/*====================================================================* - Copyright (C) 2001 Leptonica. All rights reserved. - This software is distributed in the hope that it will be - useful, but with NO WARRANTY OF ANY KIND. - No author or distributor accepts responsibility to anyone for the - consequences of using this software, or for whether it serves any - particular purpose or works at all, unless he or she says so in - writing. Everyone is granted permission to copy, modify and - redistribute this source code, for commercial or non-commercial - purposes, with the following restrictions: (1) the origin of this - source code must not be misrepresented; (2) modified versions must - be plainly marked as such; and (3) this notice may not be removed - or altered from any source or modified source distribution. *====================================================================*/ /* * adaptmap.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. * =================================================================== * * Adaptive background normalization (top-level functions) * PIX *pixBackgroundNormSimple() 8 and 32 bpp * PIX *pixBackgroundNorm() 8 and 32 bpp * PIX *pixBackgroundNormMorph() 8 and 32 bpp * * Arrays of inverted background values for normalization (16 bpp) * l_int32 pixBackgroundNormGrayArray() 8 bpp input * l_int32 pixBackgroundNormRGBArrays() 32 bpp input * l_int32 pixBackgroundNormGrayArrayMorph() 8 bpp input * l_int32 pixBackgroundNormRGBArraysMorph() 32 bpp input * * Measurement of local background * l_int32 pixGetBackgroundGrayMap() 8 bpp * l_int32 pixGetBackgroundRGBMap() 32 bpp * l_int32 pixGetBackgroundGrayMapMorph() 8 bpp * l_int32 pixGetBackgroundRGBMapMorph() 32 bpp * l_int32 pixFillMapHoles() * PIX *pixExtendByReplication() 8 bpp * l_int32 pixSmoothConnectedRegions() 8 bpp * * Measurement of local foreground * l_int32 pixGetForegroundGrayMap() 8 bpp * * Generate inverted background map for each component * PIX *pixGetInvBackgroundMap() 16 bpp * * Apply inverse background map to image * PIX *pixApplyInvBackgroundGrayMap() 8 bpp * PIX *pixApplyInvBackgroundRGBMap() 32 bpp * * Apply variable map * PIX *pixApplyVariableGrayMap() 8 bpp * * Non-adaptive (global) mapping * PIX *pixGlobalNormRGB() 32 bpp or cmapped * PIX *pixGlobalNormNoSatRGB() 32 bpp * * Adaptive threshold spread normalization * l_int32 pixThresholdSpreadNorm() 8 bpp * * Adaptive background normalization (flexible adaptaption) * PIX *pixBackgroundNormFlex() 8 bpp * * Adaptive contrast normalization * PIX *pixContrastNorm() 8 bpp * l_int32 pixMinMaxTiles() * l_int32 pixSetLowContrast() * PIX *pixLinearTRCTiled() * static l_int32 *iaaGetLinearTRC() * * Background normalization is done by generating a reduced map (or set * of maps) representing the estimated background value of the * input image, and using this to shift the pixel values so that * this background value is set to some constant value. * * Specifically, normalization has 3 steps: * (1) Generate a background map at a reduced scale. * (2) Make the array of inverted background values by inverting * the map. The result is an array of local multiplicative factors. * (3) Apply this inverse background map to the image * * The inverse background arrays can be generated in two different ways here: * (1) Remove the 'foreground' pixels and average over the remaining * pixels in each tile. Propagate values into tiles where * values have not been assigned, either because there was not * enough background in the tile or because the tile is covered * by a foreground region described by an image mask. * After the background map is made, the inverse map is generated by * smoothing over some number of adjacent tiles * (block convolution) and then inverting. * (2) Remove the foreground pixels using a morphological closing * on a subsampled version of the image. Propagate values * into pixels covered by an optional image mask. Invert the * background map without preconditioning by convolutional smoothing. * * Note: Several of these functions make an implicit assumption about RGB * component ordering. * * Other methods for adaptively normalizing the image are also given here. * * (1) pixThresholdSpreadNorm() computes a local threshold over the image * and normalizes the input pixel values so that this computed threshold * is a constant across the entire image. * * (2) pixContrastNorm() computes and applies a local TRC so that the * local dynamic range is expanded to the full 8 bits, where the * darkest pixels are mapped to 0 and the lightest to 255. This is * useful for improving the appearance of pages with very light * foreground or very dark background, and where the local TRC * function doesn't change rapidly with position. */ #include #include #include "allheaders.h" /* Default input parameters for pixBackgroundNormSimple() * Note: * (1) mincount must never exceed the tile area (width * height) * (2) bgval must be sufficiently below 255 to avoid accidental * saturation; otherwise it should be large to avoid * shrinking the dynamic range * (3) results should otherwise not be sensitive to these values */ static const l_int32 DEFAULT_TILE_WIDTH = 10; static const l_int32 DEFAULT_TILE_HEIGHT = 15; static const l_int32 DEFAULT_FG_THRESHOLD = 60; static const l_int32 DEFAULT_MIN_COUNT = 40; static const l_int32 DEFAULT_BG_VAL = 200; static const l_int32 DEFAULT_X_SMOOTH_SIZE = 2; static const l_int32 DEFAULT_Y_SMOOTH_SIZE = 1; static l_int32 *iaaGetLinearTRC(l_int32 **iaa, l_int32 diff); #ifndef NO_CONSOLE_IO #define DEBUG_GLOBAL 0 #endif /* ~NO_CONSOLE_IO */ /*------------------------------------------------------------------* * Adaptive background normalization * *------------------------------------------------------------------*/ /*! * pixBackgroundNormSimple() * * Input: pixs (8 bpp grayscale or 32 bpp rgb) * pixim ( 1 bpp 'image' mask; can be null) * pixg ( 8 bpp grayscale version; can be null) * Return: pixd (8 bpp or 32 bpp rgb), or null on error * * Notes: * (1) This is a simplified interface to pixBackgroundNorm(), * where seven parameters are defaulted. * (2) The input image is either grayscale or rgb. * (3) See pixBackgroundNorm() for usage and function. */ PIX * pixBackgroundNormSimple(PIX *pixs, PIX *pixim, PIX *pixg) { return pixBackgroundNorm(pixs, pixim, pixg, DEFAULT_TILE_WIDTH, DEFAULT_TILE_HEIGHT, DEFAULT_FG_THRESHOLD, DEFAULT_MIN_COUNT, DEFAULT_BG_VAL, DEFAULT_X_SMOOTH_SIZE, DEFAULT_Y_SMOOTH_SIZE); } /*! * pixBackgroundNorm() * * Input: pixs (8 bpp grayscale or 32 bpp rgb) * pixim ( 1 bpp 'image' mask; can be null) * pixg ( 8 bpp grayscale version; 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) * Return: pixd (8 bpp or 32 bpp rgb), or null on error * * Notes: * (1) This is a top-level interface for normalizing the image intensity * by mapping the image so that the background is near the input * value 'bgval'. * (2) The input image is either grayscale or rgb. * (3) For each component in the input image, the background value * in each tile is estimated using the values in the tile that * are not part of the foreground, where the foreground is * determined by the input 'thresh' argument. * (4) An optional binary mask can be specified, with the foreground * pixels typically over image regions. The resulting background * map values will be determined by surrounding pixels that are * not under the mask foreground. The origin (0,0) of this mask * is assumed to be aligned with the origin of the input image. * This binary mask must not fully cover pixs, because then there * will be no pixels in the input image available to compute * the background. * (5) An optional grayscale version of the input pixs can be supplied. * The only reason to do this is if the input is RGB and this * grayscale version can be used elsewhere. If the input is RGB * and this is not supplied, it is made internally using only * the green component, and destroyed after use. * (6) The dimensions of the pixel tile (sx, sy) give the amount by * by which the map is reduced in size from the input image. * (7) The threshold is used to binarize the input image, in order to * locate the foreground components. If this is set too low, * some actual foreground may be used to determine the maps; * if set too high, there may not be enough background * to determine the map values accurately. Typically, it's * better to err by setting the threshold too high. * (8) A 'mincount' threshold is a minimum count of pixels in a * tile for which a background reading is made, in order for that * pixel in the map to be valid. This number should perhaps be * at least 1/3 the size of the tile. * (9) A 'bgval' target background value for the normalized image. This * should be at least 128. If set too close to 255, some * clipping will occur in the result. * (10) Two factors, 'smoothx' and 'smoothy', are input for smoothing * the map. Each low-pass filter kernel dimension is * is 2 * (smoothing factor) + 1, so a * value of 0 means no smoothing. A value of 1 or 2 is recommended. */ PIX * pixBackgroundNorm(PIX *pixs, PIX *pixim, PIX *pixg, l_int32 sx, l_int32 sy, l_int32 thresh, l_int32 mincount, l_int32 bgval, l_int32 smoothx, l_int32 smoothy) { l_int32 d, allfg; PIX *pixm, *pixmi, *pixd; PIX *pixmr, *pixmg, *pixmb, *pixmri, *pixmgi, *pixmbi; PROCNAME("pixBackgroundNorm"); if (!pixs) return (PIX *)ERROR_PTR("pixs not defined", procName, NULL); d = pixGetDepth(pixs); if (d != 8 && d != 32) return (PIX *)ERROR_PTR("pixs not 8 or 32 bpp", 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", procName); mincount = (sx * sy) / 3; } /* If pixim exists, verify that it is not all foreground. */ if (pixim) { pixInvert(pixim, pixim); pixZero(pixim, &allfg); pixInvert(pixim, pixim); if (allfg) return (PIX *)ERROR_PTR("pixim all foreground", procName, NULL); } pixd = NULL; if (d == 8) { pixm = NULL; pixGetBackgroundGrayMap(pixs, pixim, sx, sy, thresh, mincount, &pixm); if (!pixm) { L_WARNING("map not made; returning a copy of the source", procName); return pixCopy(NULL, pixs); } pixmi = pixGetInvBackgroundMap(pixm, bgval, smoothx, smoothy); if (!pixmi) ERROR_PTR("pixmi not made", procName, NULL); else pixd = pixApplyInvBackgroundGrayMap(pixs, pixmi, sx, sy); pixDestroy(&pixm); pixDestroy(&pixmi); } else { pixmr = pixmg = pixmb = NULL; pixGetBackgroundRGBMap(pixs, pixim, pixg, sx, sy, thresh, mincount, &pixmr, &pixmg, &pixmb); if (!pixmr || !pixmg || !pixmb) { pixDestroy(&pixmr); pixDestroy(&pixmg); pixDestroy(&pixmb); L_WARNING("map not made; returning a copy of the source", procName); return pixCopy(NULL, pixs); } pixmri = pixGetInvBackgroundMap(pixmr, bgval, smoothx, smoothy); pixmgi = pixGetInvBackgroundMap(pixmg, bgval, smoothx, smoothy); pixmbi = pixGetInvBackgroundMap(pixmb, bgval, smoothx, smoothy); if (!pixmri || !pixmgi || !pixmbi) ERROR_PTR("not all pixm*i are made", procName, NULL); else pixd = pixApplyInvBackgroundRGBMap(pixs, pixmri, pixmgi, pixmbi, sx, sy); pixDestroy(&pixmr); pixDestroy(&pixmg); pixDestroy(&pixmb); pixDestroy(&pixmri); pixDestroy(&pixmgi); pixDestroy(&pixmbi); } if (!pixd) ERROR_PTR("pixd not made", procName, NULL); return pixd; } /*! * pixBackgroundNormMorph() * * Input: pixs (8 bpp grayscale or 32 bpp rgb) * pixim ( 1 bpp 'image' mask; can be null) * reduction (at which morph closings are done; between 2 and 16) * size (of square Sel for the closing; use an odd number) * bgval (target bg val; typ. > 128) * Return: pixd (8 bpp), or null on error * * Notes: * (1) This is a top-level interface for normalizing the image intensity * by mapping the image so that the background is near the input * value 'bgval'. * (2) The input image is either grayscale or rgb. * (3) For each component in the input image, the background value * is estimated using a grayscale closing; hence the 'Morph' * in the function name. * (4) An optional binary mask can be specified, with the foreground * pixels typically over image regions. The resulting background * map values will be determined by surrounding pixels that are * not under the mask foreground. The origin (0,0) of this mask * is assumed to be aligned with the origin of the input image. * This binary mask must not fully cover pixs, because then there * will be no pixels in the input image available to compute * the background. * (5) The map is computed at reduced size (given by 'reduction') * from the input pixs and optional pixim. At this scale, * pixs is closed to remove the background, using a square Sel * of odd dimension. The product of reduction * size should be * large enough to remove most of the text foreground. * (6) No convolutional smoothing needs to be done on the map before * inverting it. * (7) A 'bgval' target background value for the normalized image. This * should be at least 128. If set too close to 255, some * clipping will occur in the result. */ PIX * pixBackgroundNormMorph(PIX *pixs, PIX *pixim, l_int32 reduction, l_int32 size, l_int32 bgval) { l_int32 d, allfg; PIX *pixm, *pixmi, *pixd; PIX *pixmr, *pixmg, *pixmb, *pixmri, *pixmgi, *pixmbi; PROCNAME("pixBackgroundNormMorph"); if (!pixs) return (PIX *)ERROR_PTR("pixs not defined", procName, NULL); d = pixGetDepth(pixs); if (d != 8 && d != 32) return (PIX *)ERROR_PTR("pixs not 8 or 32 bpp", procName, NULL); if (reduction < 2 || reduction > 16) return (PIX *)ERROR_PTR("reduction must be between 2 and 16", procName, NULL); /* If pixim exists, verify that it is not all foreground. */ if (pixim) { pixInvert(pixim, pixim); pixZero(pixim, &allfg); pixInvert(pixim, pixim); if (allfg) return (PIX *)ERROR_PTR("pixim all foreground", procName, NULL); } pixd = NULL; if (d == 8) { pixGetBackgroundGrayMapMorph(pixs, pixim, reduction, size, &pixm); if (!pixm) return (PIX *)ERROR_PTR("pixm not made", procName, NULL); pixmi = pixGetInvBackgroundMap(pixm, bgval, 0, 0); if (!pixmi) ERROR_PTR("pixmi not made", procName, NULL); else pixd = pixApplyInvBackgroundGrayMap(pixs, pixmi, reduction, reduction); pixDestroy(&pixm); pixDestroy(&pixmi); } else { /* d == 32 */ pixmr = pixmg = pixmb = NULL; pixGetBackgroundRGBMapMorph(pixs, pixim, reduction, size, &pixmr, &pixmg, &pixmb); if (!pixmr || !pixmg || !pixmb) { pixDestroy(&pixmr); pixDestroy(&pixmg); pixDestroy(&pixmb); return (PIX *)ERROR_PTR("not all pixm*", procName, NULL); } pixmri = pixGetInvBackgroundMap(pixmr, bgval, 0, 0); pixmgi = pixGetInvBackgroundMap(pixmg, bgval, 0, 0); pixmbi = pixGetInvBackgroundMap(pixmb, bgval, 0, 0); if (!pixmri || !pixmgi || !pixmbi) ERROR_PTR("not all pixm*i are made", procName, NULL); else pixd = pixApplyInvBackgroundRGBMap(pixs, pixmri, pixmgi, pixmbi, reduction, reduction); pixDestroy(&pixmr); pixDestroy(&pixmg); pixDestroy(&pixmb); pixDestroy(&pixmri); pixDestroy(&pixmgi); pixDestroy(&pixmbi); } if (!pixd) ERROR_PTR("pixd not made", procName, NULL); return pixd; } /*-------------------------------------------------------------------------* * Arrays of inverted background values for normalization * *-------------------------------------------------------------------------* * Notes for these four functions: * * (1) They are useful if you need to save the actual mapping array. * * (2) They could be used in the top-level functions but are * * not because their use makes those functions less clear. * * (3) Each component in the input pixs generates a 16 bpp pix array. * *-------------------------------------------------------------------------*/ /*! * pixBackgroundNormGrayArray() * * Input: pixs (8 bpp grayscale) * pixim ( 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) * &pixd ( 16 bpp array of inverted background value) * Return: 0 if OK, 1 on error * * Notes: * (1) See notes in pixBackgroundNorm(). * (2) This returns a 16 bpp pix that can be used by * pixApplyInvBackgroundGrayMap() to generate a normalized version * of the input pixs. */ l_int32 pixBackgroundNormGrayArray(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, PIX **ppixd) { l_int32 allfg; PIX *pixm; PROCNAME("pixBackgroundNormGrayArray"); if (!ppixd) return ERROR_INT("&pixd not defined", procName, 1); *ppixd = NULL; if (!pixs) return ERROR_INT("pixs not defined", procName, 1); if (pixGetDepth(pixs) != 8) return ERROR_INT("pixs not 8 bpp", procName, 1); if (pixim && pixGetDepth(pixim) != 1) return ERROR_INT("pixim not 1 bpp", procName, 1); if (sx < 4 || sy < 4) return ERROR_INT("sx and sy must be >= 4", procName, 1); if (mincount > sx * sy) { L_WARNING("mincount too large for tile size", procName); mincount = (sx * sy) / 3; } /* If pixim exists, verify that it is not all foreground. */ if (pixim) { pixInvert(pixim, pixim); pixZero(pixim, &allfg); pixInvert(pixim, pixim); if (allfg) return ERROR_INT("pixim all foreground", procName, 1); } pixGetBackgroundGrayMap(pixs, pixim, sx, sy, thresh, mincount, &pixm); if (!pixm) return ERROR_INT("pixm not made", procName, 1); *ppixd = pixGetInvBackgroundMap(pixm, bgval, smoothx, smoothy); pixDestroy(&pixm); return 0; } /*! * pixBackgroundNormRGBArrays() * * Input: pixs (32 bpp rgb) * pixim ( 1 bpp 'image' mask; can be null) * pixg ( 8 bpp grayscale version; 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) * &pixr ( 16 bpp array of inverted R background value) * &pixg ( 16 bpp array of inverted G background value) * &pixb ( 16 bpp array of inverted B background value) * Return: 0 if OK, 1 on error * * Notes: * (1) See notes in pixBackgroundNorm(). * (2) This returns a set of three 16 bpp pix that can be used by * pixApplyInvBackgroundGrayMap() to generate a normalized version * of each component of the input pixs. */ l_int32 pixBackgroundNormRGBArrays(PIX *pixs, PIX *pixim, PIX *pixg, l_int32 sx, l_int32 sy, l_int32 thresh, l_int32 mincount, l_int32 bgval, l_int32 smoothx, l_int32 smoothy, PIX **ppixr, PIX **ppixg, PIX **ppixb) { l_int32 allfg; PIX *pixmr, *pixmg, *pixmb; PROCNAME("pixBackgroundNormRGBArrays"); if (!ppixr || !ppixg || !ppixb) return ERROR_INT("&pixr, &pixg, &pixb not all defined", procName, 1); *ppixr = *ppixg = *ppixb = NULL; if (!pixs) return ERROR_INT("pixs not defined", procName, 1); if (pixGetDepth(pixs) != 32) return ERROR_INT("pixs not 32 bpp", procName, 1); if (pixim && pixGetDepth(pixim) != 1) return ERROR_INT("pixim not 1 bpp", procName, 1); if (sx < 4 || sy < 4) return ERROR_INT("sx and sy must be >= 4", procName, 1); if (mincount > sx * sy) { L_WARNING("mincount too large for tile size", procName); mincount = (sx * sy) / 3; } /* If pixim exists, verify that it is not all foreground. */ if (pixim) { pixInvert(pixim, pixim); pixZero(pixim, &allfg); pixInvert(pixim, pixim); if (allfg) return ERROR_INT("pixim all foreground", procName, 1); } pixGetBackgroundRGBMap(pixs, pixim, pixg, sx, sy, thresh, mincount, &pixmr, &pixmg, &pixmb); if (!pixmr || !pixmg || !pixmb) { pixDestroy(&pixmr); pixDestroy(&pixmg); pixDestroy(&pixmb); return ERROR_INT("not all pixm* made", procName, 1); } *ppixr = pixGetInvBackgroundMap(pixmr, bgval, smoothx, smoothy); *ppixg = pixGetInvBackgroundMap(pixmg, bgval, smoothx, smoothy); *ppixb = pixGetInvBackgroundMap(pixmb, bgval, smoothx, smoothy); pixDestroy(&pixmr); pixDestroy(&pixmg); pixDestroy(&pixmb); return 0; } /*! * pixBackgroundNormGrayArrayMorph() * * Input: pixs (8 bpp grayscale) * pixim ( 1 bpp 'image' mask; can be null) * reduction (at which morph closings are done; between 2 and 16) * size (of square Sel for the closing; use an odd number) * bgval (target bg val; typ. > 128) * &pixd ( 16 bpp array of inverted background value) * Return: 0 if OK, 1 on error * * Notes: * (1) See notes in pixBackgroundNormMorph(). * (2) This returns a 16 bpp pix that can be used by * pixApplyInvBackgroundGrayMap() to generate a normalized version * of the input pixs. */ l_int32 pixBackgroundNormGrayArrayMorph(PIX *pixs, PIX *pixim, l_int32 reduction, l_int32 size, l_int32 bgval, PIX **ppixd) { l_int32 allfg; PIX *pixm; PROCNAME("pixBackgroundNormGrayArrayMorph"); if (!ppixd) return ERROR_INT("&pixd not defined", procName, 1); *ppixd = NULL; if (!pixs) return ERROR_INT("pixs not defined", procName, 1); if (pixGetDepth(pixs) != 8) return ERROR_INT("pixs not 8 bpp", procName, 1); if (pixim && pixGetDepth(pixim) != 1) return ERROR_INT("pixim not 1 bpp", procName, 1); if (reduction < 2 || reduction > 16) return ERROR_INT("reduction must be between 2 and 16", procName, 1); /* If pixim exists, verify that it is not all foreground. */ if (pixim) { pixInvert(pixim, pixim); pixZero(pixim, &allfg); pixInvert(pixim, pixim); if (allfg) return ERROR_INT("pixim all foreground", procName, 1); } pixGetBackgroundGrayMapMorph(pixs, pixim, reduction, size, &pixm); if (!pixm) return ERROR_INT("pixm not made", procName, 1); *ppixd = pixGetInvBackgroundMap(pixm, bgval, 0, 0); pixDestroy(&pixm); return 0; } /*! * pixBackgroundNormRGBArraysMorph() * * Input: pixs (32 bpp rgb) * pixim ( 1 bpp 'image' mask; can be null) * reduction (at which morph closings are done; between 2 and 16) * size (of square Sel for the closing; use an odd number) * bgval (target bg val; typ. > 128) * &pixr ( 16 bpp array of inverted R background value) * &pixg ( 16 bpp array of inverted G background value) * &pixb ( 16 bpp array of inverted B background value) * Return: 0 if OK, 1 on error * * Notes: * (1) See notes in pixBackgroundNormMorph(). * (2) This returns a set of three 16 bpp pix that can be used by * pixApplyInvBackgroundGrayMap() to generate a normalized version * of each component of the input pixs. */ l_int32 pixBackgroundNormRGBArraysMorph(PIX *pixs, PIX *pixim, l_int32 reduction, l_int32 size, l_int32 bgval, PIX **ppixr, PIX **ppixg, PIX **ppixb) { l_int32 allfg; PIX *pixmr, *pixmg, *pixmb; PROCNAME("pixBackgroundNormRGBArraysMorph"); if (!ppixr || !ppixg || !ppixb) return ERROR_INT("&pixr, &pixg, &pixb not all defined", procName, 1); *ppixr = *ppixg = *ppixb = NULL; if (!pixs) return ERROR_INT("pixs not defined", procName, 1); if (pixGetDepth(pixs) != 32) return ERROR_INT("pixs not 32 bpp", procName, 1); if (pixim && pixGetDepth(pixim) != 1) return ERROR_INT("pixim not 1 bpp", procName, 1); if (reduction < 2 || reduction > 16) return ERROR_INT("reduction must be between 2 and 16", procName, 1); /* If pixim exists, verify that it is not all foreground. */ if (pixim) { pixInvert(pixim, pixim); pixZero(pixim, &allfg); pixInvert(pixim, pixim); if (allfg) return ERROR_INT("pixim all foreground", procName, 1); } pixGetBackgroundRGBMapMorph(pixs, pixim, reduction, size, &pixmr, &pixmg, &pixmb); if (!pixmr || !pixmg || !pixmb) { pixDestroy(&pixmr); pixDestroy(&pixmg); pixDestroy(&pixmb); return ERROR_INT("not all pixm* made", procName, 1); } *ppixr = pixGetInvBackgroundMap(pixmr, bgval, 0, 0); *ppixg = pixGetInvBackgroundMap(pixmg, bgval, 0, 0); *ppixb = pixGetInvBackgroundMap(pixmb, bgval, 0, 0); pixDestroy(&pixmr); pixDestroy(&pixmg); pixDestroy(&pixmb); return 0; } /*------------------------------------------------------------------* * Measurement of local background * *------------------------------------------------------------------*/ /*! * pixGetBackgroundGrayMap() * * Input: pixs (8 bpp) * pixim ( 1 bpp 'image' mask; can be null; it * should not have all foreground pixels) * sx, sy (tile size in pixels) * thresh (threshold for determining foreground) * mincount (min threshold on counts in a tile) * &pixd ( 8 bpp grayscale map) * Return: 0 if OK, 1 on error * * Notes: * (1) The background is measured in regions that don't have * images. It is then propagated into the image regions, * and finally smoothed in each image region. */ l_int32 pixGetBackgroundGrayMap(PIX *pixs, PIX *pixim, l_int32 sx, l_int32 sy, l_int32 thresh, l_int32 mincount, PIX **ppixd) { l_int32 w, h, wd, hd, wim, him, wpls, wplim, wpld, wplf; l_int32 xim, yim, delx, nx, ny, i, j, k, m; l_int32 count, sum, val8; l_int32 empty, fgpixels; l_uint32 *datas, *dataim, *datad, *dataf, *lines, *lineim, *lined, *linef; l_float32 scalex, scaley; PIX *pixd, *piximi, *pixb, *pixf, *pixims; PROCNAME("pixGetBackgroundGrayMap"); if (!ppixd) return ERROR_INT("&pixd not defined", procName, 1); *ppixd = NULL; if (!pixs) return ERROR_INT("pixs not defined", procName, 1); if (pixGetDepth(pixs) != 8) return ERROR_INT("pixs not 8 bpp", procName, 1); if (pixim && pixGetDepth(pixim) != 1) return ERROR_INT("pixim not 1 bpp", procName, 1); if (sx < 4 || sy < 4) return ERROR_INT("sx and sy must be >= 4", procName, 1); if (mincount > sx * sy) { L_WARNING("mincount too large for tile size", procName); mincount = (sx * sy) / 3; } /* Evaluate the 'image' mask, pixim, and make sure * it is not all fg. */ fgpixels = 0; /* boolean for existence of fg pixels in the image mask. */ if (pixim) { piximi = pixInvert(NULL, pixim); /* set non-'image' pixels to 1 */ pixZero(piximi, &empty); pixDestroy(&piximi); if (empty) return ERROR_INT("pixim all fg; no background", procName, 1); pixZero(pixim, &empty); if (!empty) /* there are fg pixels in pixim */ fgpixels = 1; } /* Generate the foreground mask, pixf, which is at * full resolution. These pixels will be ignored when * computing the background values. */ pixb = pixThresholdToBinary(pixs, thresh); pixf = pixMorphSequence(pixb, "d7.1 + d1.7", 0); pixDestroy(&pixb); /* ------------- Set up the output map pixd --------------- */ /* Generate pixd, which is reduced by the factors (sx, sy). */ w = pixGetWidth(pixs); h = pixGetHeight(pixs); wd = (w + sx - 1) / sx; hd = (h + sy - 1) / sy; pixd = pixCreate(wd, hd, 8); /* Note: we only compute map values in tiles that are complete. * In general, tiles at right and bottom edges will not be * complete, and we must fill them in later. */ nx = w / sx; ny = h / sy; wpls = pixGetWpl(pixs); datas = pixGetData(pixs); wpld = pixGetWpl(pixd); datad = pixGetData(pixd); wplf = pixGetWpl(pixf); dataf = pixGetData(pixf); for (i = 0; i < ny; i++) { lines = datas + sy * i * wpls; linef = dataf + sy * i * wplf; lined = datad + i * wpld; for (j = 0; j < nx; j++) { delx = j * sx; sum = 0; count = 0; for (k = 0; k < sy; k++) { for (m = 0; m < sx; m++) { if (GET_DATA_BIT(linef + k * wplf, delx + m) == 0) { sum += GET_DATA_BYTE(lines + k * wpls, delx + m); count++; } } } if (count >= mincount) { val8 = sum / count; SET_DATA_BYTE(lined, j, val8); } } } pixDestroy(&pixf); /* If there is an optional mask with fg pixels, erase the previous * calculation for the corresponding map pixels, setting the * map values to 0. Then, when all the map holes are filled, * these erased pixels will be set by the surrounding map values. * * The calculation here is relatively efficient: for each pixel * in pixd (which corresponds to a tile of mask pixels in pixim) * we look only at the pixel in pixim that is at the center * of the tile. If the mask pixel is ON, we reset the map * pixel in pixd to 0, so that it can later be filled in. */ pixims = NULL; if (pixim && fgpixels) { wim = pixGetWidth(pixim); him = pixGetHeight(pixim); dataim = pixGetData(pixim); wplim = pixGetWpl(pixim); for (i = 0; i < ny; i++) { yim = i * sy + sy / 2; if (yim >= him) break; lineim = dataim + yim * wplim; for (j = 0; j < nx; j++) { xim = j * sx + sx / 2; if (xim >= wim) break; if (GET_DATA_BIT(lineim, xim)) pixSetPixel(pixd, j, i, 0); } } } /* Fill all the holes in the map. */ if (pixFillMapHoles(pixd, nx, ny, L_FILL_BLACK)) { pixDestroy(&pixd); L_WARNING("can't make the map", procName); return 1; } /* Finally, for each connected region corresponding to the * 'image' mask, reset all pixels to their average value. * Each of these components represents an image (or part of one) * in the input, and this smooths the background values * in each of these regions. */ if (pixim && fgpixels) { scalex = 1. / (l_float32)sx; scaley = 1. / (l_float32)sy; pixims = pixScaleBySampling(pixim, scalex, scaley); pixSmoothConnectedRegions(pixd, pixims, 2); pixDestroy(&pixims); } *ppixd = pixd; return 0; } /*! * pixGetBackgroundRGBMap() * * Input: pixs (32 bpp rgb) * pixim ( 1 bpp 'image' mask; can be null; it * should not have all foreground pixels) * pixg ( 8 bpp grayscale version; can be null) * sx, sy (tile size in pixels) * thresh (threshold for determining foreground) * mincount (min threshold on counts in a tile) * &pixmr, &pixmg, &pixmb ( rgb maps) * Return: 0 if OK, 1 on error * * Notes: * (1) If pixg, which is a grayscale version of pixs, is provided, * use this internally to generate the foreground mask. * Otherwise, a grayscale version of pixs will be generated * from the green component only, used, and destroyed. */ l_int32 pixGetBackgroundRGBMap(PIX *pixs, PIX *pixim, PIX *pixg, l_int32 sx, l_int32 sy, l_int32 thresh, l_int32 mincount, PIX **ppixmr, PIX **ppixmg, PIX **ppixmb) { l_int32 w, h, wm, hm, wim, him, wpls, wplim, wplf; l_int32 xim, yim, delx, nx, ny, i, j, k, m; l_int32 count, rsum, gsum, bsum, rval, gval, bval; l_int32 empty, fgpixels; l_uint32 pixel; l_uint32 *datas, *dataim, *dataf, *lines, *lineim, *linef; l_float32 scalex, scaley; PIX *piximi, *pixgc, *pixb, *pixf, *pixims; PIX *pixmr, *pixmg, *pixmb; PROCNAME("pixGetBackgroundRGBMap"); if (!ppixmr || !ppixmg || !ppixmb) return ERROR_INT("&pixm* not all defined", procName, 1); *ppixmr = *ppixmg = *ppixmb = NULL; if (!pixs) return ERROR_INT("pixs not defined", procName, 1); if (pixGetDepth(pixs) != 32) return ERROR_INT("pixs not 32 bpp", procName, 1); if (pixim && pixGetDepth(pixim) != 1) return ERROR_INT("pixim not 1 bpp", procName, 1); if (sx < 4 || sy < 4) return ERROR_INT("sx and sy must be >= 4", procName, 1); if (mincount > sx * sy) { L_WARNING("mincount too large for tile size", procName); mincount = (sx * sy) / 3; } /* Evaluate the mask pixim and make sure it is not all foreground */ fgpixels = 0; /* boolean for existence of fg mask pixels */ if (pixim) { piximi = pixInvert(NULL, pixim); /* set non-'image' pixels to 1 */ pixZero(piximi, &empty); pixDestroy(&piximi); if (empty) return ERROR_INT("pixim all fg; no background", procName, 1); pixZero(pixim, &empty); if (!empty) /* there are fg pixels in pixim */ fgpixels = 1; } /* Generate the foreground mask. These pixels will be * ignored when computing the background values. */ if (pixg) /* use the input grayscale version if it is provided */ pixgc = pixClone(pixg); else pixgc = pixConvertRGBToGrayFast(pixs); pixb = pixThresholdToBinary(pixgc, thresh); pixf = pixMorphSequence(pixb, "d7.1 + d1.7", 0); pixDestroy(&pixgc); pixDestroy(&pixb); /* Generate the output mask images */ w = pixGetWidth(pixs); h = pixGetHeight(pixs); wm = (w + sx - 1) / sx; hm = (h + sy - 1) / sy; pixmr = pixCreate(wm, hm, 8); pixmg = pixCreate(wm, hm, 8); pixmb = pixCreate(wm, hm, 8); /* ------------- Set up the mapping images --------------- */ /* Note: we only compute map values in tiles that are complete. * In general, tiles at right and bottom edges will not be * complete, and we must fill them in later. */ nx = w / sx; ny = h / sy; wpls = pixGetWpl(pixs); datas = pixGetData(pixs); wplf = pixGetWpl(pixf); dataf = pixGetData(pixf); for (i = 0; i < ny; i++) { lines = datas + sy * i * wpls; linef = dataf + sy * i * wplf; for (j = 0; j < nx; j++) { delx = j * sx; rsum = gsum = bsum = 0; count = 0; for (k = 0; k < sy; k++) { for (m = 0; m < sx; m++) { if (GET_DATA_BIT(linef + k * wplf, delx + m) == 0) { pixel = *(lines + k * wpls + delx + m); rsum += (pixel >> 24); gsum += ((pixel >> 16) & 0xff); bsum += ((pixel >> 8) & 0xff); count++; } } } if (count >= mincount) { rval = rsum / count; gval = gsum / count; bval = bsum / count; pixSetPixel(pixmr, j, i, rval); pixSetPixel(pixmg, j, i, gval); pixSetPixel(pixmb, j, i, bval); } } } pixDestroy(&pixf); /* If there is an optional mask with fg pixels, erase the previous * calculation for the corresponding map pixels, setting the * map values in each of the 3 color maps to 0. Then, when * all the map holes are filled, these erased pixels will * be set by the surrounding map values. */ if (pixim) { wim = pixGetWidth(pixim); him = pixGetHeight(pixim); dataim = pixGetData(pixim); wplim = pixGetWpl(pixim); for (i = 0; i < ny; i++) { yim = i * sy + sy / 2; if (yim >= him) break; lineim = dataim + yim * wplim; for (j = 0; j < nx; j++) { xim = j * sx + sx / 2; if (xim >= wim) break; if (GET_DATA_BIT(lineim, xim)) { pixSetPixel(pixmr, j, i, 0); pixSetPixel(pixmg, j, i, 0); pixSetPixel(pixmb, j, i, 0); } } } } /* ----------------- Now fill in the holes ----------------------- */ if (pixFillMapHoles(pixmr, nx, ny, L_FILL_BLACK) || pixFillMapHoles(pixmg, nx, ny, L_FILL_BLACK) || pixFillMapHoles(pixmb, nx, ny, L_FILL_BLACK)) { pixDestroy(&pixmr); pixDestroy(&pixmg); pixDestroy(&pixmb); L_WARNING("can't make the maps", procName); return 1; } /* Finally, for each connected region corresponding to the * fg mask, reset all pixels to their average value. */ if (pixim && fgpixels) { scalex = 1. / (l_float32)sx; scaley = 1. / (l_float32)sy; pixims = pixScaleBySampling(pixim, scalex, scaley); pixSmoothConnectedRegions(pixmr, pixims, 2); pixSmoothConnectedRegions(pixmg, pixims, 2); pixSmoothConnectedRegions(pixmb, pixims, 2); pixDestroy(&pixims); } *ppixmr = pixmr; *ppixmg = pixmg; *ppixmb = pixmb; return 0; } /*! * pixGetBackgroundGrayMapMorph() * * Input: pixs (8 bpp) * pixim ( 1 bpp 'image' mask; can be null; it * should not have all foreground pixels) * reduction (factor at which closing is performed) * size (of square Sel for the closing; use an odd number) * &pixm ( grayscale map) * Return: 0 if OK, 1 on error */ l_int32 pixGetBackgroundGrayMapMorph(PIX *pixs, PIX *pixim, l_int32 reduction, l_int32 size, PIX **ppixm) { l_int32 nx, ny, empty, fgpixels; l_float32 scale; PIX *pixm, *pixt1, *pixt2, *pixt3, *pixims; PROCNAME("pixGetBackgroundGrayMapMorph"); if (!ppixm) return ERROR_INT("&pixm not defined", procName, 1); *ppixm = NULL; if (!pixs) return ERROR_INT("pixs not defined", procName, 1); if (pixGetDepth(pixs) != 8) return ERROR_INT("pixs not 8 bpp", procName, 1); if (pixim && pixGetDepth(pixim) != 1) return ERROR_INT("pixim not 1 bpp", procName, 1); /* Evaluate the mask pixim and make sure it is not all foreground. */ fgpixels = 0; /* boolean for existence of fg mask pixels */ if (pixim) { pixInvert(pixim, pixim); /* set background pixels to 1 */ pixZero(pixim, &empty); if (empty) return ERROR_INT("pixim all fg; no background", procName, 1); pixInvert(pixim, pixim); /* revert to original mask */ pixZero(pixim, &empty); if (!empty) /* there are fg pixels in pixim */ fgpixels = 1; } /* Downscale as requested and do the closing to get the background. */ scale = 1. / (l_float32)reduction; pixt1 = pixScaleBySampling(pixs, scale, scale); pixt2 = pixCloseGray(pixt1, size, size); pixt3 = pixExtendByReplication(pixt2, 1, 1); /* Downscale the image mask, if any, and remove it from the * background. These pixels will be filled in (twice). */ pixims = NULL; if (pixim) { pixims = pixScale(pixim, scale, scale); pixm = pixConvertTo8(pixims, FALSE); pixAnd(pixm, pixm, pixt3); } else pixm = pixClone(pixt3); pixDestroy(&pixt1); pixDestroy(&pixt2); pixDestroy(&pixt3); /* Fill all the holes in the map. */ nx = pixGetWidth(pixs) / reduction; ny = pixGetHeight(pixs) / reduction; if (pixFillMapHoles(pixm, nx, ny, L_FILL_BLACK)) { pixDestroy(&pixm); L_WARNING("can't make the map", procName); return 1; } /* Finally, for each connected region corresponding to the * fg mask, reset all pixels to their average value. */ if (pixim && fgpixels) { pixSmoothConnectedRegions(pixm, pixims, 2); pixDestroy(&pixims); } *ppixm = pixm; return 0; } /*! * pixGetBackgroundRGBMapMorph() * * Input: pixs (32 bpp rgb) * pixim ( 1 bpp 'image' mask; can be null; it * should not have all foreground pixels) * reduction (factor at which closing is performed) * size (of square Sel for the closing; use an odd number) * &pixmr ( red component map) * &pixmg ( green component map) * &pixmb ( blue component map) * Return: 0 if OK, 1 on error */ l_int32 pixGetBackgroundRGBMapMorph(PIX *pixs, PIX *pixim, l_int32 reduction, l_int32 size, PIX **ppixmr, PIX **ppixmg, PIX **ppixmb) { l_int32 nx, ny, empty, fgpixels; l_float32 scale; PIX *pixm, *pixmr, *pixmg, *pixmb, *pixt1, *pixt2, *pixt3, *pixims; PROCNAME("pixGetBackgroundRGBMapMorph"); if (!ppixmr || !ppixmg || !ppixmb) return ERROR_INT("&pixm* not all defined", procName, 1); *ppixmr = *ppixmg = *ppixmb = NULL; if (!pixs) return ERROR_INT("pixs not defined", procName, 1); if (pixGetDepth(pixs) != 32) return ERROR_INT("pixs not 32 bpp", procName, 1); if (pixim && pixGetDepth(pixim) != 1) return ERROR_INT("pixim not 1 bpp", procName, 1); /* Generate an 8 bpp version of the image mask, if it exists */ scale = 1. / (l_float32)reduction; pixm = NULL; if (pixim) { pixims = pixScale(pixim, scale, scale); pixm = pixConvertTo8(pixims, FALSE); } /* Evaluate the mask pixim and make sure it is not all foreground. */ fgpixels = 0; /* boolean for existence of fg mask pixels */ if (pixim) { pixInvert(pixim, pixim); /* set background pixels to 1 */ pixZero(pixim, &empty); if (empty) return ERROR_INT("pixim all fg; no background", procName, 1); pixInvert(pixim, pixim); /* revert to original mask */ pixZero(pixim, &empty); if (!empty) /* there are fg pixels in pixim */ fgpixels = 1; } /* Downscale as requested and do the closing to get the background. * Then remove the image mask pixels from the background. They * will be filled in (twice) later. Do this for all 3 components. */ pixt1 = pixScaleRGBToGrayFast(pixs, reduction, COLOR_RED); pixt2 = pixCloseGray(pixt1, size, size); pixt3 = pixExtendByReplication(pixt2, 1, 1); if (pixim) pixmr = pixAnd(NULL, pixm, pixt3); else pixmr = pixClone(pixt3); pixDestroy(&pixt1); pixDestroy(&pixt2); pixDestroy(&pixt3); pixt1 = pixScaleRGBToGrayFast(pixs, reduction, COLOR_GREEN); pixt2 = pixCloseGray(pixt1, size, size); pixt3 = pixExtendByReplication(pixt2, 1, 1); if (pixim) pixmg = pixAnd(NULL, pixm, pixt3); else pixmg = pixClone(pixt3); pixDestroy(&pixt1); pixDestroy(&pixt2); pixDestroy(&pixt3); pixt1 = pixScaleRGBToGrayFast(pixs, reduction, COLOR_BLUE); pixt2 = pixCloseGray(pixt1, size, size); pixt3 = pixExtendByReplication(pixt2, 1, 1); if (pixim) pixmb = pixAnd(NULL, pixm, pixt3); else pixmb = pixClone(pixt3); pixDestroy(&pixm); pixDestroy(&pixt1); pixDestroy(&pixt2); pixDestroy(&pixt3); /* Fill all the holes in the three maps. */ nx = pixGetWidth(pixs) / reduction; ny = pixGetHeight(pixs) / reduction; if (pixFillMapHoles(pixmr, nx, ny, L_FILL_BLACK) || pixFillMapHoles(pixmg, nx, ny, L_FILL_BLACK) || pixFillMapHoles(pixmb, nx, ny, L_FILL_BLACK)) { pixDestroy(&pixmr); pixDestroy(&pixmg); pixDestroy(&pixmb); L_WARNING("can't make the maps", procName); return 1; } /* Finally, for each connected region corresponding to the * fg mask in each component, reset all pixels to their * average value. */ if (pixim && fgpixels) { pixSmoothConnectedRegions(pixmr, pixims, 2); pixSmoothConnectedRegions(pixmg, pixims, 2); pixSmoothConnectedRegions(pixmb, pixims, 2); pixDestroy(&pixims); } *ppixmr = pixmr; *ppixmg = pixmg; *ppixmb = pixmb; return 0; } /*! * pixFillMapHoles() * * Input: pix (8 bpp; a map, with one pixel for each tile in * a larger image) * nx (number of horizontal pixel tiles that are entirely * covered with pixels in the original source image) * ny (ditto for the number of vertical pixel tiles) * filltype (L_FILL_WHITE or L_FILL_BLACK) * Return: 0 if OK, 1 on error * * Notes: * (1) This is an in-place operation on pix (the map). pix is * typically a low-resolution version of some other image * from which it was derived, where each pixel in pix * corresponds to a rectangular tile (say, m x n) of pixels * in the larger image. All we need to know about the larger * image is whether or not the rightmost column and bottommost * row of pixels in pix correspond to tiles that are * only partially covered by pixels in the larger image. * (2) Typically, some number of pixels in the input map are * not known, and their values must be determined by near * pixels that are known. These unknown pixels are the 'holes'. * They can take on only two values, 0 and 255, and the * instruction about which to fill is given by the filltype flag. * (3) The "holes" can come from two sources. The first is when there * are not enough foreground or background pixels in a tile; * the second is when a tile is at least partially covered * by an image mask. If we're filling holes in a fg mask, * the holes are initialized to black (0) and use L_FILL_BLACK. * For filling holes in a bg mask, initialize the holes to * white (255) and use L_FILL_WHITE. * (4) If w is the map width, nx = w or nx = w - 1; ditto for h and ny. */ l_int32 pixFillMapHoles(PIX *pix, l_int32 nx, l_int32 ny, l_int32 filltype) { l_int32 w, h, d, y, nmiss, goodcol, i, j, found, ival, valtest; l_uint32 val, lastval; NUMA *na; /* indicates if there is any data in the column */ PIX *pixt; PROCNAME("pixFillMapHoles"); if (!pix) return ERROR_INT("pix not defined", procName, 1); pixGetDimensions(pix, &w, &h, &d); if (d != 8) return ERROR_INT("pix not 8 bpp", procName, 1); /* ------------- Fill holes in the mapping image columns ----------- */ na = numaCreate(0); /* holds flag for which columns have data */ nmiss = 0; valtest = (filltype == L_FILL_WHITE) ? 255 : 0; for (j = 0; j < nx; j++) { /* do it by columns */ found = FALSE; for (i = 0; i < ny; i++) { pixGetPixel(pix, j, i, &val); if (val != valtest) { y = i; found = TRUE; break; } } if (found == FALSE) { numaAddNumber(na, 0); /* no data in the column */ nmiss++; } else { numaAddNumber(na, 1); /* data in the column */ for (i = y - 1; i >= 0; i--) /* replicate upwards to top */ pixSetPixel(pix, j, i, val); pixGetPixel(pix, j, 0, &lastval); for (i = 1; i < h; i++) { /* set going down to bottom */ pixGetPixel(pix, j, i, &val); if (val == valtest) pixSetPixel(pix, j, i, lastval); else lastval = val; } } } numaAddNumber(na, 0); /* last column */ if (nmiss == nx) { /* no data in any column! */ numaDestroy(&na); L_WARNING("no bg found; no data in any column", procName); return 1; } /* ---------- Fill in missing columns by replication ----------- */ if (nmiss > 0) { /* replicate columns */ pixt = pixCopy(NULL, pix); /* Find the first good column */ goodcol = 0; for (j = 0; j < w; j++) { numaGetIValue(na, j, &ival); if (ival == 1) { goodcol = j; break; } } if (goodcol > 0) { /* copy cols backward */ for (j = goodcol - 1; j >= 0; j--) { pixRasterop(pix, j, 0, 1, h, PIX_SRC, pixt, j + 1, 0); pixRasterop(pixt, j, 0, 1, h, PIX_SRC, pix, j, 0); } } for (j = goodcol + 1; j < w; j++) { /* copy cols forward */ numaGetIValue(na, j, &ival); if (ival == 0) { /* Copy the column to the left of j */ pixRasterop(pix, j, 0, 1, h, PIX_SRC, pixt, j - 1, 0); pixRasterop(pixt, j, 0, 1, h, PIX_SRC, pix, j, 0); } } pixDestroy(&pixt); } if (w > nx) { /* replicate the last column */ for (i = 0; i < h; i++) { pixGetPixel(pix, w - 2, i, &val); pixSetPixel(pix, w - 1, i, val); } } numaDestroy(&na); return 0; } /*! * pixExtendByReplication() * * Input: pixs (8 bpp) * addw (number of extra pixels horizontally to add) * addh (number of extra pixels vertically to add) * Return: pixd (extended with replicated pixel values), or null on error * * Notes: * (1) The pixel values are extended to the left and down, as required. */ PIX * pixExtendByReplication(PIX *pixs, l_int32 addw, l_int32 addh) { l_int32 w, h, i, j; l_uint32 val; PIX *pixd; PROCNAME("pixExtendByReplication"); if (!pixs) return (PIX *)ERROR_PTR("pixs not defined", procName, NULL); if (pixGetDepth(pixs) != 8) return (PIX *)ERROR_PTR("pixs not 8 bpp", procName, NULL); if (addw == 0 && addh == 0) return pixCopy(NULL, pixs); w = pixGetWidth(pixs); h = pixGetHeight(pixs); if ((pixd = pixCreate(w + addw, h + addh, 8)) == NULL) return (PIX *)ERROR_PTR("pixd not made", procName, NULL); pixRasterop(pixd, 0, 0, w, h, PIX_SRC, pixs, 0, 0); if (addw > 0) { for (i = 0; i < h; i++) { pixGetPixel(pixd, w - 1, i, &val); for (j = 0; j < addw; j++) pixSetPixel(pixd, w + j, i, val); } } if (addh > 0) { for (j = 0; j < w + addw; j++) { pixGetPixel(pixd, j, h - 1, &val); for (i = 0; i < addh; i++) pixSetPixel(pixd, j, h + i, val); } } return pixd; } /*! * pixSmoothConnectedRegions() * * Input: pixs (8 bpp) * pixm ( 1 bpp; if null, this is a no-op) * factor (subsampling factor for getting average; >= 1) * Return: 0 if OK, 1 on error * * Notes: * (1) The pixels in pixs corresponding to those in each * 8-connected region in the mask are set to the average value. * (2) This is required for adaptive mapping to avoid the * generation of stripes in the background map, due to * variations in the pixel values near the edges of mask regions. * (3) This function is optimized for background smoothing, where * there are a relatively small number of components. It will * be inefficient if used where there are many small components. */ l_int32 pixSmoothConnectedRegions(PIX *pixs, PIX *pixm, l_int32 factor) { l_int32 empty, i, n, x, y; l_float32 aveval; BOXA *boxa; PIX *pixmc; PIXA *pixa; PROCNAME("pixSmoothConnectedRegions"); if (!pixs) return ERROR_INT("pixs not defined", procName, 1); if (pixGetDepth(pixs) != 8) return ERROR_INT("pixs not 8 bpp", procName, 1); if (!pixm) { L_INFO("pixm not defined", procName); return 0; } if (pixGetDepth(pixm) != 1) return ERROR_INT("pixm not 1 bpp", procName, 1); pixZero(pixm, &empty); if (empty) { L_INFO("pixm has no fg pixels; nothing to do", procName); return 0; } boxa = pixConnComp(pixm, &pixa, 8); n = boxaGetCount(boxa); for (i = 0; i < n; i++) { if ((pixmc = pixaGetPix(pixa, i, L_CLONE)) == NULL) { L_WARNING("missing pixmc!", procName); continue; } boxaGetBoxGeometry(boxa, i, &x, &y, NULL, NULL); pixGetAverageMasked(pixs, pixmc, x, y, factor, L_MEAN_ABSVAL, &aveval); pixPaintThroughMask(pixs, pixmc, x, y, (l_int32)aveval); pixDestroy(&pixmc); } boxaDestroy(&boxa); pixaDestroy(&pixa); return 0; } /*------------------------------------------------------------------* * Measurement of local foreground * *------------------------------------------------------------------*/ #if 0 /* Not working properly: do not use */ /*! * pixGetForegroundGrayMap() * * Input: pixs (8 bpp) * pixim ( 1 bpp 'image' mask; can be null) * sx, sy (src tile size, in pixels) * thresh (threshold for determining foreground) * &pixd ( 8 bpp grayscale map) * Return: 0 if OK, 1 on error * * Notes: * (1) Each (sx, sy) tile of pixs gets mapped to one pixel in pixd. * (2) pixd is the estimate of the fg (darkest) value within each tile. * (3) All pixels in pixd that are in 'image' regions, as specified * by pixim, are given the background value 0. * (4) For pixels in pixd that can't directly be given a fg value, * the value is inferred by propagating from neighboring pixels. * (5) In practice, pixd can be used to normalize the fg, and * it can be done after background normalization. * (6) The overall procedure is: * - reduce 2x by sampling * - paint all 'image' pixels white, so that they don't * participate in the Min reduction * - do a further (sx, sy) Min reduction -- think of * it as a large opening followed by subsampling by the * reduction factors * - threshold the result to identify fg, and set the * bg pixels to 255 (these are 'holes') * - fill holes by propagation from fg values * - replicatively expand by 2x, arriving at the final * resolution of pixd * - smooth with a 17x17 kernel * - paint the 'image' regions black */ l_int32 pixGetForegroundGrayMap(PIX *pixs, PIX *pixim, l_int32 sx, l_int32 sy, l_int32 thresh, PIX **ppixd) { l_int32 w, h, d, wd, hd; l_int32 empty, fgpixels; PIX *pixd, *piximi, *pixim2, *pixims, *pixs2, *pixb, *pixt1, *pixt2, *pixt3; PROCNAME("pixGetForegroundGrayMap"); if (!ppixd) return ERROR_INT("&pixd not defined", procName, 1); *ppixd = NULL; if (!pixs) return ERROR_INT("pixs not defined", procName, 1); pixGetDimensions(pixs, &w, &h, &d); if (d != 8) return ERROR_INT("pixs not 8 bpp", procName, 1); if (pixim && pixGetDepth(pixim) != 1) return ERROR_INT("pixim not 1 bpp", procName, 1); if (sx < 2 || sy < 2) return ERROR_INT("sx and sy must be >= 2", procName, 1); /* Generate pixd, which is reduced by the factors (sx, sy). */ wd = (w + sx - 1) / sx; hd = (h + sy - 1) / sy; pixd = pixCreate(wd, hd, 8); *ppixd = pixd; /* Evaluate the 'image' mask, pixim. If it is all fg, * the output pixd has all pixels with value 0. */ fgpixels = 0; /* boolean for existence of fg pixels in the image mask. */ if (pixim) { piximi = pixInvert(NULL, pixim); /* set non-image pixels to 1 */ pixZero(piximi, &empty); pixDestroy(&piximi); if (empty) /* all 'image'; return with all pixels set to 0 */ return 0; pixZero(pixim, &empty); if (!empty) /* there are fg pixels in pixim */ fgpixels = 1; } /* 2x subsampling; paint white through 'image' mask. */ pixs2 = pixScaleBySampling(pixs, 0.5, 0.5); if (pixim && fgpixels) { pixim2 = pixReduceBinary2(pixim, NULL); pixPaintThroughMask(pixs2, pixim2, 0, 0, 255); pixDestroy(&pixim2); } /* Min (erosion) downscaling; total reduction (4 sx, 4 sy). */ pixt1 = pixScaleGrayMinMax(pixs2, sx, sy, L_CHOOSE_MIN); /* pixDisplay(pixt1, 300, 200); */ /* Threshold to identify fg; paint bg pixels to white. */ pixb = pixThresholdToBinary(pixt1, thresh); /* fg pixels */ pixInvert(pixb, pixb); pixPaintThroughMask(pixt1, pixb, 0, 0, 255); pixDestroy(&pixb); /* Replicative expansion by 2x to (sx, sy). */ pixt2 = pixExpandReplicate(pixt1, 2); /* pixDisplay(pixt2, 500, 200); */ /* Fill holes in the fg by propagation */ pixFillMapHoles(pixt2, w / sx, h / sy, L_FILL_WHITE); /* pixDisplay(pixt2, 700, 200); */ /* Smooth with 17x17 kernel. */ pixt3 = pixBlockconv(pixt2, 8, 8); pixRasterop(pixd, 0, 0, wd, hd, PIX_SRC, pixt3, 0, 0); /* Paint the image parts black. */ pixims = pixScaleBySampling(pixim, 1. / sx, 1. / sy); pixPaintThroughMask(pixd, pixims, 0, 0, 0); pixDestroy(&pixs2); pixDestroy(&pixt1); pixDestroy(&pixt2); pixDestroy(&pixt3); return 0; } #endif /* Not working properly: do not use */ /*------------------------------------------------------------------* * Generate inverted background map * *------------------------------------------------------------------*/ /*! * pixGetInvBackgroundMap() * * Input: pixs (8 bpp) * bgval (target bg val; typ. > 128) * smoothx (half-width of block convolution kernel width) * smoothy (half-width of block convolution kernel height) * Return: pixd (16 bpp), or null on error * * Note: * - bgval should typically be > 120 and < 240 * - pixd is a normalization image; the original image is * multiplied by pixd and the result is divided by 256. */ PIX * pixGetInvBackgroundMap(PIX *pixs, l_int32 bgval, l_int32 smoothx, l_int32 smoothy) { l_int32 w, h, wplsm, wpld, i, j; l_int32 val, val16; l_uint32 *datasm, *datad, *linesm, *lined; PIX *pixsm, *pixd; PROCNAME("pixGetInvBackgroundMap"); if (!pixs) return (PIX *)ERROR_PTR("pixs not defined", procName, NULL); if (pixGetDepth(pixs) != 8) return (PIX *)ERROR_PTR("pixs not 8 bpp", procName, NULL); w = pixGetWidth(pixs); h = pixGetHeight(pixs); if (w < 5 || h < 5) return (PIX *)ERROR_PTR("w and h must be >= 5", procName, NULL); /* smooth the map image */ pixsm = pixBlockconv(pixs, smoothx, smoothy); datasm = pixGetData(pixsm); wplsm = pixGetWpl(pixsm); /* invert the map image, scaling up to preserve dynamic range */ pixd = pixCreate(w, h, 16); datad = pixGetData(pixd); wpld = pixGetWpl(pixd); for (i = 0; i < h; i++) { linesm = datasm + i * wplsm; lined = datad + i * wpld; for (j = 0; j < w; j++) { val = GET_DATA_BYTE(linesm, j); if (val > 0) val16 = (256 * bgval) / val; else { /* shouldn't happen */ L_WARNING("smoothed bg has 0 pixel!", procName); val16 = bgval / 2; } SET_DATA_TWO_BYTES(lined, j, val16); } } pixDestroy(&pixsm); return pixd; } /*------------------------------------------------------------------* * Apply background map to image * *------------------------------------------------------------------*/ /*! * pixApplyInvBackgroundGrayMap() * * Input: pixs (8 bpp) * pixm (16 bpp, inverse background map) * sx (tile width in pixels) * sy (tile height in pixels) * Return: pixd (8 bpp), or null on error */ PIX * pixApplyInvBackgroundGrayMap(PIX *pixs, PIX *pixm, l_int32 sx, l_int32 sy) { l_int32 w, h, wm, hm, wpls, wpld, i, j, k, m, xoff, yoff; l_int32 vals, vald; l_uint32 val16; l_uint32 *datas, *datad, *lines, *lined, *flines, *flined; PIX *pixd; PROCNAME("pixApplyInvBackgroundGrayMap"); if (!pixs) return (PIX *)ERROR_PTR("pixs not defined", procName, NULL); if (pixGetDepth(pixs) != 8) return (PIX *)ERROR_PTR("pixs not 8 bpp", procName, NULL); if (!pixm) return (PIX *)ERROR_PTR("pixm not defined", procName, NULL); if (pixGetDepth(pixm) != 16) return (PIX *)ERROR_PTR("pixm not 16 bpp", procName, NULL); if (sx == 0 || sy == 0) return (PIX *)ERROR_PTR("invalid sx and/or sy", procName, NULL); datas = pixGetData(pixs); wpls = pixGetWpl(pixs); w = pixGetWidth(pixs); h = pixGetHeight(pixs); wm = pixGetWidth(pixm); hm = pixGetHeight(pixm); pixd = pixCreateTemplate(pixs); datad = pixGetData(pixd); wpld = pixGetWpl(pixd); for (i = 0; i < hm; i++) { lines = datas + sy * i * wpls; lined = datad + sy * i * wpld; yoff = sy * i; for (j = 0; j < wm; j++) { pixGetPixel(pixm, j, i, &val16); xoff = sx * j; for (k = 0; k < sy && yoff + k < h; k++) { flines = lines + k * wpls; flined = lined + k * wpld; for (m = 0; m < sx && xoff + m < w; m++) { vals = GET_DATA_BYTE(flines, xoff + m); vald = (vals * val16) / 256; vald = L_MIN(vald, 255); SET_DATA_BYTE(flined, xoff + m, vald); } } } } return pixd; } /*! * pixApplyInvBackgroundRGBMap() * * Input: pixs (32 bpp rbg) * pixmr (16 bpp, red inverse background map) * pixmg (16 bpp, green inverse background map) * pixmb (16 bpp, blue inverse background map) * sx (tile width in pixels) * sy (tile height in pixels) * Return: pixd (32 bpp rbg), or null on error */ PIX * pixApplyInvBackgroundRGBMap(PIX *pixs, PIX *pixmr, PIX *pixmg, PIX *pixmb, l_int32 sx, l_int32 sy) { l_int32 w, h, wm, hm, wpls, wpld, i, j, k, m, xoff, yoff; l_int32 rvald, gvald, bvald; l_uint32 vals; l_uint32 rval16, gval16, bval16; l_uint32 *datas, *datad, *lines, *lined, *flines, *flined; PIX *pixd; PROCNAME("pixApplyInvBackgroundRGBMap"); if (!pixs) return (PIX *)ERROR_PTR("pixs not defined", procName, NULL); if (pixGetDepth(pixs) != 32) return (PIX *)ERROR_PTR("pixs not 32 bpp", procName, NULL); if (!pixmr || !pixmg || !pixmb) return (PIX *)ERROR_PTR("pix maps not all defined", procName, NULL); if (pixGetDepth(pixmr) != 16 || pixGetDepth(pixmg) != 16 || pixGetDepth(pixmb) != 16) return (PIX *)ERROR_PTR("pix maps not all 16 bpp", procName, NULL); if (sx == 0 || sy == 0) return (PIX *)ERROR_PTR("invalid sx and/or sy", procName, NULL); datas = pixGetData(pixs); wpls = pixGetWpl(pixs); w = pixGetWidth(pixs); h = pixGetHeight(pixs); wm = pixGetWidth(pixmr); hm = pixGetHeight(pixmr); pixd = pixCreateTemplate(pixs); datad = pixGetData(pixd); wpld = pixGetWpl(pixd); for (i = 0; i < hm; i++) { lines = datas + sy * i * wpls; lined = datad + sy * i * wpld; yoff = sy * i; for (j = 0; j < wm; j++) { pixGetPixel(pixmr, j, i, &rval16); pixGetPixel(pixmg, j, i, &gval16); pixGetPixel(pixmb, j, i, &bval16); xoff = sx * j; for (k = 0; k < sy && yoff + k < h; k++) { flines = lines + k * wpls; flined = lined + k * wpld; for (m = 0; m < sx && xoff + m < w; m++) { vals = *(flines + xoff + m); rvald = ((vals >> 24) * rval16) / 256; rvald = L_MIN(rvald, 255); gvald = (((vals >> 16) & 0xff) * gval16) / 256; gvald = L_MIN(gvald, 255); bvald = (((vals >> 8) & 0xff) * bval16) / 256; bvald = L_MIN(bvald, 255); composeRGBPixel(rvald, gvald, bvald, flined + xoff + m); } } } } return pixd; } /*------------------------------------------------------------------* * Apply variable map * *------------------------------------------------------------------*/ /*! * pixApplyVariableGrayMap() * * Input: pixs (8 bpp) * pixg (8 bpp, variable map) * target (typ. 128 for threshold) * Return: pixd (8 bpp), or null on error * * Notes: * (1) Suppose you have an image that you want to transform based * on some photometric measurement at each point, such as the * threshold value for binarization. Representing the photometric * measurement as an image pixg, you can threshold in input image * using pixVarThresholdToBinary(). Alternatively, you can map * the input image pointwise so that the threshold over the * entire image becomes a constant, such as 128. For example, * if a pixel in pixg is 150 and the target is 128, the * corresponding pixel in pixs is mapped linearly to a value * (128/150) of the input value. If the resulting mapped image * pixd were then thresholded at 128, you would obtain the * same result as a direct binarization using pixg with * pixVarThresholdToBinary(). * (2) The sizes of pixs and pixg must be equal. */ PIX * pixApplyVariableGrayMap(PIX *pixs, PIX *pixg, l_int32 target) { l_int32 i, j, w, h, d, wpls, wplg, wpld, vals, valg, vald; l_uint8 *lut; l_uint32 *datas, *datag, *datad, *lines, *lineg, *lined; l_float32 fval; PIX *pixd; PROCNAME("pixApplyVariableGrayMap"); if (!pixs) return (PIX *)ERROR_PTR("pixs not defined", procName, NULL); if (!pixg) return (PIX *)ERROR_PTR("pixg not defined", procName, NULL); if (!pixSizesEqual(pixs, pixg)) return (PIX *)ERROR_PTR("pix sizes not equal", procName, NULL); pixGetDimensions(pixs, &w, &h, &d); if (d != 8) return (PIX *)ERROR_PTR("depth not 8 bpp", procName, NULL); /* Generate a LUT for the mapping if the image is large enough * to warrant the overhead. The LUT is of size 2^16. For the * index to the table, get the MSB from pixs and the LSB from pixg. * Note: this LUT is bigger than the typical 32K L1 cache, so * we expect cache misses. L2 latencies are about 5ns. But * division is slooooow. For large images, this function is about * 4x faster when using the LUT. C'est la vie. */ lut = NULL; if (w * h > 100000) { /* more pixels than 2^16 */ if ((lut = (l_uint8 *)CALLOC(0x10000, sizeof(l_uint8))) == NULL) return (PIX *)ERROR_PTR("lut not made", procName, NULL); for (i = 0; i < 256; i++) { for (j = 0; j < 256; j++) { fval = (l_float32)(i * target) / (j + 0.5); lut[(i << 8) + j] = L_MIN(255, (l_int32)(fval + 0.5)); } } } pixd = pixCreateNoInit(w, h, 8); datad = pixGetData(pixd); wpld = pixGetWpl(pixd); datas = pixGetData(pixs); wpls = pixGetWpl(pixs); datag = pixGetData(pixg); wplg = pixGetWpl(pixg); for (i = 0; i < h; i++) { lines = datas + i * wpls; lineg = datag + i * wplg; lined = datad + i * wpld; if (lut) { for (j = 0; j < w; j++) { vals = GET_DATA_BYTE(lines, j); valg = GET_DATA_BYTE(lineg, j); vald = lut[(vals << 8) + valg]; SET_DATA_BYTE(lined, j, vald); } } else { for (j = 0; j < w; j++) { vals = GET_DATA_BYTE(lines, j); valg = GET_DATA_BYTE(lineg, j); fval = (l_float32)(vals * target) / (valg + 0.5); vald = L_MIN(255, (l_int32)(fval + 0.5)); SET_DATA_BYTE(lined, j, vald); } } } if (lut) FREE(lut); return pixd; } /*------------------------------------------------------------------* * Non-adaptive (global) mapping * *------------------------------------------------------------------*/ /*! * pixGlobalNormRGB() * * Input: pixd ( null, existing or equal to pixs) * pixs (32 bpp rgb, or colormapped) * rval, gval, bval (pixel values in pixs that are * linearly mapped to mapval) * mapval (use 255 for mapping to white) * Return: pixd (32 bpp rgb or colormapped), or null on error * * Notes: * (1) The value of pixd determines if the results are written to a * new pix (use NULL), in-place to pixs (use pixs), or to some * other existing pix. * (2) This does a global normalization of an image where the * r,g,b color components are not balanced. Thus, white in pixs is * represented by a set of r,g,b values that are not all 255. * (3) The input values (rval, gval, bval) should be chosen to * represent the gray color (mapval, mapval, mapval) in src. * Thus, this function will map (rval, gval, bval) to that gray color. * (4) Typically, mapval = 255, so that (rval, gval, bval) * corresponds to the white point of src. In that case, these * parameters should be chosen so that few pixels have higher values. * (5) In all cases, we do a linear TRC separately on each of the * components, saturating at 255. * (6) If the input pix is 8 bpp without a colormap, you can get * this functionality with mapval = 255 by calling: * pixGammaTRC(pixd, pixs, 1.0, 0, bgval); * where bgval is the value you want to be mapped to 255. * Or more generally, if you want bgval to be mapped to mapval: * pixGammaTRC(pixd, pixs, 1.0, 0, 255 * bgval / mapval); */ PIX * pixGlobalNormRGB(PIX *pixd, PIX *pixs, l_int32 rval, l_int32 gval, l_int32 bval, l_int32 mapval) { l_int32 w, h, d, i, j, ncolors, rv, gv, bv, wpl; l_int32 *rarray, *garray, *barray; l_uint32 *data, *line; NUMA *nar, *nag, *nab; PIXCMAP *cmap; PROCNAME("pixGlobalNormRGB"); if (!pixs) return (PIX *)ERROR_PTR("pixs not defined", procName, NULL); cmap = pixGetColormap(pixs); pixGetDimensions(pixs, &w, &h, &d); if (!cmap && d != 32) return (PIX *)ERROR_PTR("pixs not cmapped or 32 bpp", procName, NULL); if (mapval <= 0) { L_WARNING("mapval must be > 0; setting to 255", procName); mapval = 255; } /* Prepare pixd to be a copy of pixs */ if ((pixd = pixCopy(pixd, pixs)) == NULL) return (PIX *)ERROR_PTR("pixd not made", procName, NULL); /* Generate the TRC maps for each component. Make sure the * upper range for each color is greater than zero. */ nar = numaGammaTRC(1.0, 0, L_MAX(1, 255 * rval / mapval)); nag = numaGammaTRC(1.0, 0, L_MAX(1, 255 * gval / mapval)); nab = numaGammaTRC(1.0, 0, L_MAX(1, 255 * bval / mapval)); if (!nar || !nag || !nab) return (PIX *)ERROR_PTR("trc maps not all made", procName, pixd); /* Extract copies of the internal arrays */ rarray = numaGetIArray(nar); garray = numaGetIArray(nag); barray = numaGetIArray(nab); if (!rarray || !garray || !barray) return (PIX *)ERROR_PTR("*arrays not all made", procName, pixd); if (cmap) { ncolors = pixcmapGetCount(cmap); for (i = 0; i < ncolors; i++) { pixcmapGetColor(cmap, i, &rv, &gv, &bv); pixcmapResetColor(cmap, i, rarray[rv], garray[gv], barray[bv]); } } else { data = pixGetData(pixd); wpl = pixGetWpl(pixd); for (i = 0; i < h; i++) { line = data + i * wpl; for (j = 0; j < w; j++) { extractRGBValues(line[j], &rv, &gv, &bv); composeRGBPixel(rarray[rv], garray[gv], barray[bv], line + j); } } } numaDestroy(&nar); numaDestroy(&nag); numaDestroy(&nab); FREE(rarray); FREE(garray); FREE(barray); return pixd; } /*! * pixGlobalNormNoSatRGB() * * Input: pixd ( null, existing or equal to pixs) * pixs (32 bpp rgb) * rval, gval, bval (pixel values in pixs that are * linearly mapped to mapval; but see below) * factor (subsampling factor; integer >= 1) * rank (between 0.0 and 1.0; typ. use a value near 1.0) * Return: pixd (32 bpp rgb), or null on error * * Notes: * (1) This is a version of pixGlobalNormRGB(), where the output * intensity is scaled back so that a controlled fraction of * pixel components is allowed to saturate. See comments in * pixGlobalNormRGB(). * (2) The value of pixd determines if the results are written to a * new pix (use NULL), in-place to pixs (use pixs), or to some * other existing pix. * (3) This does a global normalization of an image where the * r,g,b color components are not balanced. Thus, white in pixs is * represented by a set of r,g,b values that are not all 255. * (4) The input values (rval, gval, bval) can be chosen to be the * color that, after normalization, becomes white background. * For images that are mostly background, the closer these values * are to the median component values, the closer the resulting * background will be to gray, becoming white at the brightest places. * (5) The mapval used in pixGlobalNormRGB() is computed here to * avoid saturation of any component in the image (save for a * fraction of the pixels given by the input rank value). */ PIX * pixGlobalNormNoSatRGB(PIX *pixd, PIX *pixs, l_int32 rval, l_int32 gval, l_int32 bval, l_int32 factor, l_float32 rank) { l_int32 mapval; l_float32 rankrval, rankgval, rankbval; l_float32 rfract, gfract, bfract, maxfract; PROCNAME("pixGlobalNormNoSatRGB"); if (!pixs) return (PIX *)ERROR_PTR("pixs not defined", procName, NULL); if (pixGetDepth(pixs) != 32) return (PIX *)ERROR_PTR("pixs not 32 bpp", procName, NULL); if (factor < 1) return (PIX *)ERROR_PTR("sampling factor < 1", procName, NULL); if (rank < 0.0 || rank > 1.0) return (PIX *)ERROR_PTR("rank not in [0.0 ... 1.0]", procName, NULL); if (rval <= 0 || gval <= 0 || bval <= 0) return (PIX *)ERROR_PTR("invalid estim. color values", procName, NULL); /* The max value for each component may be larger than the * input estimated background value. In that case, mapping * for those pixels would saturate. To prevent saturation, * we compute the fraction for each component by which we * would oversaturate. Then take the max of these, and * reduce, uniformly over all components, the output intensity * by this value. Then no component will saturate. * In practice, if rank < 1.0, a fraction of pixels * may have a component saturate. By keeping rank close to 1.0, * that fraction can be made arbitrarily small. */ pixGetRankValueMaskedRGB(pixs, NULL, 0, 0, factor, rank, &rankrval, &rankgval, &rankbval); rfract = rankrval / (l_float32)rval; gfract = rankgval / (l_float32)gval; bfract = rankbval / (l_float32)bval; maxfract = L_MAX(rfract, gfract); maxfract = L_MAX(maxfract, bfract); #if DEBUG_GLOBAL fprintf(stderr, "rankrval = %7.2f, rankgval = %7.2f, rankbval = %7.2f\n", rankrval, rankgval, rankbval); fprintf(stderr, "rfract = %7.4f, gfract = %7.4f, bfract = %7.4f\n", rfract, gfract, bfract); #endif /* DEBUG_GLOBAL */ mapval = (l_int32)(255. / maxfract); pixd = pixGlobalNormRGB(pixd, pixs, rval, gval, bval, mapval); return pixd; } /*------------------------------------------------------------------* * Adaptive threshold spread normalization * *------------------------------------------------------------------*/ /*! * pixThresholdSpreadNorm() * * Input: pixs (8 bpp) * filtertype (L_SOBEL_EDGE or L_TWO_SIDED_EDGE); * edgethresh (threshold on magnitude of edge filter; typ 10-20) * smoothx, smoothy (half-width of convolution kernel applied to * spread threshold: use 0 for no smoothing) * gamma (gamma correction; typ. about 0.7) * minval (input value that gives 0 for output; typ. -25) * maxval (input value that gives 255 for output; typ. 255) * targetthresh (target threshold for normalization) * &pixth ( computed local threshold value) * &pixb ( thresholded normalized image) * &pixd ( normalized image) * Return: 0 if OK, 1 on error * * Notes: * (1) The basis of this approach is the use of seed spreading * on a (possibly) sparse set of estimates for the local threshold. * The resulting dense estimates are smoothed by convolution * and used to either threshold the input image or normalize it * with a local transformation that linearly maps the pixels so * that the local threshold estimate becomes constant over the * resulting image. This approach is one of several that * have been suggested (and implemented) by Ray Smith. * (2) You can use either the Sobel or TwoSided edge filters. * The results appear to be similar, using typical values * of edgethresh in the rang 10-20. * (3) To skip the trc enhancement, use gamma = 1.0, minval = 0 * and maxval = 255. * (4) For the normalized image pixd, each pixel is linearly mapped * in such a way that the local threshold is equal to targetthresh. * (5) The full width and height of the convolution kernel * are (2 * smoothx + 1) and (2 * smoothy + 1). * (6) This function can be used with the pixtiling utility if the * images are too large. See pixOtsuAdaptiveThreshold() for * an example of this. */ l_int32 pixThresholdSpreadNorm(PIX *pixs, l_int32 filtertype, l_int32 edgethresh, l_int32 smoothx, l_int32 smoothy, l_float32 gamma, l_int32 minval, l_int32 maxval, l_int32 targetthresh, PIX **ppixth, PIX **ppixb, PIX **ppixd) { l_int32 w, h, d; PIX *pixe, *pixet, *pixsd, *pixg1, *pixg2, *pixth; PROCNAME("pixThresholdSpreadNorm"); if (ppixth) *ppixth = NULL; if (ppixb) *ppixb = NULL; if (ppixd) *ppixd = NULL; if (!pixs) return ERROR_INT("pixs not defined", procName, 1); pixGetDimensions(pixs, &w, &h, &d); if (d != 8) return ERROR_INT("pixs not 8 bpp", procName, 1); if (!ppixth && !ppixb && !ppixd) return ERROR_INT("no output requested", procName, 1); if (filtertype != L_SOBEL_EDGE && filtertype != L_TWO_SIDED_EDGE) return ERROR_INT("invalid filter type", procName, 1); /* Get the thresholded edge pixels. These are the ones * that have values in pixs near the local optimal fg/bg threshold. */ if (filtertype == L_SOBEL_EDGE) pixe = pixSobelEdgeFilter(pixs, L_VERTICAL_EDGES); else /* L_TWO_SIDED_EDGE */ pixe = pixTwoSidedEdgeFilter(pixs, L_VERTICAL_EDGES); pixet = pixThresholdToBinary(pixe, edgethresh); pixInvert(pixet, pixet); /* Build a seed image whose only nonzero values are those * values of pixs corresponding to pixels in the fg of pixet. */ pixsd = pixCreateTemplate(pixs); pixCombineMasked(pixsd, pixs, pixet); /* Spread the seed and optionally smooth to reduce noise */ pixg1 = pixSeedspread(pixsd, 4); pixg2 = pixBlockconv(pixg1, smoothx, smoothy); /* Optionally do a gamma enhancement */ pixth = pixGammaTRC(NULL, pixg2, gamma, minval, maxval); /* Do the mapping and thresholding */ if (ppixd) { *ppixd = pixApplyVariableGrayMap(pixs, pixth, targetthresh); if (ppixb) *ppixb = pixThresholdToBinary(*ppixd, targetthresh); } else if (ppixb) *ppixb = pixVarThresholdToBinary(pixs, pixth); if (ppixth) *ppixth = pixth; else pixDestroy(&pixth); pixDestroy(&pixe); pixDestroy(&pixet); pixDestroy(&pixsd); pixDestroy(&pixg1); pixDestroy(&pixg2); return 0; } /*------------------------------------------------------------------* * Adaptive background normalization (flexible adaptaption) * *------------------------------------------------------------------*/ /*! * pixBackgroundNormFlex() * * Input: pixs (8 bpp) * sx, sy (desired tile dimensions; actual size may vary; use * values between 3 and 10) * smoothx, smoothy (half-width of convolution kernel applied to * threshold array: use values between 1 and 3) * delta (difference parameter in basin filling; use 0 * to skip) * Return: pixd (8 bpp, background-normalized), or null on error) * * Notes: * (1) This does adaptation flexibly to a quickly varying background. * For that reason, all input parameters should be small. * (2) sx and sy give the tile size; they should be in [5 - 7]. * (3) The full width and height of the convolution kernel * are (2 * smoothx + 1) and (2 * smoothy + 1). They * should be in [1 - 2]. * (4) Basin filling is used to fill the large fg regions. The * parameter @delta measures the height that the black * background is raised from the local minima. By raising * the background, it is possible to threshold the large * fg regions to foreground. If @delta is too large, * bg regions will be lifted, causing thickening of * the fg regions. Use 0 to skip. */ PIX * pixBackgroundNormFlex(PIX *pixs, l_int32 sx, l_int32 sy, l_int32 smoothx, l_int32 smoothy, l_int32 delta) { l_float32 scalex, scaley; PIX *pixt, *pixsd, *pixmin, *pixbg, *pixbgi, *pixd; PROCNAME("pixBackgroundNormFlex"); if (!pixs || pixGetDepth(pixs) != 8) return (PIX *)ERROR_PTR("pixs undefined or not 8 bpp", procName, NULL); if (sx < 3 || sy < 3) return (PIX *)ERROR_PTR("sx and/or sy less than 3", procName, NULL); if (sx > 10 || sy > 10) return (PIX *)ERROR_PTR("sx and/or sy exceed 10", procName, NULL); if (smoothx < 1 || smoothy < 1) return (PIX *)ERROR_PTR("smooth params less than 1", procName, NULL); if (smoothx > 3 || smoothy > 3) return (PIX *)ERROR_PTR("smooth params exceed 3", procName, NULL); /* Generate the bg estimate using smoothed average with subsampling */ scalex = 1. / (l_float32)sx; scaley = 1. / (l_float32)sy; pixt = pixScaleSmooth(pixs, scalex, scaley); /* Do basin filling on the bg estimate if requested */ if (delta <= 0) pixsd = pixClone(pixt); else { pixLocalExtrema(pixt, 0, 0, &pixmin, NULL); pixsd = pixSeedfillGrayBasin(pixmin, pixt, delta, 4); pixDestroy(&pixmin); } pixbg = pixExtendByReplication(pixsd, 1, 1); /* Map the bg to 200 */ pixbgi = pixGetInvBackgroundMap(pixbg, 200, smoothx, smoothy); pixd = pixApplyInvBackgroundGrayMap(pixs, pixbgi, sx, sy); pixDestroy(&pixt); pixDestroy(&pixsd); pixDestroy(&pixbg); pixDestroy(&pixbgi); return pixd; } /*------------------------------------------------------------------* * Adaptive contrast normalization * *------------------------------------------------------------------*/ /*! * pixContrastNorm() * * Input: pixd ( 8 bpp; null or equal to pixs) * pixs (8 bpp, not colormapped) * sx, sy (tile dimensions) * mindiff (minimum difference to accept as valid) * smoothx, smoothy (half-width of convolution kernel applied to * min and max arrays: use 0 for no smoothing) * Return: pixd always * * Notes: * (1) This function adaptively attempts to expand the contrast * to the full dynamic range in each tile. If the contrast in * a tile is smaller than @mindiff, it uses the min and max * pixel values from neighboring tiles. It also can use * convolution to smooth the min and max values from * neighboring tiles. After all that processing, it is * possible that the actual pixel values in the tile are outside * the computed [min ... max] range for local contrast * normalization. Such pixels are taken to be at either 0 * (if below the min) or 255 (if above the max). * (2) pixd can be equal to pixs (in-place operation) or * null (makes a new pixd). * (3) sx and sy give the tile size; they are typically at least 20. * (4) mindiff is used to eliminate results for tiles where it is * likely that either fg or bg is missing. A value around 50 * or more is reasonable. * (5) The full width and height of the convolution kernel * are (2 * smoothx + 1) and (2 * smoothy + 1). Some smoothing * is typically useful, and we limit the smoothing half-widths * to the range from 0 to 8. * (6) A linear TRC (gamma = 1.0) is applied to increase the contrast * in each tile. The result can subsequently be globally corrected, * by applying pixGammaTRC() with arbitrary values of gamma * and the 0 and 255 points of the mapping. */ PIX * pixContrastNorm(PIX *pixd, PIX *pixs, l_int32 sx, l_int32 sy, l_int32 mindiff, l_int32 smoothx, l_int32 smoothy) { PIX *pixmin, *pixmax; PROCNAME("pixContrastNorm"); if (!pixs || pixGetDepth(pixs) != 8) return (PIX *)ERROR_PTR("pixs undefined or not 8 bpp", procName, pixd); if (pixd && pixd != pixs) return (PIX *)ERROR_PTR("pixd not null or == pixs", procName, pixd); if (pixGetColormap(pixs)) return (PIX *)ERROR_PTR("pixs is colormapped", procName, pixd); if (sx < 5 || sy < 5) return (PIX *)ERROR_PTR("sx and/or sy less than 5", procName, pixd); if (smoothx < 0 || smoothy < 0) return (PIX *)ERROR_PTR("smooth params less than 0", procName, pixd); if (smoothx > 8 || smoothy > 8) return (PIX *)ERROR_PTR("smooth params exceed 8", procName, pixd); /* Get the min and max pixel values in each tile, and represent * each value as a pixel in pixmin and pixmax, respectively. */ pixMinMaxTiles(pixs, sx, sy, mindiff, smoothx, smoothy, &pixmin, &pixmax); /* For each tile, do a linear expansion of the dynamic range * of pixels so that the min value is mapped to 0 and the * max value is mapped to 255. */ pixd = pixLinearTRCTiled(pixd, pixs, sx, sy, pixmin, pixmax); pixDestroy(&pixmin); pixDestroy(&pixmax); return pixd; } /*! * pixMinMaxTiles() * * Input: pixs (8 bpp, not colormapped) * sx, sy (tile dimensions) * mindiff (minimum difference to accept as valid) * smoothx, smoothy (half-width of convolution kernel applied to * min and max arrays: use 0 for no smoothing) * &pixmin ( tiled minima) * &pixmax ( tiled maxima) * Return: 0 if OK, 1 on error * * Notes: * (1) This computes filtered and smoothed values for the min and * max pixel values in each tile of the image. * (2) See pixContrastNorm() for usage. */ l_int32 pixMinMaxTiles(PIX *pixs, l_int32 sx, l_int32 sy, l_int32 mindiff, l_int32 smoothx, l_int32 smoothy, PIX **ppixmin, PIX **ppixmax) { l_int32 w, h; PIX *pixmin1, *pixmax1, *pixmin2, *pixmax2; PROCNAME("pixMinMaxTiles"); if (!ppixmin || !ppixmax) return ERROR_INT("&pixmin or &pixmax undefined", procName, 1); *ppixmin = *ppixmax = 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 colormapped", procName, 1); if (sx < 5 || sy < 5) return ERROR_INT("sx and/or sy less than 3", procName, 1); if (smoothx < 0 || smoothy < 0) return ERROR_INT("smooth params less than 0", procName, 1); if (smoothx > 5 || smoothy > 5) return ERROR_INT("smooth params exceed 5", procName, 1); /* Get the min and max values in each tile */ pixmin1 = pixScaleGrayMinMax(pixs, sx, sy, L_CHOOSE_MIN); pixmax1 = pixScaleGrayMinMax(pixs, sx, sy, L_CHOOSE_MAX); pixmin2 = pixExtendByReplication(pixmin1, 1, 1); pixmax2 = pixExtendByReplication(pixmax1, 1, 1); pixDestroy(&pixmin1); pixDestroy(&pixmax1); /* Make sure no value is 0 */ pixAddConstantGray(pixmin2, 1); pixAddConstantGray(pixmax2, 1); /* Generate holes where the contrast is too small */ pixSetLowContrast(pixmin2, pixmax2, mindiff); /* Fill the holes (0 values) */ pixGetDimensions(pixmin2, &w, &h, NULL); pixFillMapHoles(pixmin2, w, h, L_FILL_BLACK); pixFillMapHoles(pixmax2, w, h, L_FILL_BLACK); /* Smooth if requested */ if (smoothx > 0 || smoothy > 0) { smoothx = L_MIN(smoothx, (w - 1) / 2); smoothy = L_MIN(smoothy, (h - 1) / 2); *ppixmin = pixBlockconv(pixmin2, smoothx, smoothy); *ppixmax = pixBlockconv(pixmax2, smoothx, smoothy); } else { *ppixmin = pixClone(pixmin2); *ppixmax = pixClone(pixmax2); } pixDestroy(&pixmin2); pixDestroy(&pixmax2); return 0; } /*! * pixSetLowContrast() * * Input: pixs1 (8 bpp) * pixs2 (8 bpp) * mindiff (minimum difference to accept as valid) * Return: 0 if OK; 1 if no pixel diffs are large enough, or on error * * Notes: * (1) This compares corresponding pixels in pixs1 and pixs2. * When they differ by less than @mindiff, set the pixel * values to 0 in each. Each pixel typically represents a tile * in a larger image, and a very small difference between * the min and max in the tile indicates that the min and max * values are not to be trusted. * (2) If contrast (pixel difference) detection is expected to fail, * caller should check return value. */ l_int32 pixSetLowContrast(PIX *pixs1, PIX *pixs2, l_int32 mindiff) { l_int32 i, j, w, h, d, wpl, val1, val2, found; l_uint32 *data1, *data2, *line1, *line2; PROCNAME("pixSetLowContrast"); if (!pixs1 || !pixs2) return ERROR_INT("pixs1 and pixs2 not both defined", procName, 1); if (pixSizesEqual(pixs1, pixs2) == 0) return ERROR_INT("pixs1 and pixs2 not equal size", procName, 1); pixGetDimensions(pixs1, &w, &h, &d); if (d != 8) return ERROR_INT("depth not 8 bpp", procName, 1); if (mindiff > 254) return 0; data1 = pixGetData(pixs1); data2 = pixGetData(pixs2); wpl = pixGetWpl(pixs1); found = 0; /* init to not finding any diffs >= mindiff */ for (i = 0; i < h; i++) { line1 = data1 + i * wpl; line2 = data2 + i * wpl; for (j = 0; j < w; j++) { val1 = GET_DATA_BYTE(line1, j); val2 = GET_DATA_BYTE(line2, j); if (L_ABS(val1 - val2) >= mindiff) { found = 1; break; } } if (found) break; } if (!found) { L_WARNING("no pixel pair diffs as large as mindiff", procName); pixClearAll(pixs1); pixClearAll(pixs2); return 1; } for (i = 0; i < h; i++) { line1 = data1 + i * wpl; line2 = data2 + i * wpl; for (j = 0; j < w; j++) { val1 = GET_DATA_BYTE(line1, j); val2 = GET_DATA_BYTE(line2, j); if (L_ABS(val1 - val2) < mindiff) { SET_DATA_BYTE(line1, j, 0); SET_DATA_BYTE(line2, j, 0); } } } return 0; } /*! * pixLinearTRCTiled() * * Input: pixd ( 8 bpp) * pixs (8 bpp, not colormapped) * sx, sy (tile dimensions) * pixmin (pix of min values in tiles) * pixmax (pix of max values in tiles) * Return: pixd always * * Notes: * (1) pixd can be equal to pixs (in-place operation) or * null (makes a new pixd). * (2) sx and sy give the tile size; they are typically at least 20. * (3) pixmin and pixmax are generated by pixMinMaxTiles() * (4) For each tile, this does a linear expansion of the dynamic * range so that the min value in the tile becomes 0 and the * max value in the tile becomes 255. * (5) The LUTs that do the mapping are generated as needed * and stored for reuse in an integer array within the ptr array iaa[]. */ PIX * pixLinearTRCTiled(PIX *pixd, PIX *pixs, l_int32 sx, l_int32 sy, PIX *pixmin, PIX *pixmax) { l_int32 i, j, k, m, w, h, wt, ht, wpl, wplt, xoff, yoff; l_int32 minval, maxval, val, sval; l_int32 *ia; l_int32 **iaa; l_uint32 *data, *datamin, *datamax, *line, *tline, *linemin, *linemax; PROCNAME("pixLinearTRCTiled"); if (!pixs || pixGetDepth(pixs) != 8) return (PIX *)ERROR_PTR("pixs undefined or not 8 bpp", procName, pixd); if (pixd && pixd != pixs) return (PIX *)ERROR_PTR("pixd not null or == pixs", procName, pixd); if (pixGetColormap(pixs)) return (PIX *)ERROR_PTR("pixs is colormapped", procName, pixd); if (!pixmin || !pixmax) return (PIX *)ERROR_PTR("pixmin & pixmax not defined", procName, pixd); if (sx < 5 || sy < 5) return (PIX *)ERROR_PTR("sx and/or sy less than 5", procName, pixd); pixd = pixCopy(pixd, pixs); iaa = (l_int32 **)CALLOC(256, sizeof(l_int32 *)); pixGetDimensions(pixd, &w, &h, NULL); data = pixGetData(pixd); wpl = pixGetWpl(pixd); datamin = pixGetData(pixmin); datamax = pixGetData(pixmax); wplt = pixGetWpl(pixmin); pixGetDimensions(pixmin, &wt, &ht, NULL); for (i = 0; i < ht; i++) { line = data + sy * i * wpl; linemin = datamin + i * wplt; linemax = datamax + i * wplt; yoff = sy * i; for (j = 0; j < wt; j++) { xoff = sx * j; minval = GET_DATA_BYTE(linemin, j); maxval = GET_DATA_BYTE(linemax, j); if (maxval == minval) { /* this is bad */ /* fprintf(stderr, "should't happen! i,j = %d,%d, minval = %d\n", i, j, minval); */ continue; } ia = iaaGetLinearTRC(iaa, maxval - minval); for (k = 0; k < sy && yoff + k < h; k++) { tline = line + k * wpl; for (m = 0; m < sx && xoff + m < w; m++) { val = GET_DATA_BYTE(tline, xoff + m); sval = val - minval; sval = L_MAX(0, sval); SET_DATA_BYTE(tline, xoff + m, ia[sval]); } } } } for (i = 0; i < 256; i++) if (iaa[i]) FREE(iaa[i]); FREE(iaa); return pixd; } /*! * iaaGetLinearTRC() * * Input: iaa (bare array of ptrs to l_int32) * diff (between min and max pixel values that are * to be mapped to 0 and 255) * Return: ia (LUT with input (val - minval) and output a * value between 0 and 255) */ static l_int32 * iaaGetLinearTRC(l_int32 **iaa, l_int32 diff) { l_int32 i; l_int32 *ia; l_float32 factor; PROCNAME("iaaGetLinearTRC"); if (!iaa) return (l_int32 *)ERROR_PTR("iaa not defined", procName, NULL); if (iaa[diff] != NULL) /* already have it */ return iaa[diff]; if ((ia = (l_int32 *)CALLOC(256, sizeof(l_int32))) == NULL) return (l_int32 *)ERROR_PTR("ia not made", procName, NULL); iaa[diff] = ia; if (diff == 0) { /* shouldn't happen */ for (i = 0; i < 256; i++) ia[i] = 128; } else { factor = 255. / (l_float32)diff; for (i = 0; i < diff + 1; i++) ia[i] = (l_int32)(factor * i + 0.5); for (i = diff + 1; i < 256; i++) ia[i] = 255; } return ia; }