bins (30)");
gplotSimple1(nah_med, GPLOT_PNG, "/tmp/lept/regout/h_30bin",
"height vs rank bins (30)");
numaDestroy(&naw_med);
numaDestroy(&nah_med);
/* Save as golden files, or check against them */
regTestCheckFile(rp, "/tmp/lept/regout/w_10bin.png"); /* 0 */
regTestCheckFile(rp, "/tmp/lept/regout/h_10bin.png"); /* 1 */
regTestCheckFile(rp, "/tmp/lept/regout/w_30bin.png"); /* 2 */
regTestCheckFile(rp, "/tmp/lept/regout/h_30bin.png"); /* 3 */
/* Display results for debugging */
pixt = pixRead("/tmp/lept/regout/w_10bin.png");
pixDisplayWithTitle(pixt, 0, 0, NULL, rp->display);
pixDestroy(&pixt);
pixt = pixRead("/tmp/lept/regout/h_10bin.png");
pixDisplayWithTitle(pixt, 650, 0, NULL, rp->display);
pixDestroy(&pixt);
pixt = pixRead("/tmp/lept/regout/w_30bin.png");
pixDisplayWithTitle(pixt, 0, 550, NULL, rp->display);
pixDestroy(&pixt);
pixt = pixRead("/tmp/lept/regout/h_30bin.png");
pixDisplayWithTitle(pixt, 650, 550, NULL, rp->display);
pixDestroy(&pixt);
pixDestroy(&pixs);
numaDestroy(&naw);
numaDestroy(&nah);
return regTestCleanup(rp);
}
