third-party/leptonica/src/binarize.c
/*====================================================================*
- Copyright (C) 2001 Leptonica. All rights reserved.
-
- Redistribution and use in source and binary forms, with or without
- modification, are permitted provided that the following conditions
- are met:
- 1. Redistributions of source code must retain the above copyright
- notice, this list of conditions and the following disclaimer.
- 2. Redistributions in binary form must reproduce the above
- copyright notice, this list of conditions and the following
- disclaimer in the documentation and/or other materials
- provided with the distribution.
-
- THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
- ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
- LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
- A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL ANY
- CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
- EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
- PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
- PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
- OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
- NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
- SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*====================================================================*/
/*
* binarize.c
*
* ===================================================================
* Image binarization algorithms are found in:
* grayquant.c: standard, simple, general grayscale quantization
* adaptmap.c: local adaptive; mostly gray-to-gray in preparation
* for binarization
* binarize.c: special binarization methods, locally adaptive and
* global.
* ===================================================================
*
* Adaptive Otsu-based thresholding
* l_int32 pixOtsuAdaptiveThreshold() 8 bpp
*
* Otsu thresholding on adaptive background normalization
* PIX *pixOtsuThreshOnBackgroundNorm() 8 bpp
*
* Masking and Otsu estimate on adaptive background normalization
* PIX *pixMaskedThreshOnBackgroundNorm() 8 bpp
*
* Sauvola local thresholding
* l_int32 pixSauvolaBinarizeTiled()
* l_int32 pixSauvolaBinarize()
* PIX *pixSauvolaGetThreshold()
* PIX *pixApplyLocalThreshold();
*
* Thresholding using connected components
* PIX *pixThresholdByConnComp()
*
* Notes:
* (1) pixOtsuAdaptiveThreshold() computes a global threshold over each
* tile and performs the threshold operation, resulting in a
* binary image for each tile. These are stitched into the
* final result.
* (2) pixOtsuThreshOnBackgroundNorm() and
* pixMaskedThreshOnBackgroundNorm() are binarization functions
* that use background normalization with other techniques.
* (3) Sauvola binarization computes a local threshold based on
* the local average and square average. It takes two constants:
* the window size for the measurment at each pixel and a
* parameter that determines the amount of normalized local
* standard deviation to subtract from the local average value.
* (4) pixThresholdByCC() uses the numbers of 4 and 8 connected
* components at different thresholding to determine if a
* global threshold can be used (for text or line-art) and the
* value it should have.
*/
#include <math.h>
#include "allheaders.h"
/*------------------------------------------------------------------*
* Adaptive Otsu-based thresholding *
*------------------------------------------------------------------*/
/*!
* pixOtsuAdaptiveThreshold()
*
* Input: pixs (8 bpp)
* sx, sy (desired tile dimensions; actual size may vary)
* smoothx, smoothy (half-width of convolution kernel applied to
* threshold array: use 0 for no smoothing)
* scorefract (fraction of the max Otsu score; typ. 0.1;
* use 0.0 for standard Otsu)
* &pixth (<optional return> array of threshold values
* found for each tile)
* &pixd (<optional return> thresholded input pixs, based on
* the threshold array)
* Return: 0 if OK, 1 on error
*
* Notes:
* (1) The Otsu method finds a single global threshold for an image.
* This function allows a locally adapted threshold to be
* found for each tile into which the image is broken up.
* (2) The array of threshold values, one for each tile, constitutes
* a highly downscaled image. This array is optionally
* smoothed using a convolution. The full width and height of the
* convolution kernel are (2 * @smoothx + 1) and (2 * @smoothy + 1).
* (3) The minimum tile dimension allowed is 16. If such small
* tiles are used, it is recommended to use smoothing, because
* without smoothing, each small tile determines the splitting
* threshold independently. A tile that is entirely in the
* image bg will then hallucinate fg, resulting in a very noisy
* binarization. The smoothing should be large enough that no
* tile is only influenced by one type (fg or bg) of pixels,
* because it will force a split of its pixels.
* (4) To get a single global threshold for the entire image, use
* input values of @sx and @sy that are larger than the image.
* For this situation, the smoothing parameters are ignored.
* (5) The threshold values partition the image pixels into two classes:
* one whose values are less than the threshold and another
* whose values are greater than or equal to the threshold.
* This is the same use of 'threshold' as in pixThresholdToBinary().
* (6) The scorefract is the fraction of the maximum Otsu score, which
* is used to determine the range over which the histogram minimum
* is searched. See numaSplitDistribution() for details on the
* underlying method of choosing a threshold.
* (7) This uses enables a modified version of the Otsu criterion for
* splitting the distribution of pixels in each tile into a
* fg and bg part. The modification consists of searching for
* a minimum in the histogram over a range of pixel values where
* the Otsu score is within a defined fraction, @scorefract,
* of the max score. To get the original Otsu algorithm, set
* @scorefract == 0.
* (8) N.B. This method is NOT recommended for images with weak text
* and significant background noise, such as bleedthrough, because
* of the problem noted in (3) above for tiling. Use Sauvola.
*/
l_int32
pixOtsuAdaptiveThreshold(PIX *pixs,
l_int32 sx,
l_int32 sy,
l_int32 smoothx,
l_int32 smoothy,
l_float32 scorefract,
PIX **ppixth,
PIX **ppixd)
{
l_int32 w, h, nx, ny, i, j, thresh;
l_uint32 val;
PIX *pixt, *pixb, *pixthresh, *pixth, *pixd;
PIXTILING *pt;
PROCNAME("pixOtsuAdaptiveThreshold");
if (!ppixth && !ppixd)
return ERROR_INT("neither &pixth nor &pixd defined", procName, 1);
if (ppixth) *ppixth = NULL;
if (ppixd) *ppixd = NULL;
if (!pixs || pixGetDepth(pixs) != 8)
return ERROR_INT("pixs not defined or not 8 bpp", procName, 1);
if (sx < 16 || sy < 16)
return ERROR_INT("sx and sy must be >= 16", procName, 1);
/* Compute the threshold array for the tiles */
pixGetDimensions(pixs, &w, &h, NULL);
nx = L_MAX(1, w / sx);
ny = L_MAX(1, h / sy);
smoothx = L_MIN(smoothx, (nx - 1) / 2);
smoothy = L_MIN(smoothy, (ny - 1) / 2);
pt = pixTilingCreate(pixs, nx, ny, 0, 0, 0, 0);
pixthresh = pixCreate(nx, ny, 8);
for (i = 0; i < ny; i++) {
for (j = 0; j < nx; j++) {
pixt = pixTilingGetTile(pt, i, j);
pixSplitDistributionFgBg(pixt, scorefract, 1, &thresh,
NULL, NULL, 0);
pixSetPixel(pixthresh, j, i, thresh); /* see note (4) */
pixDestroy(&pixt);
}
}
/* Optionally smooth the threshold array */
if (smoothx > 0 || smoothy > 0)
pixth = pixBlockconv(pixthresh, smoothx, smoothy);
else
pixth = pixClone(pixthresh);
pixDestroy(&pixthresh);
/* Optionally apply the threshold array to binarize pixs */
if (ppixd) {
pixd = pixCreate(w, h, 1);
pixCopyResolution(pixd, pixs);
for (i = 0; i < ny; i++) {
for (j = 0; j < nx; j++) {
pixt = pixTilingGetTile(pt, i, j);
pixGetPixel(pixth, j, i, &val);
pixb = pixThresholdToBinary(pixt, val);
pixTilingPaintTile(pixd, i, j, pixb, pt);
pixDestroy(&pixt);
pixDestroy(&pixb);
}
}
*ppixd = pixd;
}
if (ppixth)
*ppixth = pixth;
else
pixDestroy(&pixth);
pixTilingDestroy(&pt);
return 0;
}
/*------------------------------------------------------------------*
* Otsu thresholding on adaptive background normalization *
*------------------------------------------------------------------*/
/*!
* pixOtsuThreshOnBackgroundNorm()
*
* Input: pixs (8 bpp grayscale; not colormapped)
* pixim (<optional> 1 bpp 'image' mask; can be null)
* sx, sy (tile size in pixels)
* thresh (threshold for determining foreground)
* mincount (min threshold on counts in a tile)
* bgval (target bg val; typ. > 128)
* smoothx (half-width of block convolution kernel width)
* smoothy (half-width of block convolution kernel height)
* scorefract (fraction of the max Otsu score; typ. 0.1)
* &thresh (<optional return> threshold value that was
* used on the normalized image)
* Return: pixd (1 bpp thresholded image), or null on error
*
* Notes:
* (1) This does background normalization followed by Otsu
* thresholding. Otsu binarization attempts to split the
* image into two roughly equal sets of pixels, and it does
* a very poor job when there are large amounts of dark
* background. By doing a background normalization first,
* to get the background near 255, we remove this problem.
* Then we use a modified Otsu to estimate the best global
* threshold on the normalized image.
* (2) See pixBackgroundNorm() for meaning and typical values
* of input parameters. For a start, you can try:
* sx, sy = 10, 15
* thresh = 100
* mincount = 50
* bgval = 255
* smoothx, smoothy = 2
*/
PIX *
pixOtsuThreshOnBackgroundNorm(PIX *pixs,
PIX *pixim,
l_int32 sx,
l_int32 sy,
l_int32 thresh,
l_int32 mincount,
l_int32 bgval,
l_int32 smoothx,
l_int32 smoothy,
l_float32 scorefract,
l_int32 *pthresh)
{
l_int32 w, h;
l_uint32 val;
PIX *pixn, *pixt, *pixd;
PROCNAME("pixOtsuThreshOnBackgroundNorm");
if (pthresh) *pthresh = 0;
if (!pixs || pixGetDepth(pixs) != 8)
return (PIX *)ERROR_PTR("pixs undefined or not 8 bpp", procName, NULL);
if (pixGetColormap(pixs))
return (PIX *)ERROR_PTR("pixs is colormapped", procName, NULL);
if (sx < 4 || sy < 4)
return (PIX *)ERROR_PTR("sx and sy must be >= 4", procName, NULL);
if (mincount > sx * sy) {
L_WARNING("mincount too large for tile size\n", procName);
mincount = (sx * sy) / 3;
}
pixn = pixBackgroundNorm(pixs, pixim, NULL, sx, sy, thresh,
mincount, bgval, smoothx, smoothy);
if (!pixn)
return (PIX *)ERROR_PTR("pixn not made", procName, NULL);
/* Just use 1 tile for a global threshold, which is stored
* as a single pixel in pixt. */
pixGetDimensions(pixn, &w, &h, NULL);
pixOtsuAdaptiveThreshold(pixn, w, h, 0, 0, scorefract, &pixt, &pixd);
pixDestroy(&pixn);
if (pixt && pthresh) {
pixGetPixel(pixt, 0, 0, &val);
*pthresh = val;
}
pixDestroy(&pixt);
if (!pixd)
return (PIX *)ERROR_PTR("pixd not made", procName, NULL);
else
return pixd;
}
/*----------------------------------------------------------------------*
* Masking and Otsu estimate on adaptive background normalization *
*----------------------------------------------------------------------*/
/*!
* pixMaskedThreshOnBackgroundNorm()
*
* Input: pixs (8 bpp grayscale; not colormapped)
* pixim (<optional> 1 bpp 'image' mask; can be null)
* sx, sy (tile size in pixels)
* thresh (threshold for determining foreground)
* mincount (min threshold on counts in a tile)
* smoothx (half-width of block convolution kernel width)
* smoothy (half-width of block convolution kernel height)
* scorefract (fraction of the max Otsu score; typ. ~ 0.1)
* &thresh (<optional return> threshold value that was
* used on the normalized image)
* Return: pixd (1 bpp thresholded image), or null on error
*
* Notes:
* (1) This begins with a standard background normalization.
* Additionally, there is a flexible background norm, that
* will adapt to a rapidly varying background, and this
* puts white pixels in the background near regions with
* significant foreground. The white pixels are turned into
* a 1 bpp selection mask by binarization followed by dilation.
* Otsu thresholding is performed on the input image to get an
* estimate of the threshold in the non-mask regions.
* The background normalized image is thresholded with two
* different values, and the result is combined using
* the selection mask.
* (2) Note that the numbers 255 (for bgval target) and 190 (for
* thresholding on pixn) are tied together, and explicitly
* defined in this function.
* (3) See pixBackgroundNorm() for meaning and typical values
* of input parameters. For a start, you can try:
* sx, sy = 10, 15
* thresh = 100
* mincount = 50
* smoothx, smoothy = 2
*/
PIX *
pixMaskedThreshOnBackgroundNorm(PIX *pixs,
PIX *pixim,
l_int32 sx,
l_int32 sy,
l_int32 thresh,
l_int32 mincount,
l_int32 smoothx,
l_int32 smoothy,
l_float32 scorefract,
l_int32 *pthresh)
{
l_int32 w, h;
l_uint32 val;
PIX *pixn, *pixm, *pixd, *pixt1, *pixt2, *pixt3, *pixt4;
PROCNAME("pixMaskedThreshOnBackgroundNorm");
if (pthresh) *pthresh = 0;
if (!pixs || pixGetDepth(pixs) != 8)
return (PIX *)ERROR_PTR("pixs undefined or not 8 bpp", procName, NULL);
if (pixGetColormap(pixs))
return (PIX *)ERROR_PTR("pixs is colormapped", procName, NULL);
if (sx < 4 || sy < 4)
return (PIX *)ERROR_PTR("sx and sy must be >= 4", procName, NULL);
if (mincount > sx * sy) {
L_WARNING("mincount too large for tile size\n", procName);
mincount = (sx * sy) / 3;
}
/* Standard background normalization */
pixn = pixBackgroundNorm(pixs, pixim, NULL, sx, sy, thresh,
mincount, 255, smoothx, smoothy);
if (!pixn)
return (PIX *)ERROR_PTR("pixn not made", procName, NULL);
/* Special background normalization for adaptation to quickly
* varying background. Threshold on the very light parts,
* which tend to be near significant edges, and dilate to
* form a mask over regions that are typically text. The
* dilation size is chosen to cover the text completely,
* except for very thick fonts. */
pixt1 = pixBackgroundNormFlex(pixs, 7, 7, 1, 1, 20);
pixt2 = pixThresholdToBinary(pixt1, 240);
pixInvert(pixt2, pixt2);
pixm = pixMorphSequence(pixt2, "d21.21", 0);
pixDestroy(&pixt1);
pixDestroy(&pixt2);
/* Use Otsu to get a global threshold estimate for the image,
* which is stored as a single pixel in pixt3. */
pixGetDimensions(pixs, &w, &h, NULL);
pixOtsuAdaptiveThreshold(pixs, w, h, 0, 0, scorefract, &pixt3, NULL);
if (pixt3 && pthresh) {
pixGetPixel(pixt3, 0, 0, &val);
*pthresh = val;
}
pixDestroy(&pixt3);
/* Threshold the background normalized images differentially,
* using a high value correlated with the background normalization
* for the part of the image under the mask (i.e., near the
* darker, thicker foreground), and a value that depends on the Otsu
* threshold for the rest of the image. This gives a solid
* (high) thresholding for the foreground parts of the image,
* while allowing the background and light foreground to be
* reasonably well cleaned using a threshold adapted to the
* input image. */
pixd = pixThresholdToBinary(pixn, val + 30); /* for bg and light fg */
pixt4 = pixThresholdToBinary(pixn, 190); /* for heavier fg */
pixCombineMasked(pixd, pixt4, pixm);
pixDestroy(&pixt4);
pixDestroy(&pixm);
pixDestroy(&pixn);
if (!pixd)
return (PIX *)ERROR_PTR("pixd not made", procName, NULL);
else
return pixd;
}
/*----------------------------------------------------------------------*
* Sauvola binarization *
*----------------------------------------------------------------------*/
/*!
* pixSauvolaBinarizeTiled()
*
* Input: pixs (8 bpp grayscale, not colormapped)
* whsize (window half-width for measuring local statistics)
* factor (factor for reducing threshold due to variance; >= 0)
* nx, ny (subdivision into tiles; >= 1)
* &pixth (<optional return> Sauvola threshold values)
* &pixd (<optional return> thresholded image)
* Return: 0 if OK, 1 on error
*
* Notes:
* (1) The window width and height are 2 * @whsize + 1. The minimum
* value for @whsize is 2; typically it is >= 7..
* (2) For nx == ny == 1, this defaults to pixSauvolaBinarize().
* (3) Why a tiled version?
* (a) Because the mean value accumulator is a uint32, overflow
* can occur for an image with more than 16M pixels.
* (b) The mean value accumulator array for 16M pixels is 64 MB.
* The mean square accumulator array for 16M pixels is 128 MB.
* Using tiles reduces the size of these arrays.
* (c) Each tile can be processed independently, in parallel,
* on a multicore processor.
* (4) The Sauvola threshold is determined from the formula:
* t = m * (1 - k * (1 - s / 128))
* See pixSauvolaBinarize() for details.
*/
l_int32
pixSauvolaBinarizeTiled(PIX *pixs,
l_int32 whsize,
l_float32 factor,
l_int32 nx,
l_int32 ny,
PIX **ppixth,
PIX **ppixd)
{
l_int32 i, j, w, h, xrat, yrat;
PIX *pixth, *pixd, *tileth, *tiled, *pixt;
PIX **ptileth, **ptiled;
PIXTILING *pt;
PROCNAME("pixSauvolaBinarizeTiled");
if (!ppixth && !ppixd)
return ERROR_INT("no outputs", procName, 1);
if (ppixth) *ppixth = NULL;
if (ppixd) *ppixd = NULL;
if (!pixs || pixGetDepth(pixs) != 8)
return ERROR_INT("pixs undefined or not 8 bpp", procName, 1);
if (pixGetColormap(pixs))
return ERROR_INT("pixs is cmapped", procName, 1);
pixGetDimensions(pixs, &w, &h, NULL);
if (whsize < 2)
return ERROR_INT("whsize must be >= 2", procName, 1);
if (w < 2 * whsize + 3 || h < 2 * whsize + 3)
return ERROR_INT("whsize too large for image", procName, 1);
if (factor < 0.0)
return ERROR_INT("factor must be >= 0", procName, 1);
if (nx <= 1 && ny <= 1)
return pixSauvolaBinarize(pixs, whsize, factor, 1, NULL, NULL,
ppixth, ppixd);
/* Test to see if the tiles are too small. The required
* condition is that the tile dimensions must be at least
* (whsize + 2) x (whsize + 2). */
xrat = w / nx;
yrat = h / ny;
if (xrat < whsize + 2) {
nx = w / (whsize + 2);
L_WARNING("tile width too small; nx reduced to %d\n", procName, nx);
}
if (yrat < whsize + 2) {
ny = h / (whsize + 2);
L_WARNING("tile height too small; ny reduced to %d\n", procName, ny);
}
if (nx <= 1 && ny <= 1)
return pixSauvolaBinarize(pixs, whsize, factor, 1, NULL, NULL,
ppixth, ppixd);
/* We can use pixtiling for painting both outputs, if requested */
if (ppixth) {
pixth = pixCreateNoInit(w, h, 8);
*ppixth = pixth;
}
if (ppixd) {
pixd = pixCreateNoInit(w, h, 1);
*ppixd = pixd;
}
pt = pixTilingCreate(pixs, nx, ny, 0, 0, whsize + 1, whsize + 1);
pixTilingNoStripOnPaint(pt); /* pixSauvolaBinarize() does the stripping */
for (i = 0; i < ny; i++) {
for (j = 0; j < nx; j++) {
pixt = pixTilingGetTile(pt, i, j);
ptileth = (ppixth) ? &tileth : NULL;
ptiled = (ppixd) ? &tiled : NULL;
pixSauvolaBinarize(pixt, whsize, factor, 0, NULL, NULL,
ptileth, ptiled);
if (ppixth) { /* do not strip */
pixTilingPaintTile(pixth, i, j, tileth, pt);
pixDestroy(&tileth);
}
if (ppixd) {
pixTilingPaintTile(pixd, i, j, tiled, pt);
pixDestroy(&tiled);
}
pixDestroy(&pixt);
}
}
pixTilingDestroy(&pt);
return 0;
}
/*!
* pixSauvolaBinarize()
*
* Input: pixs (8 bpp grayscale; not colormapped)
* whsize (window half-width for measuring local statistics)
* factor (factor for reducing threshold due to variance; >= 0)
* addborder (1 to add border of width (@whsize + 1) on all sides)
* &pixm (<optional return> local mean values)
* &pixsd (<optional return> local standard deviation values)
* &pixth (<optional return> threshold values)
* &pixd (<optional return> thresholded image)
* Return: 0 if OK, 1 on error
*
* Notes:
* (1) The window width and height are 2 * @whsize + 1. The minimum
* value for @whsize is 2; typically it is >= 7..
* (2) The local statistics, measured over the window, are the
* average and standard deviation.
* (3) The measurements of the mean and standard deviation are
* performed inside a border of (@whsize + 1) pixels. If pixs does
* not have these added border pixels, use @addborder = 1 to add
* it here; otherwise use @addborder = 0.
* (4) The Sauvola threshold is determined from the formula:
* t = m * (1 - k * (1 - s / 128))
* where:
* t = local threshold
* m = local mean
* k = @factor (>= 0) [ typ. 0.35 ]
* s = local standard deviation, which is maximized at
* 127.5 when half the samples are 0 and half are 255.
* (5) The basic idea of Niblack and Sauvola binarization is that
* the local threshold should be less than the median value,
* and the larger the variance, the closer to the median
* it should be chosen. Typical values for k are between
* 0.2 and 0.5.
*/
l_int32
pixSauvolaBinarize(PIX *pixs,
l_int32 whsize,
l_float32 factor,
l_int32 addborder,
PIX **ppixm,
PIX **ppixsd,
PIX **ppixth,
PIX **ppixd)
{
l_int32 w, h;
PIX *pixg, *pixsc, *pixm, *pixms, *pixth, *pixd;
PROCNAME("pixSauvolaBinarize");
if (ppixm) *ppixm = NULL;
if (ppixsd) *ppixsd = NULL;
if (ppixth) *ppixth = NULL;
if (ppixd) *ppixd = NULL;
if (!ppixm && !ppixsd && !ppixth && !ppixd)
return ERROR_INT("no outputs", procName, 1);
if (!pixs || pixGetDepth(pixs) != 8)
return ERROR_INT("pixs undefined or not 8 bpp", procName, 1);
if (pixGetColormap(pixs))
return ERROR_INT("pixs is cmapped", procName, 1);
pixGetDimensions(pixs, &w, &h, NULL);
if (whsize < 2)
return ERROR_INT("whsize must be >= 2", procName, 1);
if (w < 2 * whsize + 3 || h < 2 * whsize + 3)
return ERROR_INT("whsize too large for image", procName, 1);
if (factor < 0.0)
return ERROR_INT("factor must be >= 0", procName, 1);
if (addborder) {
pixg = pixAddMirroredBorder(pixs, whsize + 1, whsize + 1,
whsize + 1, whsize + 1);
pixsc = pixClone(pixs);
} else {
pixg = pixClone(pixs);
pixsc = pixRemoveBorder(pixs, whsize + 1);
}
if (!pixg || !pixsc)
return ERROR_INT("pixg and pixsc not made", procName, 1);
/* All these functions strip off the border pixels. */
if (ppixm || ppixth || ppixd)
pixm = pixWindowedMean(pixg, whsize, whsize, 1, 1);
if (ppixsd || ppixth || ppixd)
pixms = pixWindowedMeanSquare(pixg, whsize, whsize, 1);
if (ppixth || ppixd)
pixth = pixSauvolaGetThreshold(pixm, pixms, factor, ppixsd);
if (ppixd) {
pixd = pixApplyLocalThreshold(pixsc, pixth, 1);
pixCopyResolution(pixd, pixs);
}
if (ppixm)
*ppixm = pixm;
else
pixDestroy(&pixm);
pixDestroy(&pixms);
if (ppixth)
*ppixth = pixth;
else
pixDestroy(&pixth);
if (ppixd)
*ppixd = pixd;
else
pixDestroy(&pixd);
pixDestroy(&pixg);
pixDestroy(&pixsc);
return 0;
}
/*!
* pixSauvolaGetThreshold()
*
* Input: pixm (8 bpp grayscale; not colormapped)
* pixms (32 bpp)
* factor (factor for reducing threshold due to variance; >= 0)
* &pixsd (<optional return> local standard deviation)
* Return: pixd (8 bpp, sauvola threshold values), or null on error
*
* Notes:
* (1) The Sauvola threshold is determined from the formula:
* t = m * (1 - k * (1 - s / 128))
* where:
* t = local threshold
* m = local mean
* k = @factor (>= 0) [ typ. 0.35 ]
* s = local standard deviation, which is maximized at
* 127.5 when half the samples are 0 and half are 255.
* (2) See pixSauvolaBinarize() for other details.
* (3) Important definitions and relations for computing averages:
* v == pixel value
* E(p) == expected value of p == average of p over some pixel set
* S(v) == square of v == v * v
* mv == E(v) == expected pixel value == mean value
* ms == E(S(v)) == expected square of pixel values
* == mean square value
* var == variance == expected square of deviation from mean
* == E(S(v - mv)) = E(S(v) - 2 * S(v * mv) + S(mv))
* = E(S(v)) - S(mv)
* = ms - mv * mv
* s == standard deviation = sqrt(var)
* So for evaluating the standard deviation in the Sauvola
* threshold, we take
* s = sqrt(ms - mv * mv)
*/
PIX *
pixSauvolaGetThreshold(PIX *pixm,
PIX *pixms,
l_float32 factor,
PIX **ppixsd)
{
l_int32 i, j, w, h, tabsize, wplm, wplms, wplsd, wpld, usetab;
l_int32 mv, ms, var, thresh;
l_uint32 *datam, *datams, *datasd, *datad;
l_uint32 *linem, *linems, *linesd, *lined;
l_float32 sd;
l_float32 *tab; /* of 2^16 square roots */
PIX *pixsd, *pixd;
PROCNAME("pixSauvolaGetThreshold");
if (ppixsd) *ppixsd = NULL;
if (!pixm || pixGetDepth(pixm) != 8)
return (PIX *)ERROR_PTR("pixm undefined or not 8 bpp", procName, NULL);
if (pixGetColormap(pixm))
return (PIX *)ERROR_PTR("pixm is colormapped", procName, NULL);
if (!pixms || pixGetDepth(pixms) != 32)
return (PIX *)ERROR_PTR("pixms undefined or not 32 bpp",
procName, NULL);
if (factor < 0.0)
return (PIX *)ERROR_PTR("factor must be >= 0", procName, NULL);
/* Only make a table of 2^16 square roots if there
* are enough pixels to justify it. */
pixGetDimensions(pixm, &w, &h, NULL);
usetab = (w * h > 100000) ? 1 : 0;
if (usetab) {
tabsize = 1 << 16;
tab = (l_float32 *)CALLOC(tabsize, sizeof(l_float32));
for (i = 0; i < tabsize; i++)
tab[i] = (l_float32)sqrt((l_float64)i);
}
pixd = pixCreate(w, h, 8);
if (ppixsd) {
pixsd = pixCreate(w, h, 8);
*ppixsd = pixsd;
}
datam = pixGetData(pixm);
datams = pixGetData(pixms);
if (ppixsd) datasd = pixGetData(pixsd);
datad = pixGetData(pixd);
wplm = pixGetWpl(pixm);
wplms = pixGetWpl(pixms);
if (ppixsd) wplsd = pixGetWpl(pixsd);
wpld = pixGetWpl(pixd);
for (i = 0; i < h; i++) {
linem = datam + i * wplm;
linems = datams + i * wplms;
if (ppixsd) linesd = datasd + i * wplsd;
lined = datad + i * wpld;
for (j = 0; j < w; j++) {
mv = GET_DATA_BYTE(linem, j);
ms = linems[j];
var = ms - mv * mv;
if (usetab)
sd = tab[var];
else
sd = (l_float32)sqrt((l_float32)var);
if (ppixsd) SET_DATA_BYTE(linesd, j, (l_int32)sd);
thresh = (l_int32)(mv * (1.0 - factor * (1.0 - sd / 128.)));
SET_DATA_BYTE(lined, j, thresh);
}
}
if (usetab) FREE(tab);
return pixd;
}
/*!
* pixApplyLocalThreshold()
*
* Input: pixs (8 bpp grayscale; not colormapped)
* pixth (8 bpp array of local thresholds)
* redfactor ( ... )
* Return: pixd (1 bpp, thresholded image), or null on error
*/
PIX *
pixApplyLocalThreshold(PIX *pixs,
PIX *pixth,
l_int32 redfactor)
{
l_int32 i, j, w, h, wpls, wplt, wpld, vals, valt;
l_uint32 *datas, *datat, *datad, *lines, *linet, *lined;
PIX *pixd;
PROCNAME("pixApplyLocalThreshold");
if (!pixs || pixGetDepth(pixs) != 8)
return (PIX *)ERROR_PTR("pixs undefined or not 8 bpp", procName, NULL);
if (pixGetColormap(pixs))
return (PIX *)ERROR_PTR("pixs is colormapped", procName, NULL);
if (!pixth || pixGetDepth(pixth) != 8)
return (PIX *)ERROR_PTR("pixth undefined or not 8 bpp", procName, NULL);
pixGetDimensions(pixs, &w, &h, NULL);
pixd = pixCreate(w, h, 1);
datas = pixGetData(pixs);
datat = pixGetData(pixth);
datad = pixGetData(pixd);
wpls = pixGetWpl(pixs);
wplt = pixGetWpl(pixth);
wpld = pixGetWpl(pixd);
for (i = 0; i < h; i++) {
lines = datas + i * wpls;
linet = datat + i * wplt;
lined = datad + i * wpld;
for (j = 0; j < w; j++) {
vals = GET_DATA_BYTE(lines, j);
valt = GET_DATA_BYTE(linet, j);
if (vals < valt)
SET_DATA_BIT(lined, j);
}
}
return pixd;
}
/*----------------------------------------------------------------------*
* Thresholding using connected components *
*----------------------------------------------------------------------*/
/*!
* pixThresholdByConnComp()
*
* Input: pixs (depth > 1, colormap OK)
* pixm (<optional> 1 bpp mask giving region to ignore by setting
* pixels to white; use NULL if no mask)
* start, end, incr (binarization threshold levels to test)
* thresh48 (threshold on normalized difference between the
* numbers of 4 and 8 connected components)
* threshdiff (threshold on normalized difference between the
* number of 4 cc at successive iterations)
* &globthresh (<optional return> best global threshold; 0
* if no threshold is found)
* &pixd (<optional return> image thresholded to binary, or
* null if no threshold is found)
* debugflag (1 for plotted results)
* Return: 0 if OK, 1 on error or if no threshold is found
*
* Notes:
* (1) This finds a global threshold based on connected components.
* Although slow, it is reasonable to use it in a situation where
* (a) the background in the image is relatively uniform, and
* (b) the result will be fed to an OCR program that accepts 1 bpp
* images and works best with easily segmented characters.
* The reason for (b) is that this selects a threshold with a
* minimum number of both broken characters and merged characters.
* (2) If the pix has color, it is converted to gray using the
* max component.
* (3) Input 0 to use default values for any of these inputs:
* @start, @end, @incr, @thresh48, @threshdiff.
* (4) This approach can be understood as follows. When the
* binarization threshold is varied, the numbers of c.c. identify
* four regimes:
* (a) For low thresholds, text is broken into small pieces, and
* the number of c.c. is large, with the 4 c.c. significantly
* exceeding the 8 c.c.
* (b) As the threshold rises toward the optimum value, the text
* characters coalesce and there is very little difference
* between the numbers of 4 and 8 c.c, which both go
* through a minimum.
* (c) Above this, the image background gets noisy because some
* pixels are(thresholded to foreground, and the numbers
* of c.c. quickly increase, with the 4 c.c. significantly
* larger than the 8 c.c.
* (d) At even higher thresholds, the image background noise
* coalesces as it becomes mostly foreground, and the
* number of c.c. drops quickly.
* (5) If there is no global threshold that distinguishes foreground
* text from background (e.g., weak text over a background that
* has significant variation and/or bleedthrough), this returns 1,
* which the caller should check.
*/
l_int32
pixThresholdByConnComp(PIX *pixs,
PIX *pixm,
l_int32 start,
l_int32 end,
l_int32 incr,
l_float32 thresh48,
l_float32 threshdiff,
l_int32 *pglobthresh,
PIX **ppixd,
l_int32 debugflag)
{
l_int32 i, thresh, n, n4, n8, mincounts, found, globthresh;
l_float32 count4, count8, firstcount4, prevcount4, diff48, diff4;
GPLOT *gplot;
NUMA *na4, *na8;
PIX *pix1, *pix2, *pix3;
PROCNAME("pixThresholdByConnComp");
if (pglobthresh) *pglobthresh = 0;
if (ppixd) *ppixd = NULL;
if (!pixs || pixGetDepth(pixs) == 1)
return ERROR_INT("pixs undefined or 1 bpp", procName, 1);
if (pixm && pixGetDepth(pixm) != 1)
return ERROR_INT("pixm must be 1 bpp", procName, 1);
/* Assign default values if requested */
if (start <= 0) start = 80;
if (end <= 0) end = 200;
if (incr <= 0) incr = 10;
if (thresh48 <= 0.0) thresh48 = 0.01;
if (threshdiff <= 0.0) threshdiff = 0.01;
if (start > end)
return ERROR_INT("invalid start,end", procName, 1);
/* Make 8 bpp, using the max component if color. */
if (pixGetColormap(pixs))
pix1 = pixRemoveColormap(pixs, REMOVE_CMAP_BASED_ON_SRC);
else
pix1 = pixClone(pixs);
if (pixGetDepth(pix1) == 32)
pix2 = pixConvertRGBToGrayMinMax(pix1, L_CHOOSE_MAX);
else
pix2 = pixConvertTo8(pix1, 0);
pixDestroy(&pix1);
/* Mask out any non-text regions. Do this in-place, because pix2
* can never be the same pix as pixs. */
if (pixm)
pixSetMasked(pix2, pixm, 255);
/* Make sure there are enough components to get a valid signal */
pix3 = pixConvertTo1(pix2, start);
pixCountConnComp(pix3, 4, &n4);
pixDestroy(&pix3);
mincounts = 500;
if (n4 < mincounts) {
L_INFO("Insufficient component count: %d\n", procName, n4);
pixDestroy(&pix2);
return 1;
}
/* Compute the c.c. data */
na4 = numaCreate(0);
na8 = numaCreate(0);
numaSetParameters(na4, start, incr);
numaSetParameters(na8, start, incr);
for (thresh = start, i = 0; thresh <= end; thresh += incr, i++) {
pix3 = pixConvertTo1(pix2, thresh);
pixCountConnComp(pix3, 4, &n4);
pixCountConnComp(pix3, 8, &n8);
numaAddNumber(na4, n4);
numaAddNumber(na8, n8);
pixDestroy(&pix3);
}
if (debugflag) {
gplot = gplotCreate("/tmp/threshroot", GPLOT_PNG,
"number of cc vs. threshold",
"threshold", "number of cc");
gplotAddPlot(gplot, NULL, na4, GPLOT_LINES, "plot 4cc");
gplotAddPlot(gplot, NULL, na8, GPLOT_LINES, "plot 8cc");
gplotMakeOutput(gplot);
gplotDestroy(&gplot);
}
n = numaGetCount(na4);
found = FALSE;
for (i = 0; i < n; i++) {
if (i == 0) {
numaGetFValue(na4, i, &firstcount4);
prevcount4 = firstcount4;
} else {
numaGetFValue(na4, i, &count4);
numaGetFValue(na8, i, &count8);
diff48 = (count4 - count8) / firstcount4;
diff4 = L_ABS(prevcount4 - count4) / firstcount4;
if (debugflag) {
fprintf(stderr, "diff48 = %7.3f, diff4 = %7.3f\n",
diff48, diff4);
}
if (diff48 < thresh48 && diff4 < threshdiff) {
found = TRUE;
break;
}
prevcount4 = count4;
}
}
numaDestroy(&na4);
numaDestroy(&na8);
if (found) {
globthresh = start + i * incr;
if (pglobthresh) *pglobthresh = globthresh;
if (ppixd) {
*ppixd = pixConvertTo1(pix2, globthresh);
pixCopyResolution(*ppixd, pixs);
}
if (debugflag) fprintf(stderr, "global threshold = %d\n", globthresh);
pixDestroy(&pix2);
return 0;
}
if (debugflag) fprintf(stderr, "no global threshold found\n");
pixDestroy(&pix2);
return 1;
}