NAME
Image::Leptonica::Func::adaptmap
VERSION
version 0.03
adaptmap.c
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.FUNCTIONS
pixApplyInvBackgroundGrayMap
PIX * pixApplyInvBackgroundGrayMap ( PIX *pixs, PIX *pixm, l_int32 sx, l_int32 sy )
pixApplyInvBackgroundGrayMap()
Input: pixs (8 bpp grayscale; no colormap)
pixm (16 bpp, inverse background map)
sx (tile width in pixels)
sy (tile height in pixels)
Return: pixd (8 bpp), or null on errorpixApplyInvBackgroundRGBMap
PIX * pixApplyInvBackgroundRGBMap ( PIX *pixs, PIX *pixmr, PIX *pixmg, PIX *pixmb, l_int32 sx, l_int32 sy )
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 errorpixApplyVariableGrayMap
PIX * pixApplyVariableGrayMap ( PIX *pixs, PIX *pixg, l_int32 target )
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.pixBackgroundNorm
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 )
pixBackgroundNorm()
Input: pixs (8 bpp grayscale or 32 bpp rgb)
pixim (<optional> 1 bpp 'image' mask; can be null)
pixg (<optional> 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.pixBackgroundNormFlex
PIX * pixBackgroundNormFlex ( PIX *pixs, l_int32 sx, l_int32 sy, l_int32 smoothx, l_int32 smoothy, l_int32 delta )
pixBackgroundNormFlex()
Input: pixs (8 bpp grayscale; not colormapped)
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.pixBackgroundNormGrayArray
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 )
pixBackgroundNormGrayArray()
Input: pixs (8 bpp grayscale)
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)
&pixd (<return> 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.pixBackgroundNormGrayArrayMorph
l_int32 pixBackgroundNormGrayArrayMorph ( PIX *pixs, PIX *pixim, l_int32 reduction, l_int32 size, l_int32 bgval, PIX **ppixd )
pixBackgroundNormGrayArrayMorph()
Input: pixs (8 bpp grayscale)
pixim (<optional> 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 (<return> 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.pixBackgroundNormMorph
PIX * pixBackgroundNormMorph ( PIX *pixs, PIX *pixim, l_int32 reduction, l_int32 size, l_int32 bgval )
pixBackgroundNormMorph()
Input: pixs (8 bpp grayscale or 32 bpp rgb)
pixim (<optional> 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.pixBackgroundNormRGBArrays
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 )
pixBackgroundNormRGBArrays()
Input: pixs (32 bpp rgb)
pixim (<optional> 1 bpp 'image' mask; can be null)
pixg (<optional> 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 (<return> 16 bpp array of inverted R background value)
&pixg (<return> 16 bpp array of inverted G background value)
&pixb (<return> 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.pixBackgroundNormRGBArraysMorph
l_int32 pixBackgroundNormRGBArraysMorph ( PIX *pixs, PIX *pixim, l_int32 reduction, l_int32 size, l_int32 bgval, PIX **ppixr, PIX **ppixg, PIX **ppixb )
pixBackgroundNormRGBArraysMorph()
Input: pixs (32 bpp rgb)
pixim (<optional> 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 (<return> 16 bpp array of inverted R background value)
&pixg (<return> 16 bpp array of inverted G background value)
&pixb (<return> 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.pixBackgroundNormSimple
#ifndef LEPTONICA_ALLHEADERS_H #define LEPTONICA_ALLHEADERS_H
#define LIBLEPT_MAJOR_VERSION 1 #define LIBLEPT_MINOR_VERSION 70
#include "alltypes.h"
#ifndef NO_PROTOS
#ifdef __cplusplus extern "C" { #endif
LEPT_DLL extern PIX * pixBackgroundNormSimple ( PIX *pixs, PIX *pixim, PIX *pixg )
pixBackgroundNormSimple()
Input: pixs (8 bpp grayscale or 32 bpp rgb)
pixim (<optional> 1 bpp 'image' mask; can be null)
pixg (<optional> 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.pixContrastNorm
PIX * pixContrastNorm ( PIX *pixd, PIX *pixs, l_int32 sx, l_int32 sy, l_int32 mindiff, l_int32 smoothx, l_int32 smoothy )
pixContrastNorm()
Input: pixd (<optional> 8 bpp; null or equal to pixs)
pixs (8 bpp grayscale; 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.pixExtendByReplication
PIX * pixExtendByReplication ( PIX *pixs, l_int32 addw, l_int32 addh )
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.pixFillMapHoles
l_int32 pixFillMapHoles ( PIX *pix, l_int32 nx, l_int32 ny, l_int32 filltype )
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.pixGetBackgroundGrayMap
l_int32 pixGetBackgroundGrayMap ( PIX *pixs, PIX *pixim, l_int32 sx, l_int32 sy, l_int32 thresh, l_int32 mincount, PIX **ppixd )
pixGetBackgroundGrayMap()
Input: pixs (8 bpp grayscale; not cmapped)
pixim (<optional> 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 (<return> 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.pixGetBackgroundGrayMapMorph
l_int32 pixGetBackgroundGrayMapMorph ( PIX *pixs, PIX *pixim, l_int32 reduction, l_int32 size, PIX **ppixm )
pixGetBackgroundGrayMapMorph()
Input: pixs (8 bpp grayscale; not cmapped)
pixim (<optional> 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 (<return> grayscale map)
Return: 0 if OK, 1 on errorpixGetBackgroundRGBMap
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 )
pixGetBackgroundRGBMap()
Input: pixs (32 bpp rgb)
pixim (<optional> 1 bpp 'image' mask; can be null; it
should not have all foreground pixels)
pixg (<optional> 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 (<return> 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.pixGetBackgroundRGBMapMorph
l_int32 pixGetBackgroundRGBMapMorph ( PIX *pixs, PIX *pixim, l_int32 reduction, l_int32 size, PIX **ppixmr, PIX **ppixmg, PIX **ppixmb )
pixGetBackgroundRGBMapMorph()
Input: pixs (32 bpp rgb)
pixim (<optional> 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 (<return> red component map)
&pixmg (<return> green component map)
&pixmb (<return> blue component map)
Return: 0 if OK, 1 on errorpixGetInvBackgroundMap
PIX * pixGetInvBackgroundMap ( PIX *pixs, l_int32 bgval, l_int32 smoothx, l_int32 smoothy )
pixGetInvBackgroundMap()
Input: pixs (8 bpp grayscale; no colormap)
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.pixGlobalNormNoSatRGB
PIX * pixGlobalNormNoSatRGB ( PIX *pixd, PIX *pixs, l_int32 rval, l_int32 gval, l_int32 bval, l_int32 factor, l_float32 rank )
pixGlobalNormNoSatRGB()
Input: pixd (<optional> 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).pixGlobalNormRGB
PIX * pixGlobalNormRGB ( PIX *pixd, PIX *pixs, l_int32 rval, l_int32 gval, l_int32 bval, l_int32 mapval )
pixGlobalNormRGB()
Input: pixd (<optional> 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);pixLinearTRCTiled
PIX * pixLinearTRCTiled ( PIX *pixd, PIX *pixs, l_int32 sx, l_int32 sy, PIX *pixmin, PIX *pixmax )
pixLinearTRCTiled()
Input: pixd (<optional> 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[].pixMinMaxTiles
l_int32 pixMinMaxTiles ( PIX *pixs, l_int32 sx, l_int32 sy, l_int32 mindiff, l_int32 smoothx, l_int32 smoothy, PIX **ppixmin, PIX **ppixmax )
pixMinMaxTiles()
Input: pixs (8 bpp grayscale; 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 (<return> tiled minima)
&pixmax (<return> 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.pixSetLowContrast
l_int32 pixSetLowContrast ( PIX *pixs1, PIX *pixs2, l_int32 mindiff )
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.pixSmoothConnectedRegions
l_int32 pixSmoothConnectedRegions ( PIX *pixs, PIX *pixm, l_int32 factor )
pixSmoothConnectedRegions()
Input: pixs (8 bpp grayscale; no colormap)
pixm (<optional> 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.pixThresholdSpreadNorm
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 )
pixThresholdSpreadNorm()
Input: pixs (8 bpp grayscale; not colormapped)
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 (<optional return> computed local threshold value)
&pixb (<optional return> thresholded normalized image)
&pixd (<optional return> 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.AUTHOR
Zakariyya Mughal <zmughal@cpan.org>
COPYRIGHT AND LICENSE
This software is copyright (c) 2014 by Zakariyya Mughal.
This is free software; you can redistribute it and/or modify it under the same terms as the Perl 5 programming language system itself.