Файл src/convolve.c

#include <stdio.h>
#include <stdlib.h>
#include "allheaders.h"

Функции

PIXpixBlockconv (PIX *pix, l_int32 wc, l_int32 hc)
PIXpixBlockconvGray (PIX *pixs, PIX *pixacc, l_int32 wc, l_int32 hc)
PIXpixBlockconvAccum (PIX *pixs)
PIXpixBlockconvGrayUnnormalized (PIX *pixs, l_int32 wc, l_int32 hc)
PIXpixBlockconvTiled (PIX *pix, l_int32 wc, l_int32 hc, l_int32 nx, l_int32 ny)
PIXpixBlockconvGrayTile (PIX *pixs, PIX *pixacc, l_int32 wc, l_int32 hc)
PIXpixWindowedMean (PIX *pixs, l_int32 wc, l_int32 hc, l_int32 normflag)
PIXpixWindowedMeanSquare (PIX *pixs, l_int32 size)
DPIXpixMeanSquareAccum (PIX *pixs)
PIXpixBlockrank (PIX *pixs, PIX *pixacc, l_int32 wc, l_int32 hc, l_float32 rank)
PIXpixBlocksum (PIX *pixs, PIX *pixacc, l_int32 wc, l_int32 hc)
PIXpixWoodfillTransform (PIX *pixs, l_int32 halfsize, PIX *pixacc)
PIXpixConvolve (PIX *pixs, L_KERNEL *kel, l_int32 outdepth, l_int32 normflag)
PIXpixConvolveSep (PIX *pixs, L_KERNEL *kelx, L_KERNEL *kely, l_int32 outdepth, l_int32 normflag)
FPIXfpixConvolve (FPIX *fpixs, L_KERNEL *kel, l_int32 normflag)
FPIXfpixConvolveSep (FPIX *fpixs, L_KERNEL *kelx, L_KERNEL *kely, l_int32 normflag)

Функции

FPIX* fpixConvolve ( FPIX fpixs,
L_KERNEL kel,
l_int32  normflag 
)

fpixConvolve()

Input: fpixs (32 bit float array) kernel normflag (1 to normalize kernel to unit sum; 0 otherwise) Return: fpixd (32 bit float array)

Notes: (1) This gives a float convolution with an arbitrary kernel. (2) If normflag == 1, the result is normalized by scaling all kernel values for a unit sum. Do not normalize if the kernel has null sum, such as a DoG. (3) With the FPix, there are no issues about negative array or kernel values. The convolution is performed with single precision arithmetic. (4) This uses a mirrored border to avoid special casing on the boundaries.

FPIX* fpixConvolveSep ( FPIX fpixs,
L_KERNEL kelx,
L_KERNEL kely,
l_int32  normflag 
)

fpixConvolveSep()

Input: fpixs (32 bit float array) kelx (x-dependent kernel) kely (y-dependent kernel) normflag (1 to normalize kernel to unit sum; 0 otherwise) Return: fpixd (32 bit float array)

Notes: (1) This does a convolution with a separable kernel that is is a sequence of convolutions in x and y. The two one-dimensional kernel components must be input separately; the full kernel is the product of these components. The support for the full kernel is thus a rectangular region. (2) The normflag parameter is used as in fpixConvolve(). (3) This uses mirrored borders to avoid special casing on the boundaries.

PIX* pixBlockconv ( PIX pix,
l_int32  wc,
l_int32  hc 
)

pixBlockconv()

Input: pix (8 or 32 bpp; or 2, 4 or 8 bpp with colormap) wc, hc (half width/height of convolution kernel) Return: pixd, or null on error

Notes: (1) The full width and height of the convolution kernel are (2 * wc + 1) and (2 * hc + 1) (2) Returns a copy if both wc and hc are 0 (3) Require that w >= 2 * wc + 1 and h >= 2 * hc + 1, where (w,h) are the dimensions of pixs.

PIX* pixBlockconvAccum ( PIX pixs  ) 

pixBlockconvAccum()

Input: pixs (1, 8 or 32 bpp) Return: accum pix (32 bpp), or null on error.

Notes: (1) The general recursion relation is a(i,j) = v(i,j) + a(i-1, j) + a(i, j-1) - a(i-1, j-1) For the first line, this reduces to the special case a(i,j) = v(i,j) + a(i, j-1) For the first column, the special case is a(i,j) = v(i,j) + a(i-1, j)

PIX* pixBlockconvGray ( PIX pixs,
PIX pixacc,
l_int32  wc,
l_int32  hc 
)

pixBlockconvGray()

Input: pix (8 bpp) accum pix (32 bpp; can be null) wc, hc (half width/height of convolution kernel) Return: pix (8 bpp), or null on error

Notes: (1) If accum pix is null, make one and destroy it before returning; otherwise, just use the input accum pix. (2) The full width and height of the convolution kernel are (2 * wc + 1) and (2 * hc + 1). (3) Returns a copy if both wc and hc are 0. (4) Require that w >= 2 * wc + 1 and h >= 2 * hc + 1, where (w,h) are the dimensions of pixs.

PIX* pixBlockconvGrayTile ( PIX pixs,
PIX pixacc,
l_int32  wc,
l_int32  hc 
)

pixBlockconvGrayTile()

Input: pixs (8 bpp gray) pixacc (32 bpp accum pix) wc, hc (half width/height of convolution kernel) Return: pixd, or null on error

Notes: (1) The full width and height of the convolution kernel are (2 * wc + 1) and (2 * hc + 1) (2) Assumes that the input pixs is padded with (wc + 1) pixels on left and right, and with (hc + 1) pixels on top and bottom. The returned pix has these stripped off; they are only used for computation. (3) Returns a copy if both wc and hc are 0 (4) Require that w > 2 * wc + 1 and h > 2 * hc + 1, where (w,h) are the dimensions of pixs.

PIX* pixBlockconvGrayUnnormalized ( PIX pixs,
l_int32  wc,
l_int32  hc 
)

pixBlockconvGrayUnnormalized()

Input: pixs (8 bpp) wc, hc (half width/height of convolution kernel) Return: pix (32 bpp; containing the convolution without normalizing for the window size), or null on error

Notes: (1) The full width and height of the convolution kernel are (2 * wc + 1) and (2 * hc + 1). (2) Require that w >= 2 * wc + 1 and h >= 2 * hc + 1, where (w,h) are the dimensions of pixs. (3) Returns a copy if both wc and hc are 0. (3) Adds mirrored border to avoid treating the boundary pixels specially. Note that we add wc + 1 pixels to the left and wc to the right. The added width is 2 * wc + 1 pixels, and the particular choice simplifies the indexing in the loop. Likewise, add hc + 1 pixels to the top and hc to the bottom. (4) To get the normalized result, divide by the area of the convolution kernel: (2 * wc + 1) * (2 * hc + 1) Specifically, do this: pixc = pixBlockconvGrayUnnormalized(pixs, wc, hc); fract = 1. / ((2 * wc + 1) * (2 * hc + 1)); pixMultConstantGray(pixc, fract); pixd = pixGetRGBComponent(pixc, L_ALPHA_CHANNEL); (5) Unlike pixBlockconvGray(), this always computes the accumulation pix because its size is tied to wc and hc. (6) Compare this implementation with pixBlockconvGray(), where most of the code in blockconvLow() is special casing for efficiently handling the boundary. Here, the use of mirrored borders and destination indexing makes the implementation very simple.

PIX* pixBlockconvTiled ( PIX pix,
l_int32  wc,
l_int32  hc,
l_int32  nx,
l_int32  ny 
)

pixBlockconvTiled()

Input: pix (8 or 32 bpp; or 2, 4 or 8 bpp with colormap) wc, hc (half width/height of convolution kernel) nx, ny (subdivision into tiles) Return: pixd, or null on error

Notes: (1) The full width and height of the convolution kernel are (2 * wc + 1) and (2 * hc + 1) (2) Returns a copy if both wc and hc are 0 (3) Require that w >= 2 * wc + 1 and h >= 2 * hc + 1, where (w,h) are the dimensions of pixs. (4) For nx == ny == 1, this defaults to pixBlockconv(), which is typically about twice as fast, and gives nearly identical results as pixBlockconvGrayTile(). (5) If the tiles are too small, nx and/or ny are reduced a minimum amount so that the tiles are expanded to the smallest workable size in the problematic direction(s). (6) Why a tiled version? Three reasons: (a) Because the accumulator is a uint32, overflow can occur for an image with more than 16M pixels. (b) The accumulator array for 16M pixels is 64 MB; using tiles reduces the size of this array. (c) Each tile can be processed independently, in parallel, on a multicore processor.

PIX* pixBlockrank ( PIX pixs,
PIX pixacc,
l_int32  wc,
l_int32  hc,
l_float32  rank 
)

pixBlockrank()

Input: pixs (1 bpp) accum pix (<optional> 32 bpp) wc, hc (half width/height of block sum/rank kernel) rank (between 0.0 and 1.0; 0.5 is median filter) Return: pixd (1 bpp)

Notes: (1) The full width and height of the convolution kernel are (2 * wc + 1) and (2 * hc + 1) (2) This returns a pixd where each pixel is a 1 if the neighborhood (2 * wc + 1) x (2 * hc + 1)) pixels contains the rank fraction of 1 pixels. Otherwise, the returned pixel is 0. Note that the special case of rank = 0.0 is always satisfied, so the returned pixd has all pixels with value 1. (3) If accum pix is null, make one, use it, and destroy it before returning; otherwise, just use the input accum pix (4) If both wc and hc are 0, returns a copy unless rank == 0.0, in which case this returns an all-ones image. (5) Require that w >= 2 * wc + 1 and h >= 2 * hc + 1, where (w,h) are the dimensions of pixs.

PIX* pixBlocksum ( PIX pixs,
PIX pixacc,
l_int32  wc,
l_int32  hc 
)

pixBlocksum()

Input: pixs (1 bpp) accum pix (<optional> 32 bpp) wc, hc (half width/height of block sum/rank kernel) Return: pixd (8 bpp)

Notes: (1) If accum pix is null, make one and destroy it before returning; otherwise, just use the input accum pix (2) The full width and height of the convolution kernel are (2 * wc + 1) and (2 * hc + 1) (3) Use of wc = hc = 1, followed by pixInvert() on the 8 bpp result, gives a nice anti-aliased, and somewhat darkened, result on text. (4) Require that w >= 2 * wc + 1 and h >= 2 * hc + 1, where (w,h) are the dimensions of pixs. (5) Returns in each dest pixel the sum of all src pixels that are within a block of size of the kernel, centered on the dest pixel. This sum is the number of src ON pixels in the block at each location, normalized to 255 for a block containing all ON pixels. For pixels near the boundary, where the block is not entirely contained within the image, we then multiply by a second normalization factor that is greater than one, so that all results are normalized by the number of participating pixels within the block.

PIX* pixConvolve ( PIX pixs,
L_KERNEL kel,
l_int32  outdepth,
l_int32  normflag 
)

pixConvolve()

Input: pixs (8, 16, 32 bpp; no colormap) kernel outdepth (of pixd: 8, 16 or 32) normflag (1 to normalize kernel to unit sum; 0 otherwise) Return: pixd (8, 16 or 32 bpp)

Notes: (1) This gives a convolution with an arbitrary kernel. (2) The parameter determines the depth of the result. (3) If normflag == 1, the result is normalized by scaling all kernel values for a unit sum. Do not normalize if the kernel has null sum, such as a DoG. (4) If the kernel is normalized to unit sum, the output values can never exceed 255, so an output depth of 8 bpp is sufficient. If the kernel is not normalized, it may be necessary to use 16 or 32 bpp output to avoid overflow. (5) The kernel values can be positive or negative, but the result for the convolution can only be stored as a positive number. Consequently, if it goes negative, the choices are to clip to 0 or take the absolute value. We're choosing the former for now. Another possibility would be to output a second unsigned image for the negative values. (6) This uses a mirrored border to avoid special casing on the boundaries. (7) The function is slow, running at about 12 machine cycles for each pixel-op in the convolution. For example, with a 3 GHz cpu, a 1 Mpixel grayscale image, and a kernel with (sx * sy) = 25 elements, the convolution takes about 100 msec.

PIX* pixConvolveSep ( PIX pixs,
L_KERNEL kelx,
L_KERNEL kely,
l_int32  outdepth,
l_int32  normflag 
)

pixConvolveSep()

Input: pixs (8 bpp) kelx (x-dependent kernel) kely (y-dependent kernel) outdepth (of pixd: 8, 16 or 32) normflag (1 to normalize kernel to unit sum; 0 otherwise) Return: pixd (8, 16 or 32 bpp)

Notes: (1) This does a convolution with a separable kernel that is is a sequence of convolutions in x and y. The two one-dimensional kernel components must be input separately; the full kernel is the product of these components. The support for the full kernel is thus a rectangular region. (2) The and parameters are used as in pixConvolve(). (3) If the kernel is normalized to unit sum, the output values can never exceed 255, so an output depth of 8 bpp is sufficient. If the kernel is not normalized, it may be necessary to use 16 or 32 bpp output to avoid overflow. (4) The kernel values can be positive or negative, but the result for the convolution can only be stored as a positive number. Consequently, if it goes negative, the choices are to clip to 0 or take the absolute value. We're choosing the former for now. Another possibility would be to output a second unsigned image for the negative values. (5) This uses mirrored borders to avoid special casing on the boundaries.

DPIX* pixMeanSquareAccum ( PIX pixs  ) 

pixMeanSquareAccum()

Input: pixs (8 bpp grayscale) Return: dpix (64 bit array), or null on error

Notes: (1) Similar to pixBlockconvAccum(), this computes the sum of the squares of the pixel values in such a way that the value at (i,j) is the sum of all squares in the rectangle from the origin to (i,j). (2) The general recursion relation (v are squared pixel values) is a(i,j) = v(i,j) + a(i-1, j) + a(i, j-1) - a(i-1, j-1) For the first line, this reduces to the special case a(i,j) = v(i,j) + a(i, j-1) For the first column, the special case is a(i,j) = v(i,j) + a(i-1, j)

PIX* pixWindowedMean ( PIX pixs,
l_int32  wc,
l_int32  hc,
l_int32  normflag 
)

pixWindowedMean()

Input: pixs (8 or 32 bpp grayscale) wc, hc (half width/height of convolution kernel) normflag (1 for normalization to get average in window; 0 for the sum in the window (un-normalized)) Return: pixd (8 or 32 bpp, average over kernel window)

Notes: (1) The input and output depths are the same. (2) A set of border pixels of width (wc + 1) on left and right, and of height (hc + 1) on top and bottom, is included in pixs. The output pixd (after convolution) has this border removed. (3) Typically, == 1. However, if you want the sum within the window, rather than a normalized convolution, use == 0. (4) This builds a block accumulator pix, uses it here, and destroys it.

PIX* pixWindowedMeanSquare ( PIX pixs,
l_int32  size 
)

pixWindowedMeanSquare()

Input: pixs (8 bpp grayscale) size (halfwidth of convolution kernel) Return: pixd (32 bpp, average over window of size (2 * size + 1))

Notes: (1) A set of border pixels of width (size + 1) is included in pixs. The output pixd (after convolution) has this border removed. (2) The advantage is that we are unaffected by the boundary, and it is not necessary to treat pixels within of the border differently. This is because processing for pixd only takes place for pixels in pixs for which the kernel is entirely contained in pixs. (3) Why do we have an added border of width ( + 1), when we only need pixels to satisfy this condition? Answer: the accumulators are asymmetric, requiring an extra row and column of pixels at top and left to work accurately.

PIX* pixWoodfillTransform ( PIX pixs,
l_int32  halfsize,
PIX pixacc 
)

pixWoodfillTransform()

Input: pixs (8 bpp) halfsize (of square over which neighbors are averaged) accum pix (<optional> 32 bpp) Return: pixd (1 bpp)

Notes: (1) The Woodfill transform compares each pixel against the average of its neighbors (in a square of odd dimension centered on the pixel). If the pixel is greater than the average of its neighbors, the output pixel value is 1; otherwise it is 0. (2) This can be used as an encoding for an image that is fairly robust against slow illumination changes, with applications in image comparison and mosaicing. (3) The size of the convolution kernel is (2 * halfsize + 1) on a side. The halfsize parameter must be >= 1. (4) If accum pix is null, make one, use it, and destroy it before returning; otherwise, just use the input accum pix


Документация по Leptonica. Последние изменения: Fri Aug 7 20:31:35 2009. Создано системой  doxygen 1.5.9