![]() |
Leptonica
1.82.0
Image processing and image analysis suite
|
#include <string.h>
#include "allheaders.h"
Go to the source code of this file.
Functions | |
static void | scaleToGray2Low (l_uint32 *datad, l_int32 wd, l_int32 hd, l_int32 wpld, l_uint32 *datas, l_int32 wpls, l_uint32 *sumtab, l_uint8 *valtab) |
static l_uint32 * | makeSumTabSG2 (void) |
static l_uint8 * | makeValTabSG2 (void) |
static void | scaleToGray3Low (l_uint32 *datad, l_int32 wd, l_int32 hd, l_int32 wpld, l_uint32 *datas, l_int32 wpls, l_uint32 *sumtab, l_uint8 *valtab) |
static l_uint32 * | makeSumTabSG3 (void) |
static l_uint8 * | makeValTabSG3 (void) |
static void | scaleToGray4Low (l_uint32 *datad, l_int32 wd, l_int32 hd, l_int32 wpld, l_uint32 *datas, l_int32 wpls, l_uint32 *sumtab, l_uint8 *valtab) |
static l_uint32 * | makeSumTabSG4 (void) |
static l_uint8 * | makeValTabSG4 (void) |
static void | scaleToGray6Low (l_uint32 *datad, l_int32 wd, l_int32 hd, l_int32 wpld, l_uint32 *datas, l_int32 wpls, l_int32 *tab8, l_uint8 *valtab) |
static l_uint8 * | makeValTabSG6 (void) |
static void | scaleToGray8Low (l_uint32 *datad, l_int32 wd, l_int32 hd, l_int32 wpld, l_uint32 *datas, l_int32 wpls, l_int32 *tab8, l_uint8 *valtab) |
static l_uint8 * | makeValTabSG8 (void) |
static void | scaleToGray16Low (l_uint32 *datad, l_int32 wd, l_int32 hd, l_int32 wpld, l_uint32 *datas, l_int32 wpls, l_int32 *tab8) |
static l_int32 | scaleMipmapLow (l_uint32 *datad, l_int32 wd, l_int32 hd, l_int32 wpld, l_uint32 *datas1, l_int32 wpls1, l_uint32 *datas2, l_int32 wpls2, l_float32 red) |
PIX * | pixScaleToGray (PIX *pixs, l_float32 scalefactor) |
PIX * | pixScaleToGrayFast (PIX *pixs, l_float32 scalefactor) |
PIX * | pixScaleToGray2 (PIX *pixs) |
PIX * | pixScaleToGray3 (PIX *pixs) |
PIX * | pixScaleToGray4 (PIX *pixs) |
PIX * | pixScaleToGray6 (PIX *pixs) |
PIX * | pixScaleToGray8 (PIX *pixs) |
PIX * | pixScaleToGray16 (PIX *pixs) |
PIX * | pixScaleToGrayMipmap (PIX *pixs, l_float32 scalefactor) |
PIX * | pixScaleMipmap (PIX *pixs1, PIX *pixs2, l_float32 scale) |
PIX * | pixExpandReplicate (PIX *pixs, l_int32 factor) |
PIX * | pixScaleGrayMinMax (PIX *pixs, l_int32 xfact, l_int32 yfact, l_int32 type) |
PIX * | pixScaleGrayMinMax2 (PIX *pixs, l_int32 type) |
PIX * | pixScaleGrayRankCascade (PIX *pixs, l_int32 level1, l_int32 level2, l_int32 level3, l_int32 level4) |
PIX * | pixScaleGrayRank2 (PIX *pixs, l_int32 rank) |
l_ok | pixScaleAndTransferAlpha (PIX *pixd, PIX *pixs, l_float32 scalex, l_float32 scaley) |
PIX * | pixScaleWithAlpha (PIX *pixs, l_float32 scalex, l_float32 scaley, PIX *pixg, l_float32 fract) |
Variables | |
l_float32 | AlphaMaskBorderVals [2] |
Scale-to-gray (1 bpp --> 8 bpp; arbitrary downscaling) PIX *pixScaleToGray() PIX *pixScaleToGrayFast()
Scale-to-gray (1 bpp --> 8 bpp; integer downscaling) PIX *pixScaleToGray2() PIX *pixScaleToGray3() PIX *pixScaleToGray4() PIX *pixScaleToGray6() PIX *pixScaleToGray8() PIX *pixScaleToGray16()
Scale-to-gray by mipmap(1 bpp --> 8 bpp, arbitrary reduction) PIX *pixScaleToGrayMipmap()
Grayscale scaling using mipmap PIX *pixScaleMipmap()
Replicated (integer) expansion (all depths) PIX *pixExpandReplicate()
Grayscale downscaling using min and max PIX *pixScaleGrayMinMax() PIX *pixScaleGrayMinMax2()
Grayscale downscaling using rank value PIX *pixScaleGrayRankCascade() PIX *pixScaleGrayRank2()
Helper function for transferring alpha with scaling l_int32 pixScaleAndTransferAlpha()
RGB scaling including alpha (blend) component PIX *pixScaleWithAlpha()
Low-level static functions:
Scale-to-gray 2x static void scaleToGray2Low() static l_uint32 *makeSumTabSG2() static l_uint8 *makeValTabSG2()
Scale-to-gray 3x static void scaleToGray3Low() static l_uint32 *makeSumTabSG3() static l_uint8 *makeValTabSG3()
Scale-to-gray 4x static void scaleToGray4Low() static l_uint32 *makeSumTabSG4() static l_uint8 *makeValTabSG4()
Scale-to-gray 6x static void scaleToGray6Low() static l_uint8 *makeValTabSG6()
Scale-to-gray 8x static void scaleToGray8Low() static l_uint8 *makeValTabSG8()
Scale-to-gray 16x static void scaleToGray16Low()
Grayscale mipmap static l_int32 scaleMipmapLow()
Definition in file scale2.c.
|
static |
Notes: (1) Returns a table of 256 l_uint32s, giving the four output 8-bit grayscale sums corresponding to 8 input bits of a binary image, for a 2x scale-to-gray op. The sums from two adjacent scanlines are then added and transformed to output four 8 bpp pixel values, using makeValTabSG2().
|
static |
Notes: (1) Returns a table of 64 l_uint32s, giving the two output 8-bit grayscale sums corresponding to 6 input bits of a binary image, for a 3x scale-to-gray op. In practice, this would be used three times (on adjacent scanlines), and the sums would be added and then transformed to output 8 bpp pixel values, using makeValTabSG3().
|
static |
Notes: (1) Returns a table of 256 l_uint32s, giving the two output 8-bit grayscale sums corresponding to 8 input bits of a binary image, for a 4x scale-to-gray op. The sums from four adjacent scanlines are then added and transformed to output 8 bpp pixel values, using makeValTabSG4().
|
static |
Notes: (1) Returns an 8 bit value for the sum of ON pixels in a 2x2 square, according to val = 255 - (255 * sum)/4 where sum is in set {0,1,2,3,4}
|
static |
Notes: (1) Returns an 8 bit value for the sum of ON pixels in a 3x3 square, according to val = 255 - (255 * sum)/9 where sum is in [0,...,9]
|
static |
Notes: (1) Returns an 8 bit value for the sum of ON pixels in a 4x4 square, according to val = 255 - (255 * sum)/16 where sum is in [0,...,16]
|
static |
Notes: (1) Returns an 8 bit value for the sum of ON pixels in a 6x6 square, according to val = 255 - (255 * sum)/36 where sum is in [0,...,36]
|
static |
Notes: (1) Returns an 8 bit value for the sum of ON pixels in an 8x8 square, according to val = 255 - (255 * sum)/64 where sum is in [0,...,64]
[in] | pixs | 1, 2, 4, 8, 16, 32 bpp |
[in] | factor | integer scale factor for replicative expansion |
Definition at line 872 of file scale2.c.
References pixCopy(), pixCopyColormap(), pixCreate(), pixExpandBinaryReplicate(), and pixGetDimensions().
Referenced by displayHSVColorRange(), fpixaDisplayQuadtree(), and pixCompareWithTranslation().
[in] | pixd | 32 bpp, scaled image |
[in] | pixs | 32 bpp, original unscaled image |
[in] | scalex | must be > 0.0 |
[in] | scaley | must be > 0.0 |
Notes: (1) This scales the alpha component of pixs and inserts into pixd.
[in] | pixs | 8 bpp, not cmapped |
[in] | xfact | x downscaling factor; integer |
[in] | yfact | y downscaling factor; integer |
[in] | type | L_CHOOSE_MIN, L_CHOOSE_MAX, L_CHOOSE_MAXDIFF |
Notes: (1) The downscaled pixels in pixd are the min, max or (max - min) of the corresponding set of xfact * yfact pixels in pixs. (2) Using L_CHOOSE_MIN is equivalent to a grayscale erosion, using a brick Sel of size (xfact * yfact), followed by subsampling within each (xfact * yfact) cell. Using L_CHOOSE_MAX is equivalent to the corresponding dilation. (3) Using L_CHOOSE_MAXDIFF finds the difference between max and min values in each cell. (4) For the special case of downscaling by 2x in both directions, pixScaleGrayMinMax2() is about 2x more efficient.
[in] | pixs | 8 bpp, not cmapped |
[in] | type | L_CHOOSE_MIN, L_CHOOSE_MAX, L_CHOOSE_MAXDIFF |
Notes: (1) Special version for 2x reduction. The downscaled pixels in pixd are the min, max or (max - min) of the corresponding set of 4 pixels in pixs. (2) The max and min operations are a special case (for levels 1 and 4) of grayscale analog to the binary rank scaling operation pixReduceRankBinary2(). Note, however, that because of the photometric definition that higher gray values are lighter, the erosion-like L_CHOOSE_MIN will darken the resulting image, corresponding to a threshold level 1 in the binary case. Likewise, L_CHOOSE_MAX will lighten the pixd, corresponding to a threshold level of 4. (3) To choose any of the four rank levels in a 2x grayscale reduction, use pixScaleGrayRank2(). (4) This runs at about 70 MPix/sec/GHz of source data for erosion and dilation.
[in] | pixs | 8 bpp, no cmap |
[in] | rank | 1 (darkest), 2, 3, 4 (lightest) |
Notes: (1) Rank 2x reduction. If rank == 1(4), the downscaled pixels in pixd are the min(max) of the corresponding set of 4 pixels in pixs. Values 2 and 3 are intermediate. (2) This is the grayscale analog to the binary rank scaling operation pixReduceRankBinary2(). Here, because of the photometric definition that higher gray values are lighter, rank 1 gives the darkest pixel, whereas rank 4 gives the lightest pixel. This is opposite to the binary rank operation. (3) For rank = 1 and 4, this calls pixScaleGrayMinMax2(), which runs at about 70 MPix/sec/GHz of source data. For rank 2 and 3, this runs 3x slower, at about 25 MPix/sec/GHz.
PIX* pixScaleGrayRankCascade | ( | PIX * | pixs, |
l_int32 | level1, | ||
l_int32 | level2, | ||
l_int32 | level3, | ||
l_int32 | level4 | ||
) |
[in] | pixs | 8 bpp, not cmapped |
[in] | level1,level2 | ... |
[in] | level3,level4 | rank thresholds, in set {0, 1, 2, 3, 4} |
Notes: (1) This performs up to four cascaded 2x rank reductions. (2) Use level = 0 to truncate the cascade.
[in] | pixs1 | high res 8 bpp, no cmap |
[in] | pixs2 | low res – 2x reduced – 8 bpp, no cmap |
[in] | scale | reduction with respect to high res image, > 0.5 |
Notes: (1) See notes in pixScaleToGrayMipmap(). (2) This function suffers from aliasing effects that are easily seen in document images.
[in] | pixs | 1 bpp |
[in] | scalefactor | reduction: must be > 0.0 and < 1.0 |
Notes:
For faster scaling in the range of scalefactors from 0.0625 to 0.5, with very little difference in quality, use pixScaleToGrayFast().
Binary images have sharp edges, so they intrinsically have very high frequency content. To avoid aliasing, they must be low-pass filtered, which tends to blur the edges. How can we keep relatively crisp edges without aliasing? The trick is to do binary upscaling followed by a power-of-2 scaleToGray. For large reductions, where you don't end up with much detail, some corners can be cut.
The intent here is to get high quality reduced grayscale images with relatively little computation. We do binary pre-scaling followed by scaleToGrayN() for best results, esp. to avoid excess blur when the scale factor is near an inverse power of 2. Where a low-pass filter is required, we use simple convolution kernels: either the hat filter for linear interpolation or a flat filter for larger downscaling. Other choices, such as a perfect bandpass filter with infinite extent (the sinc) or various approximations to it (e.g., lanczos), are unnecessarily expensive.
The choices made are as follows: (1) Do binary upscaling before scaleToGrayN() for scalefactors > 1/8 (2) Do binary downscaling before scaleToGray8() for scalefactors between 1/16 and 1/8. (3) Use scaleToGray16() before grayscale downscaling for scalefactors less than 1/16 Another reasonable choice would be to start binary downscaling for scalefactors below 1/4, rather than below 1/8 as we do here.
The general scaling rules, not all of which are used here, go as follows: (1) For grayscale upscaling, use pixScaleGrayLI(). However, note that edges will be visibly blurred for scalefactors near (but above) 1.0. Replication will avoid edge blur, and should be considered for factors very near 1.0. (2) For grayscale downscaling with a scale factor larger than about 0.7, use pixScaleGrayLI(). For scalefactors near (but below) 1.0, you tread between Scylla and Charybdis. pixScaleGrayLI() again gives edge blurring, but pixScaleBySampling() gives visible aliasing. (3) For grayscale downscaling with a scale factor smaller than about 0.7, use pixScaleSmooth() (4) For binary input images, do as much scale to gray as possible using the special integer functions (2, 3, 4, 8 and 16). (5) It is better to upscale in binary, followed by scaleToGrayN() than to do scaleToGrayN() followed by an upscale using either LI or oversampling. (6) It may be better to downscale in binary, followed by scaleToGrayN() than to first use scaleToGrayN() followed by downscaling. For downscaling between 8x and 16x, this is a reasonable option. (7) For reductions greater than 16x, it's reasonable to use scaleToGray16() followed by further grayscale downscaling.
Definition at line 208 of file scale2.c.
Referenced by pixaDisplayTiledAndScaled().
[in] | pixs | 1 bpp |
[in] | pixs | 1 bpp |
[in] | pixs | 1 bpp |
Notes: (1) Speed is about 100 x 10^6 src-pixels/sec/GHz. Another way to express this is it processes 1 src pixel in about 10 cycles. (2) The width of pixd is truncated is truncated to a factor of 8.
[in] | pixs | 1 bpp |
Notes: (1) The width of pixd is truncated is truncated to a factor of 2.
[in] | pixs | 1 bpp |
Notes: (1) The width of pixd is truncated is truncated to a factor of 8.
[in] | pixs | 1 bpp |
[in] | pixs | 1 bpp |
[in] | scalefactor | reduction: must be > 0.0 and < 1.0 |
Notes: (1) See notes in pixScaleToGray() for the basic approach. (2) This function is considerably less expensive than pixScaleToGray() for scalefactor in the range (0.0625 ... 0.5), and the quality is nearly as good. (3) Unlike pixScaleToGray(), which does binary upscaling before downscaling for scale factors >= 0.0625, pixScaleToGrayFast() first downscales in binary for all scale factors < 0.5, and then does a 2x scale-to-gray as the final step. For scale factors < 0.0625, both do a 16x scale-to-gray, followed by further grayscale reduction.
[in] | pixs | 1 bpp |
[in] | scalefactor | reduction: must be > 0.0 and < 1.0 |
Notes:
This function is here mainly for pedagogical reasons. Mip-mapping is widely used in graphics for texture mapping, because the texture changes smoothly with scale. This is accomplished by constructing a multiresolution pyramid and, for each pixel, doing a linear interpolation between corresponding pixels in the two planes of the pyramid that bracket the desired resolution. The computation is very efficient, and is implemented in hardware in high-end graphics cards.
We can use mip-mapping for scale-to-gray by using two scale-to-gray reduced images (we don't need the entire pyramid) selected from the set {2x, 4x, ... 16x}, and interpolating. However, we get severe aliasing, probably because we are subsampling from the higher resolution image. The method is very fast, but the result is very poor. In fact, the results don't look any better than either subsampling off the higher-res grayscale image or oversampling on the lower-res image. Consequently, this method should NOT be used for generating reduced images, scale-to-gray or otherwise.
PIX* pixScaleWithAlpha | ( | PIX * | pixs, |
l_float32 | scalex, | ||
l_float32 | scaley, | ||
PIX * | pixg, | ||
l_float32 | fract | ||
) |
[in] | pixs | 32 bpp rgb or cmapped |
[in] | scalex | must be > 0.0 |
[in] | scaley | must be > 0.0 |
[in] | pixg | [optional] 8 bpp, can be null |
[in] | fract | between 0.0 and 1.0, with 0.0 fully transparent and 1.0 fully opaque |
Notes: (1) The alpha channel is transformed separately from pixs, and aligns with it, being fully transparent outside the boundary of the transformed pixs. For pixels that are fully transparent, a blending function like pixBlendWithGrayMask() will give zero weight to corresponding pixels in pixs. (2) Scaling is done with area mapping or linear interpolation, depending on the scale factors. Default sharpening is done. (3) If pixg is NULL, it is generated as an alpha layer that is partially opaque, using fract. Otherwise, it is cropped to pixs if required, and fract is ignored. The alpha channel in pixs is never used. (4) Colormaps are removed to 32 bpp. (5) The default setting for the border values in the alpha channel is 0 (transparent) for the outermost ring of pixels and (0.5 * fract * 255) for the second ring. When blended over a second image, this (a) shrinks the visible image to make a clean overlap edge with an image below, and (b) softens the edges by weakening the aliasing there. Use l_setAlphaMaskBorder() to change these values. (6) A subtle use of gamma correction is to remove gamma correction before scaling and restore it afterwards. This is done by sandwiching this function between a gamma/inverse-gamma photometric transform: pixt = pixGammaTRCWithAlpha(NULL, pixs, 1.0 / gamma, 0, 255); pixd = pixScaleWithAlpha(pixt, scalex, scaley, NULL, fract); pixGammaTRCWithAlpha(pixd, pixd, gamma, 0, 255); pixDestroy(&pixt); This has the side-effect of producing artifacts in the very dark regions.
Definition at line 1443 of file scale2.c.
References pixGetDimensions().
|
static |
Notes: (1) See notes in scale.c for pixScaleToGrayMipmap(). This function is here for pedagogical reasons. It gives poor results on document images because of aliasing.
Definition at line 2272 of file scale2.c.
References GET_DATA_BYTE, and SET_DATA_BYTE.
|
static |
[in] | datad | dest data |
[in] | wd,hd | dest width, height |
[in] | wpld | dest words/line |
[in] | datas | src data |
[in] | wpls | src words/line |
[in] | tab8 | made from makePixelSumTab8() |
Notes: (1) The output is processed one dest byte at a time, corresponding to 16 rows consisting each of 2 src bytes in the input image. This uses one lookup table, tab8, which gives the sum of ON pixels in a byte. After summing for all ON pixels in the 32 src bytes, which is between 0 and 256, this is converted to an 8 bpp grayscale value between 0 for 255 or 256 bits ON and 255 for 0 bits ON.
Definition at line 2195 of file scale2.c.
References GET_DATA_BYTE, and SET_DATA_BYTE.
|
static |
[in] | datad | dest data |
[in] | wd,hd | dest width, height |
[in] | wpld | dest words/line |
[in] | datas | src data |
[in] | wpls | src words/line |
[in] | sumtab | made from makeSumTabSG2() |
[in] | valtab | made from makeValTabSG2() |
Notes: (1) The output is processed in sets of 4 output bytes on a row, corresponding to 4 2x2 bit-blocks in the input image. Two lookup tables are used. The first, sumtab, gets the sum of ON pixels in 4 sets of two adjacent bits, storing the result in 4 adjacent bytes. After sums from two rows have been added, the second table, valtab, converts from the sum of ON pixels in the 2x2 block to an 8 bpp grayscale value between 0 for 4 bits ON and 255 for 0 bits ON.
Definition at line 1545 of file scale2.c.
References GET_DATA_BYTE, and SET_DATA_BYTE.
|
static |
[in] | datad | dest data |
[in] | wd,hd | dest width, height |
[in] | wpld | dest words/line |
[in] | datas | src data |
[in] | wpls | src words/line |
[in] | sumtab | made from makeSumTabSG3() |
[in] | valtab | made from makeValTabSG3() |
Notes: (1) Each set of 8 3x3 bit-blocks in the source image, which consist of 72 pixels arranged 24 pixels wide by 3 scanlines, is converted to a row of 8 8-bit pixels in the dest image. These 72 pixels of the input image are runs of 24 pixels in three adjacent scanlines. Each run of 24 pixels is stored in the 24 LSbits of a 32-bit word. We use 2 LUTs. The first, sumtab, takes 6 of these bits and stores sum, taken 3 bits at a time, in two bytes. (See makeSumTabSG3). This is done for each of the 3 scanlines, and the results are added. We now have the sum of ON pixels in the first two 3x3 blocks in two bytes. The valtab LUT then converts these values (which go from 0 to 9) to grayscale values between between 255 and 0. (See makeValTabSG3). This process is repeated for each of the other 3 sets of 6x3 input pixels, giving 8 output pixels in total. (2) Note: because the input image is processed in groups of 24 x 3 pixels, the process clips the input height to (h - h % 3) and the input width to (w - w % 24).
Definition at line 1687 of file scale2.c.
References GET_DATA_BYTE, and SET_DATA_BYTE.
|
static |
[in] | datad | dest data |
[in] | wd,hd | dest width, height |
[in] | wpld | dest words/line |
[in] | datas | src data |
[in] | wpls | src words/line |
[in] | sumtab | made from makeSumTabSG4() |
[in] | valtab | made from makeValTabSG4() |
Notes: (1) The output is processed in sets of 2 output bytes on a row, corresponding to 2 4x4 bit-blocks in the input image. Two lookup tables are used. The first, sumtab, gets the sum of ON pixels in two sets of four adjacent bits, storing the result in 2 adjacent bytes. After sums from four rows have been added, the second table, valtab, converts from the sum of ON pixels in the 4x4 block to an 8 bpp grayscale value between 0 for 16 bits ON and 255 for 0 bits ON.
Definition at line 1836 of file scale2.c.
References GET_DATA_BYTE, and SET_DATA_BYTE.
|
static |
[in] | datad | dest data |
[in] | wd,hd | dest width, height |
[in] | wpld | dest words/line |
[in] | datas | src data |
[in] | wpls | src words/line |
[in] | tab8 | made from makePixelSumTab8() |
[in] | valtab | made from makeValTabSG6() |
Notes: (1) Each set of 4 6x6 bit-blocks in the source image, which consist of 144 pixels arranged 24 pixels wide by 6 scanlines, is converted to a row of 4 8-bit pixels in the dest image. These 144 pixels of the input image are runs of 24 pixels in six adjacent scanlines. Each run of 24 pixels is stored in the 24 LSbits of a 32-bit word. We use 2 LUTs. The first, tab8, takes 6 of these bits and stores sum in one byte. This is done for each of the 6 scanlines, and the results are added. We now have the sum of ON pixels in the first 6x6 block. The valtab LUT then converts these values (which go from 0 to 36) to grayscale values between between 255 and 0. (See makeValTabSG6). This process is repeated for each of the other 3 sets of 6x6 input pixels, giving 4 output pixels in total. (2) Note: because the input image is processed in groups of 24 x 6 pixels, the process clips the input height to (h - h % 6) and the input width to (w - w % 24).
Definition at line 1965 of file scale2.c.
References GET_DATA_BYTE, and SET_DATA_BYTE.
|
static |
[in] | datad | dest data |
[in] | wd,hd | dest width, height |
[in] | wpld | dest words/line |
[in] | datas | src data |
[in] | wpls | src words/line |
[in] | tab8 | made from makePixelSumTab8() |
[in] | valtab | made from makeValTabSG8() |
Notes: (1) The output is processed one dest byte at a time, corresponding to 8 rows of src bytes in the input image. Two lookup tables are used. The first, tab8, gets the sum of ON pixels in a byte. After sums from 8 rows have been added, the second table, valtab, converts from this value which is between 0 and 64 to an 8 bpp grayscale value between 0 and 255: 0 for all 64 bits ON and 255 for all 64 bits OFF.
Definition at line 2103 of file scale2.c.
References GET_DATA_BYTE, and SET_DATA_BYTE.