imopv.h File Reference

Vectorized image operations. More...

#include "generic.h"

Macros
Image convolution flags
#define	VL_PAD_BY_ZERO   (0x0 << 0)
	Pad with zeroes.
#define	VL_PAD_BY_CONTINUITY   (0x1 << 0)
	Pad by continuity.
#define	VL_PAD_MASK   (0x3)
	Padding field selector.
#define	VL_TRANSPOSE   (0x1 << 2)
	Transpose result.

Functions
Image convolution
void	vl_imconvcol_vf (float *dst, vl_size dst_stride, float const *src, vl_size src_width, vl_size src_height, vl_size src_stride, float const *filt, vl_index filt_begin, vl_index filt_end, int step, unsigned int flags)
void	vl_imconvcol_vd (double *dst, vl_size dst_stride, double const *src, vl_size src_width, vl_size src_height, vl_size src_stride, double const *filt, vl_index filt_begin, vl_index filt_end, int step, unsigned int flags)
	Convolve image along columns. More...
void	vl_imconvcoltri_f (float *dest, vl_size destStride, float const *image, vl_size imageWidth, vl_size imageHeight, vl_size imageStride, vl_size filterSize, vl_size step, int unsigned flags)
	Convolve an image along the columns with a triangular kernel. More...
void	vl_imconvcoltri_d (double *dest, vl_size destStride, double const *image, vl_size imageWidth, vl_size imageHeight, vl_size imageStride, vl_size filterSize, vl_size step, int unsigned flags)
	Convolve an image along the columns with a triangular kernel. More...
Integral image
void	vl_imintegral_f (float *integral, vl_size integralStride, float const *image, vl_size imageWidth, vl_size imageHeight, vl_size imageStride)
	Compute integral image. More...
void	vl_imintegral_d (double *integral, vl_size integralStride, double const *image, vl_size imageWidth, vl_size imageHeight, vl_size imageStride)
	Compute integral image. More...
void	vl_imintegral_i32 (vl_int32 *integral, vl_size integralStride, vl_int32 const *image, vl_size imageWidth, vl_size imageHeight, vl_size imageStride)
	Compute integral image. More...
void	vl_imintegral_ui32 (vl_uint32 *integral, vl_size integralStride, vl_uint32 const *image, vl_size imageWidth, vl_size imageHeight, vl_size imageStride)
	Compute integral image. More...
Distance transform
void	vl_image_distance_transform_d (double const *image, vl_size numColumns, vl_size numRows, vl_size columnStride, vl_size rowStride, double *distanceTransform, vl_uindex *indexes, double coeff, double offset)
	Compute the distance transform of an image. More...
void	vl_image_distance_transform_f (float const *image, vl_size numColumns, vl_size numRows, vl_size columnStride, vl_size rowStride, float *distanceTransform, vl_uindex *indexes, float coeff, float offset)
Image smoothing
void	vl_imsmooth_f (float *smoothed, vl_size smoothedStride, float const *image, vl_size width, vl_size height, vl_size stride, double sigmax, double sigmay)
	Smooth an image with a Gaussian filter. More...
void	vl_imsmooth_d (double *smoothed, vl_size smoothedStride, double const *image, vl_size width, vl_size height, vl_size stride, double sigmax, double sigmay)
	Smooth an image with a Gaussian filter. More...
Image gradients
void	vl_imgradient_polar_f (float *amplitudeGradient, float *angleGradient, vl_size gradWidthStride, vl_size gradHeightStride, float const *image, vl_size imageWidth, vl_size imageHeight, vl_size imageStride)
void	vl_imgradient_polar_d (double *amplitudeGradient, double *angleGradient, vl_size gradWidthStride, vl_size gradHeightStride, double const *image, vl_size imageWidth, vl_size imageHeight, vl_size imageStride)
	Compute gradient mangitudes and directions of an image. More...
void	vl_imgradient_f (float *xGradient, float *yGradient, vl_size gradWidthStride, vl_size gradHeightStride, float const *image, vl_size imageWidth, vl_size imageHeight, vl_size imageStride)
void	vl_imgradient_d (double *xGradient, double *yGradient, vl_size gradWidthStride, vl_size gradHeightStride, double const *image, vl_size imageWidth, vl_size imageHeight, vl_size imageStride)
void	vl_imgradient_polar_f_callback (float const *sourceImage, int sourceImageWidth, int sourceImageHeight, float *dstImage, int dstWidth, int dstHeight, int octave, int level, void *params)

Detailed Description

This module provides the following image operations:

• Separable convolution. The function vl_imconvcol_vf() can be used to compute separable convolutions.
• Convolution by a triangular kernel. The function vl_imconvcoltri_vf() is an optimized convolution routine for triangular kernels.
• Distance transform. vl_image_distance_transform_f() is a linear algorithm to compute the distance transform of an image.

Remarks

Some operations are optimized to exploit possible SIMD instructions. This requires image data to be properly aligned (typically to 16 bytes). Similalry, the image stride (the number of bytes to skip to move to the next image row), must be aligned.

Author

Andrea Vedaldi

Function Documentation

vl_image_distance_transform_d()

vl_image_distance_transform_d	(	double const *	image,
		vl_size	numColumns,
		vl_size	numRows,
		vl_size	columnStride,
		vl_size	rowStride,
		double *	distanceTransform,
		vl_uindex *	indexes,
		double	coeff,
		double	offset
	)

Parameters

image	image.
numColumns	number of columns of the image.
numRows	number of rows of the image.
columnStride	offset from one column to the next.
rowStride	offset from one row to the next.
distanceTransform	distance transform (out).
indexes	nearest neighbor indexes (in/out).
coeff	quadratic cost coefficient (non-negative).
offset	quadratic cost offset.

The function computes the distance transform along the first dimension of the image image. Let \( I(u,v) \) be image. Its distance transfrom \( D(u,v) \) is given by:

\[ u^*(u,v) = \min_{u'} I(u',v) + \mathtt{coeff} (u' - u - \mathtt{offset})^2, \quad D(u,v) = I(u^*(u,v),v). \]

Notice that coeff must be non negative.

The function fills in the buffer distanceTransform with \( D \). This buffer must have the same size as image.

If indexes is not NULL, it must be a matrix of the same size o the image. The function interprets the value of this matrix as indexes of the pixels, i.e \( \mathtt{indexes}(u,v) \) is the index of pixel \( (u,v) \). On output, the matrix indexes contains \( \mathtt{indexes}(u^*(u,v),v) \). This information can be used to determine for each pixel \( (u,v) \) its “nearest neighbor&rdquo.

Notice that by swapping numRows and numColumns and columnStride and rowStride, the function can be made to operate along the other image dimension. Specifically, to compute the distance transform along columns and rows, call the functinon twice:

for (i = 0 ; i < numColumns * numRows ; ++i) indexes[i] = i ;
vl_image_distance_transform_d(image,numColumns,numRows,1,numColumns,
                              distanceTransform,indexes,u_coeff,u_offset) ;
vl_image_distance_transform_d(distanceTransform,numRows,numColumns,numColumns,1,
                              distanceTransform,indexes,u_coeff,u_offset) ;

Algorithm

The function implements the algorithm described in: P. F. Felzenszwalb and D. P. Huttenlocher, Distance Transforms of Sampled Functions, Technical Report, Cornell University, 2004.

Since the algorithm operates along one dimension per time, consider the 1D version of the problem for simplicity:

\[ d(y) = \min_{x} g(y;x), \quad g(y;x) = f(x) + \alpha (y - x - \beta)^2, \quad x,y \in \{0,1,\dots,N-1\}. \]

Hence the distance transform \( d(y) \) is the lower envelope of the family of parabolas \( g(y;x) \) indexed by \( x \). Notice that all parabolas have the same curvature and that their centers are located at \( x + \beta, \) \( x=0,\dots,N-1 \). The algorithm considers one parabola per time, from left to right, and finds the interval for which the parabola belongs to the lower envelope (if any).

Initially, only the leftmost parabola \( g(y;0) \) has been considered, and its validity interval is \((-\infty, \infty) \). Then the second parabola \( g(y;1) \) is considered. As long as \( \alpha > 0 \), the two parabolas \( g(y;0),\ g(y;1) \) intersect at a unique point \( \bar y \). Then the first parabola belongs to the envelope in the interval \( (-\infty, \bar y] \) and the second one in the interval \( (\bar y, +\infty] \). When the third parabola \( g(y;2) \) is considered, the intersection point \( \hat y \) with the previously added parabola \( g(y;1) \) is found. Now two cases may arise:

• \( \hat y > \bar y \), in which case all three parabolas belong to the envelope in the intervals \( (-\infty,\bar y], (\bar y, \hat y], (\hat y, +\infty] \).
• \( \hat y \leq \bar y \), in which case the second parabola \( g(y;1) \) has no point beloning to the envelope, and it is removed. One then remains with the two parabolas \( g(y;0),\ g(y;2) \) and the algorithm is re-iterated.

The algorithm proceeds in this fashion. Every time a new parabola is considered, its intersection point with the previously added parabola on the left is computed, and that parabola is potentially removed. The cost of an iteration is 1 plus the number of deleted parabolas. Since there are \( N \) iterations and at most \( N \) parabolas to delete overall, the complexity is linear, i.e. \( O(N) \).

vl_image_distance_transform_f()

vl_image_distance_transform_f	(	float const *	image,
		vl_size	numColumns,
		vl_size	numRows,
		vl_size	columnStride,
		vl_size	rowStride,
		float *	distanceTransform,
		vl_uindex *	indexes,
		float	coeff,
		float	offset
	)

See also

vl_image_distance_transform_d

vl_imconvcol_vd()

vl_imconvcol_vd	(	double *	dst,
		vl_size	dst_stride,
		double const *	src,
		vl_size	src_width,
		vl_size	src_height,
		vl_size	src_stride,
		double const *	filt,
		vl_index	filt_begin,
		vl_index	filt_end,
		int	step,
		unsigned int	flags
	)

Parameters

dst	destination image.
dst_stride	width of the destination image including padding.
src	source image.
src_width	width of the source image.
src_height	height of the source image.
src_stride	width of the source image including padding.
filt	filter kernel.
filt_begin	coordinate of the first filter element.
filt_end	coordinate of the last filter element.
step	sub-sampling step.
flags	operation modes.

The function convolves the column of the image src by the filter filt and saves the result to the image dst. The size of dst must be equal to the size of src. Formally, this results in the calculation

\[ \mathrm{dst} [x,y] = \sum_{p=y-\mathrm{filt\_end}}^{y-\mathrm{filt\_begin}} \mathrm{src}[x,y] \mathrm{filt}[y - p - \mathrm{filt\_begin}] \]

The function subsamples the image along the columns according to the parameter step. Setting step to 1 (one) computes the elements \(\mathrm{dst}[x,y]\) for all pairs (x,0), (x,1), (x,2) and so on. Setting step two 2 (two) computes only (x,0), (x,2) and so on (in this case the height of the destination image is floor(src_height/step)+1).

Calling twice the function can be used to compute 2-D separable convolutions. Use the flag VL_TRANSPOSE to transpose the result (in this case dst has transposed dimension as well).

The function allows the support of the filter to be any range. Usually the support is filt_end = -filt_begin.

The convolution operation may pick up values outside the image boundary. To cope with this edge cases, the function either pads the image by zero (VL_PAD_BY_ZERO) or with the values at the boundary (VL_PAD_BY_CONTINUITY).

vl_imconvcol_vf()

vl_imconvcol_vf	(	float *	dst,
		vl_size	dst_stride,
		float const *	src,
		vl_size	src_width,
		vl_size	src_height,
		vl_size	src_stride,
		float const *	filt,
		vl_index	filt_begin,
		vl_index	filt_end,
		int	step,
		unsigned int	flags
	)

See also

vl_imconvcol_vd

vl_imconvcoltri_d()

vl_imconvcoltri_d	(	double *	dest,
		vl_size	destStride,
		double const *	image,
		vl_size	imageWidth,
		vl_size	imageHeight,
		vl_size	imageStride,
		vl_size	filterSize,
		vl_size	step,
		int unsigned	flags
	)

Parameters

dest	destination image.
destStride	destination image stride.
image	image to convolve.
imageWidth	width of the image.
imageHeight	height of the image.
imageStride	width of the image including padding.
filterSize	size of the triangular filter.
step	sub-sampling step.
flags	operation modes.

The function convolves the columns of the image image with the triangular kernel

\[ k(t) = \frac{1}{\Delta^2} \max\{ \Delta - |t|, 0 \}, \quad t \in \mathbb{Z} \]

The paramter \( \Delta \), equal to the function argument filterSize, controls the width of the kernel. Notice that the support of \( k(x) \) as a continuous function of \( x \) is the open interval \( (-\Delta,\Delta) \), which has length \( 2\Delta \). However, \( k(x) \) restricted to the ingeter domain \( x \in \mathcal{Z} \) has support \( \{ -\Delta + 1, \Delta +2, \dots, \Delta-1 \} \), which counts \( 2 \Delta - 1 \) elements only. In particular, the discrete kernel is symmetric about the origin for all values of \( \Delta \).

The normalization factor \( 1 / \Delta^2 \) guaratnees that the filter is normalized to one, i.e.:

\[ \sum_{t=-\infty}^{+\infty} k(t) = 1 \]

Algorithm

The function exploits the fact that convolution by a triangular kernel can be expressed as the repeated convolution by a rectangular kernel, and that the latter can be performed in time indepenedent on the fiter width by using an integral-image type trick. Overall, the algorithm complexity is independent on the parameter filterSize and linear in the nubmer of image pixels.

See also

vl_imconvcol_vd for details on the meaning of the other parameters.

vl_imconvcoltri_f()

vl_imconvcoltri_f	(	float *	dest,
		vl_size	destStride,
		float const *	image,
		vl_size	imageWidth,
		vl_size	imageHeight,
		vl_size	imageStride,
		vl_size	filterSize,
		vl_size	step,
		int unsigned	flags
	)

See also

vl_imconvcoltri_d()

vl_imgradient_polar_d()

vl_imgradient_polar_d	(	double *	amplitudeGradient,
		double *	angleGradient,
		vl_size	gradWidthStride,
		vl_size	gradHeightStride,
		double const *	image,
		vl_size	imageWidth,
		vl_size	imageHeight,
		vl_size	imageStride
	)

Parameters

amplitudeGradient	Pointer to amplitude gradient plane
angleGradient	Pointer to angle gradient plane
gradWidthStride	Width of the gradient plane including padding
gradHeightStride	Height of the gradient plane including padding
image	Pointer to the source image
imageWidth	Source image width
imageHeight	Source image height
imageStride	Width of the source image including padding.

This functions computes the amplitudes and angles of input image gradient.

Gradient is computed simple by gradient kernel \( (-1 ~ 1) \), \( (-1 ~ 1)^T \) for border pixels and with sobel filter kernel \( (-0.5 ~ 0 ~ 0.5) \), \( (-0.5 ~ 0 ~ 0.5)^T \) otherwise on the input image image yielding x-gradient \( dx \), stored in xGradient and y-gradient \( dy \), stored in yGradient, respectively.

The amplitude of the gradient, stored in plane amplitudeGradient, is then calculated as \( \sqrt(dx^2+dy^2) \) and the angle of the gradient, stored in angleGradient is \( atan(\frac{dy}{dx}) \) normalised into interval 0 and \( 2\pi \).

This function also allows to process only part of the input image defining the imageStride as original image width and width as width of the sub-image.

Also it allows to easily align the output data by definition of the gradWidthStride and gradHeightStride .

vl_imgradient_polar_f()

vl_imgradient_polar_f	(	float *	amplitudeGradient,
		float *	angleGradient,
		vl_size	gradWidthStride,
		vl_size	gradHeightStride,
		float const *	image,
		vl_size	imageWidth,
		vl_size	imageHeight,
		vl_size	imageStride
	)

See also

vl_imgradient_polar_d

vl_imintegral_d()

vl_imintegral_d	(	double *	integral,
		vl_size	integralStride,
		double const *	image,
		vl_size	imageWidth,
		vl_size	imageHeight,
		vl_size	imageStride
	)

Parameters

integral	integral image.
integralStride	integral image stride.
image	source image.
imageWidth	source image width.
imageHeight	source image height.
imageStride	source image stride.

Let \( I(x,y), (x,y) \in [0, W-1] \times [0, H-1] \). The function computes the integral image \( J(x,y) \) of \( I(x,g) \):

\[ J(x,y) = \sum_{x'=0}^{x} \sum_{y'=0}^{y} I(x',y') \]

The integral image \( J(x,y) \) can be used to compute quickly the integral of of \( I(x,y) \) in a rectangular region \( R = [x',x'']\times[y',y''] \):

\[ \sum_{(x,y)\in[x',x'']\times[y',y'']} I(x,y) = (J(x'',y'') - J(x'-1, y'')) - (J(x'',y'-1) - J(x'-1,y'-1)). \]

Note that the order of operations is important when the integral image has an unsigned data type (e.g. vl_uint32). The formula is easily derived as follows:

\begin{eqnarray*} \sum_{(x,y)\in R} I(x,y) &=& \sum_{x=x'}^{x''} \sum_{y=y'}^{y''} I(x,y)\\ &=& \sum_{x=0}^{x''} \sum_{y=y'}^{y''} I(x,y) - \sum_{x=0}^{x'-1} \sum_{y=y'}^{y''} I(x,y)\\ &=& \sum_{x=0}^{x''} \sum_{y=0}^{y''} I(x,y) - \sum_{x=0}^{x''} \sum_{y=0}^{y'-1} I(x,y) - \sum_{x=0}^{x'-1} \sum_{y=0}^{y''} I(x,y) + \sum_{x=0}^{x'-1} \sum_{y=0}^{y'-1} I(x,y)\\ &=& J(x'',y'') - J(x'-1,y'') - J(x'',y'-1) + J(x'-1,y'-1). \end{eqnarray*}

vl_imintegral_f()

vl_imintegral_f	(	float *	integral,
		vl_size	integralStride,
		float const *	image,
		vl_size	imageWidth,
		vl_size	imageHeight,
		vl_size	imageStride
	)

See also

vl_imintegral_d.

vl_imintegral_i32()

vl_imintegral_i32	(	vl_int32 *	integral,
		vl_size	integralStride,
		vl_int32 const *	image,
		vl_size	imageWidth,
		vl_size	imageHeight,
		vl_size	imageStride
	)

See also

vl_imintegral_d.

vl_imintegral_ui32()

vl_imintegral_ui32	(	vl_uint32 *	integral,
		vl_size	integralStride,
		vl_uint32 const *	image,
		vl_size	imageWidth,
		vl_size	imageHeight,
		vl_size	imageStride
	)

See also

vl_imintegral_d.

vl_imsmooth_d()

vl_imsmooth_d	(	double *	smoothed,
		vl_size	smoothedStride,
		double const *	image,
		vl_size	width,
		vl_size	height,
		vl_size	stride,
		double	sigmax,
		double	sigmay
	)

Parameters

smoothed
smoothedStride
image
width
height
stride
sigmax
sigmay

vl_imsmooth_f()

vl_imsmooth_f	(	float *	smoothed,
		vl_size	smoothedStride,
		float const *	image,
		vl_size	width,
		vl_size	height,
		vl_size	stride,
		double	sigmax,
		double	sigmay
	)

See also

vl_imsmooth_d