convolution

convolution — convolve and correlate images

Stability Level

Stable, unless otherwise indicated

Functions

int vips_conv ()
int vips_convf ()
int vips_convi ()
int vips_conva ()
int vips_convsep ()
int vips_convasep ()
int vips_compass ()
int vips_gaussblur ()
int vips_sharpen ()
int vips_spcor ()
int vips_fastcor ()

Types and Values

Includes

#include <vips/vips.h>

Description

These operations convolve an image in some way, or are operations based on simple convolution, or are useful with convolution.

Functions

vips_conv ()

int
vips_conv (VipsImage *in,
           VipsImage **out,
           VipsImage *mask,
           ...);

Optional arguments:

  • precision : VipsPrecision, calculation accuracy

  • layers : gint, number of layers for approximation

  • cluster : gint, cluster lines closer than this distance

Convolution.

Perform a convolution of in with mask . Each output pixel is calculated as:

1
sigma[i]{pixel[i] * mask[i]} / scale + offset

where scale and offset are part of mask .

By default, precision is VIPS_PRECISION_FLOAT. The output image is always VIPS_FORMAT_FLOAT unless in is VIPS_FORMAT_DOUBLE, in which case out is also VIPS_FORMAT_DOUBLE.

If precision is VIPS_PRECISION_INTEGER, then elements of mask are converted to integers before convolution, using rint(), and the output image always has the same VipsBandFormat as the input image.

For VIPS_FORMAT_UCHAR images and VIPS_PRECISION_INTEGER precision , vips_conv() uses a fast vector path based on fixed-point arithmetic. This can produce slightly different results. Disable the vector path with --vips-novector or VIPS_NOVECTOR or vips_vector_set_enabled().

If precision is VIPS_PRECISION_APPROXIMATE then, like VIPS_PRECISION_INTEGER, mask is converted to int before convolution, and the output image always has the same VipsBandFormat as the input image.

Larger values for layers give more accurate results, but are slower. As layers approaches the mask radius, the accuracy will become close to exact convolution and the speed will drop to match. For many large masks, such as Gaussian, n_layers need be only 10% of this value and accuracy will still be good.

Smaller values of cluster will give more accurate results, but be slower and use more memory. 10% of the mask radius is a good rule of thumb.

See also: vips_convsep().

[method]

Parameters

in

input image

 

out

output image.

[out]

mask

convolve with this mask

 

...

NULL-terminated list of optional named arguments

 

Returns

0 on success, -1 on error


vips_convf ()

int
vips_convf (VipsImage *in,
            VipsImage **out,
            VipsImage *mask,
            ...);

Convolution. This is a low-level operation, see vips_conv() for something more convenient.

Perform a convolution of in with mask . Each output pixel is calculated as sigma[i]{pixel[i] * mask[i]} / scale + offset, where scale and offset are part of mask .

The convolution is performed with floating-point arithmetic. The output image is always VIPS_FORMAT_FLOAT unless in is VIPS_FORMAT_DOUBLE, in which case out is also VIPS_FORMAT_DOUBLE.

See also: vips_conv().

[method]

Parameters

in

input image

 

out

output image.

[out]

mask

convolve with this mask

 

...

NULL-terminated list of optional named arguments

 

Returns

0 on success, -1 on error


vips_convi ()

int
vips_convi (VipsImage *in,
            VipsImage **out,
            VipsImage *mask,
            ...);

Integer convolution. This is a low-level operation, see vips_conv() for something more convenient.

mask is converted to an integer mask with rint() of each element, rint of scale and rint of offset. Each output pixel is then calculated as

1
sigma[i]{pixel[i] * mask[i]} / scale + offset

The output image always has the same VipsBandFormat as the input image.

For VIPS_FORMAT_UCHAR images, vips_convi() uses a fast vector path based on half-float arithmetic. This can produce slightly different results. Disable the vector path with --vips-novector or VIPS_NOVECTOR or vips_vector_set_enabled().

See also: vips_conv().

[method]

Parameters

in

input image

 

out

output image.

[out]

mask

convolve with this mask

 

...

NULL-terminated list of optional named arguments

 

Returns

0 on success, -1 on error


vips_conva ()

int
vips_conva (VipsImage *in,
            VipsImage **out,
            VipsImage *mask,
            ...);

Optional arguments:

  • layers : gint, number of layers for approximation

  • cluster : gint, cluster lines closer than this distance

Perform an approximate integer convolution of in with mask . This is a low-level operation, see vips_conv() for something more convenient.

The output image always has the same VipsBandFormat as the input image. Elements of mask are converted to integers before convolution.

Larger values for layers give more accurate results, but are slower. As layers approaches the mask radius, the accuracy will become close to exact convolution and the speed will drop to match. For many large masks, such as Gaussian, layers need be only 10% of this value and accuracy will still be good.

Smaller values of cluster will give more accurate results, but be slower and use more memory. 10% of the mask radius is a good rule of thumb.

See also: vips_conv().

[method]

Parameters

in

input image

 

out

output image.

[out]

mask

convolution mask

 

...

NULL-terminated list of optional named arguments

 

Returns

0 on success, -1 on error


vips_convsep ()

int
vips_convsep (VipsImage *in,
              VipsImage **out,
              VipsImage *mask,
              ...);

Optional arguments:

  • precision : calculation accuracy

  • layers : number of layers for approximation

  • cluster : cluster lines closer than this distance

Perform a separable convolution of in with mask . See vips_conv() for a detailed description.

The mask must be 1xn or nx1 elements.

The image is convolved twice: once with mask and then again with mask rotated by 90 degrees. This is much faster for certain types of mask (gaussian blur, for example) than doing a full 2D convolution.

See also: vips_conv(), vips_gaussmat().

[method]

Parameters

in

input image

 

out

output image.

[out]

mask

convolution mask

 

...

NULL-terminated list of optional named arguments

 

Returns

0 on success, -1 on error


vips_convasep ()

int
vips_convasep (VipsImage *in,
               VipsImage **out,
               VipsImage *mask,
               ...);

Optional arguments:

  • layers : gint, number of layers for approximation

Approximate separable integer convolution. This is a low-level operation, see vips_convsep() for something more convenient.

The image is convolved twice: once with mask and then again with mask rotated by 90 degrees. mask must be 1xn or nx1 elements. Elements of mask are converted to integers before convolution.

Larger values for layers give more accurate results, but are slower. As layers approaches the mask radius, the accuracy will become close to exact convolution and the speed will drop to match. For many large masks, such as Gaussian, layers need be only 10% of this value and accuracy will still be good.

The output image always has the same VipsBandFormat as the input image.

See also: vips_convsep().

[method]

Parameters

in

input image

 

out

output image.

[out]

mask

convolve with this mask

 

...

NULL-terminated list of optional named arguments

 

Returns

0 on success, -1 on error


vips_compass ()

int
vips_compass (VipsImage *in,
              VipsImage **out,
              VipsImage *mask,
              ...);

Optional arguments:

  • times : gint, how many times to rotate and convolve

  • angle : VipsAngle45, rotate mask by this much between colvolutions

  • combine : VipsCombine, combine results like this

  • precision : VipsPrecision, precision for blur, default float

  • layers : gint, number of layers for approximation

  • cluster : gint, cluster lines closer than this distance

This convolves in with mask times times, rotating mask by angle each time. By default, it comvolves twice, rotating by 90 degrees, taking the maximum result.

See also: vips_conv().

[method]

Parameters

in

input image

 

out

output image.

[out]

mask

convolve with this mask

 

...

NULL-terminated list of optional named arguments

 

Returns

0 on success, -1 on error.


vips_gaussblur ()

int
vips_gaussblur (VipsImage *in,
                VipsImage **out,
                double sigma,
                ...);

Optional arguments:

  • precision : VipsPrecision, precision for blur, default int

  • min_ampl : minimum amplitude, default 0.2

This operator runs vips_gaussmat() and vips_convsep() for you on an image. Set min_ampl smaller to generate a larger, more accurate mask. Set sigma larger to make the blur more blurry.

See also: vips_gaussmat(), vips_convsep().

[method]

Parameters

in

input image

 

out

output image.

[out]

sigma

how large a mask to use

 

...

NULL-terminated list of optional named arguments

 

Returns

0 on success, -1 on error.


vips_sharpen ()

int
vips_sharpen (VipsImage *in,
              VipsImage **out,
              ...);

Optional arguments:

  • sigma : sigma of gaussian

  • x1 : flat/jaggy threshold

  • y2 : maximum amount of brightening

  • y3 : maximum amount of darkening

  • m1 : slope for flat areas

  • m2 : slope for jaggy areas

Selectively sharpen the L channel of a LAB image. The input image is transformed to VIPS_INTERPRETATION_LABS.

The operation performs a gaussian blur of radius radius and subtracts from in to generate a high-frequency signal. This signal is passed through a lookup table formed from the five parameters and added back to in .

The lookup table is formed like this:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
.                     ^
.                  y2 |- - - - - -----------
.                     |         / 
.                     |        / slope m2
.                     |    .../    
.             -x1     | ...   |    
. -------------------...---------------------->
.             |   ... |      x1           
.             |... slope m1
.             /       |
.            / m2     |
.           /         |
.          /          |
.         /           |
.        /            |
. ______/ _ _ _ _ _ _ | -y3
.                     |

For screen output, we suggest the following settings (the defaults):

1
2
3
4
5
6
sigma == 0.5
x1 == 2
y2 == 10         (don't brighten by more than 10 L*)
y3 == 20         (can darken by up to 20 L*)
m1 == 0          (no sharpening in flat areas)
m2 == 3          (some sharpening in jaggy areas)

If you want more or less sharpening, we suggest you just change the m2 parameter.

The sigma parameter changes the width of the fringe and can be adjusted according to the output printing resolution. As an approximate guideline, use 0.5 for 4 pixels/mm (display resolution), 1.0 for 12 pixels/mm and 1.5 for 16 pixels/mm (300 dpi == 12 pixels/mm). These figures refer to the image raster, not the half-tone resolution.

See also: vips_conv().

[method]

Parameters

in

input image

 

out

output image.

[out]

...

NULL-terminated list of optional named arguments

 

Returns

0 on success, -1 on error.


vips_spcor ()

int
vips_spcor (VipsImage *in,
            VipsImage *ref,
            VipsImage **out,
            ...);

Calculate a correlation surface.

ref is placed at every position in in and the correlation coefficient calculated. The output image is always float.

The output image is the same size as the input. Extra input edge pixels are made by copying the existing edges outwards.

The correlation coefficient is calculated as:

1
2
3
4
sumij (ref(i,j)-mean(ref))(inkl(i,j)-mean(inkl))
c(k,l) = ------------------------------------------------
sqrt(sumij (ref(i,j)-mean(ref))^2) *
            sqrt(sumij (inkl(i,j)-mean(inkl))^2)

where inkl is the area of in centred at position (k,l).

from Niblack "An Introduction to Digital Image Processing", Prentice/Hall, pp 138.

If the number of bands differs, one of the images must have one band. In this case, an n-band image is formed from the one-band image by joining n copies of the one-band image together, and then the two n-band images are operated upon.

The output image is always float, unless either of the two inputs is double, in which case the output is also double.

See also: vips_fastcor().

[method]

Parameters

in

input image

 

ref

reference image

 

out

output image.

[out]

...

NULL-terminated list of optional named arguments

 

Returns

0 on success, -1 on error


vips_fastcor ()

int
vips_fastcor (VipsImage *in,
              VipsImage *ref,
              VipsImage **out,
              ...);

Calculate a fast correlation surface.

ref is placed at every position in in and the sum of squares of differences calculated.

The output image is the same size as the input. Extra input edge pixels are made by copying the existing edges outwards.

If the number of bands differs, one of the images must have one band. In this case, an n-band image is formed from the one-band image by joining n copies of the one-band image together, and then the two n-band images are operated upon.

The output type is uint if both inputs are integer, float if both are float or complex, and double if either is double or double complex. In other words, the output type is just large enough to hold the whole range of possible values.

See also: vips_spcor().

[method]

Parameters

in

input image

 

ref

reference image

 

out

output image.

[out]

...

NULL-terminated list of optional named arguments

 

Returns

0 on success, -1 on error

Types and Values

enum VipsCombine

How to combine values. See vips_compass(), for example.

Members

VIPS_COMBINE_MAX

take the maximum of the possible values

 

VIPS_COMBINE_SUM

sum all the values

 

VIPS_COMBINE_MIN

take the minimum value

 

VIPS_COMBINE_LAST