/ ciao4.17 / ahelp / aconvolve.html

Jump to: Description · Examples · Parameters · Changes in CIAO 4.17 · Changes in CIAO 4.17 · Bugs · See Also

AHELP for CIAO 4.17

aconvolve

Synopsis

Convolve an N-dimensional image with a kernel

Syntax

aconvolve  infile outfile kernelspec [writekernel] [kernelfile]
[writefft] [fftroot] [method] [edges] [const] [pad] [center]
[normkernel] [clobber] [verbose]

Description

`aconvolve' convolves an N-dimensional image with a kernel (specified by the parameter 'kernelspec'). The kernel can be specified in one of several flexible ways: as either a file (image), a function, from the library, or specified on the command line. Other than memory restrictions of the machine, there are no limits to the size or dimensionality of either the kernel or image (i.e. the kernel could be larger than the image). One of the most common applications of this tool is smoothing of an image by convolving with a gaussian.

The convolution of a data set with a kernel yields a measure of "how much" the data set looks like the kernel at every location in the data set. Another way to think about the convolution operation is a filtering of the dataset by the kernel.

Two forms of convolution are available:

spatial domain (sliding cell);
frequency domain (FFT).

Frequency-domain convolution is the standard technique: take two functions, Fourier transform them (using the "FFTW" routine from MIT), take their product in Fourier space, and then back transforming the product to the spatial domain to produce the convolved result. When using method="fft", any size image can be used but the algorithm is faster when the data is "padded" in size to take advantage of certain algorithmic optimisations; see the pad parameter.

In spatial-domain convolution `aconvolve' can treat edge-located pixels in one of 5 ways:

"wrap" the edges;
"mirror" them;
set them to a "constant";
"renorm"-alize them;
use the "nearest" neighbor value.

Examples

Example 1

aconvolve test_data.fits outfile.fits "lib:gaus(2,5,1,10,10)" method=fft

Convolve (smooth) the image in test_data.fits with a Gaussian that

has 2 dimensions;
is embedded in an array 5 sigma in size;
is normalized to 1;
has a sigma of 10 pixels along each axis.

The convolution is done by the method of FFTing the input arrays, multiplying, and taking the inverse. It will work with any size array but it is much faster if the size is a power of 2.

Example 2

aconvolve test_data.fits outfile.fits "file:conv.fits" edges=wrap pad=no

Convolve the image in test_data.fits with the image in the file conv.fits, wrapping the edges and storing the data in outfile.fits. The input data are not padded.

Example 3

aconvolve test_data.fits outfile.fits "lib:box(2,1,3,3):(1,1)" pad=yes
method=fft writefft=no

Convolve (smooth) the image test_data.fits with a box that is normalized to 1, has two dimensions, and is 3 units long in each direction. The center of the box is a (1,1). The convolution is done by the method of FFTing the input arrays, multiplying, and taking the inverse. The data are padded to the next power of 2**N before doing the convolution. It will work with any size array but it is much faster if the size is a power of 2.

Example 4

aconvolve test_data.fits outfile.fits
"txt:((1,1,1),(1,1,1),(1,1,1)):(1,1)" edges=constant const=0 pad=yes

Same as Example 3. The "box" is implemented from the command line as a text array. The edges are padded with a constant = 0.

Example 5

aconvolve img.fits shifted_img.fits "txt:((1)):(1,1)" meth=slide
edge=wrap

This shifts both the X and Y by +1.

It is also possible to change the offset (kernel origin) of each axis individually, e.g. to shift X by +512 but not change Y (shift=0):

txt:((1)):(512,0)

Parameters

name	type	ftype	def	min	max	reqd	autoname
infile	file	input				yes
outfile	file	output				yes	yes
kernelspec	string					yes
writekernel	boolean		no
kernelfile	file	output	./.				yes
writefft	boolean		no
fftroot	file	output	./.				yes
method	string		slide
edges	string		wrap
const	real		0
pad	boolean		no
center	boolean		no
normkernel	string		area
clobber	boolean		no
verbose	integer		0	0	5

Detailed Parameter Descriptions

Parameter=infile (file required filetype=input)

Input image file.

The input is an image and can have any number of dimensions. The input can have the following data types: "short" (BITPIX=16), "long" (BITPIX=32), "float" (BITPIX=-32), and "double" (BITPIX=-64).

The image can be a virtual image as defined by the datamodel. Thus one could specify a virtual image by using the "bin" syntax like my_file.fits[EVENTS][bin x=1:100:1, y=1:100:1] to specify a 2D image binned on X and Y columns in EVENTS extension of my_file.fits file.

Parameter=outfile (file required filetype=output autoname=yes)

Output image file.

The output file is an image and has the same dimensions as the input image. The output data type is always "float" (BITPIX=-32). The autoname filename suffix is "_conv".

Parameter=kernelspec (string required)

Kernel specification.

The generic syntax for the kernelspec specification is: [key]:[parameters]:[origin]

Where -

[key] = file : This tells the program to read the kernel from the image stored in the file [parameters]. The format of the specified file can be a FITS image of any of the data types specified for the input file.
[key] = txt : This tells the program to parse the string [parameters] to create the kernel. Ex. see example #3 (example of a 2D kernel).
[key] = lib : This tells the program to parse the string [parameters] for two things a) which library and b) parameters needed for that library call. The parameters needed follow the library specification in "()"'s.

Currently supported libraries are:

box

A multi-dimensional array with constant value. The parameters are (D, N, D1, D2,...DD) where:

D = number of dimensions
N = normalization (constant value)
D1 - DD = length of box in each dimension.

gaus

A multi-dimensional, non-rotated Gaussian. The parameters are (D, M, N, S1, S2, ... SD) where:

D = number of dimensions
M = number of sigma to extend in each direction
N = normalization
S1-SD = size of sigma d in pixels in each direction (s1 = size of 1 sigma in pixels in each direction, s2 is size of 2 sigma in pixels, etc.)

tophat

A two-dimensional non-rotated elliptical top-hat function. The parameters are (D, N, D1, D2) where:

D = number of dimension (ONLY = 2)!
N = normalization (what value)
D1 D2 = radii of ellipse axes along 1st and 2nd axes.

mexhat

A multi-dimensional, mexican hat function.

mexhat = Norm * exp( -0.5 * (x/s1)^2 + (y/s2)^2 + ...)

The parameters are (D, M, N, S1, S2, ... SD) where:

D = number of dimensions
M = number of sigma to extend in each direction
N = normalization
S1-SD = sigmas (width of mexican hat; units = pixels)

exp

A multi-dimensional exponential function.

exp = Norm * exp(-abs(x/s1)) * exp(-abs(y/s2)) * ...

The parameters are (D, M, N, S1, S2, ... SD) where:

D = number of dimensions
M = number of sigma to extend in each direction
N = normalization
S1 - SD = sigmas (width of exponential function; units = pixels)

power

A multi-dimensional power law function.

power = Norm * abs(x)^s1 * abs(y)^s2 * ...

The parameters are (D, M, N, S1, S2, ... SD) where:

D = number of dimensions
M = number of sigma to extend in each direction
N = normalization
S1 - SD = sigmas (width of power law function; units = pixels)

beta / lorentz

A multi-dimensional beta/lorentzian function.

beta = Norm / ((x^2 +y^2+...)/ s1^2)

The parameters are (D, M, N, S1) where:

D = number of dimensions
M = number of sigma to extend in each direction
N = normalization
S1 = sigma (width of beta/lorentzian function; units = pixels)

sinc

A multi-dimensional sinc function.

sinc= Norm * sin( 2*pi * x^2/s1 * y^2/s1 * ...

The parameters are (D, M, N, S1) where:

D = number of dimensions
M = number of sigma to extend in each direction
N = normalization
S1 = sigma (width of sinc function; units = pixels)

cone

A multi-dimensional cone function.

cone = Norm * 1 - sqrt(x^2+y^2+...)  / s1

The parameters are (D, N, S1)

D = number of dimensions
N = normalization
S1 = radius of the base of the cone (units; pixels)

pyramid

A multi-dimensional stepped pyramid function.

pyramid = Norm * 1 - MIN(x,y,...)  / s1

The parameters are (D, N, S1)

D = number of dimensions
N = normalization
S1 = radius of the base of the pyramid (units; pixels)

sphere

A hemisphere/ball kernel

sphere = sqrt( (x - s1)^2 + ( y - s1 ) ^ + ... )

The parameters are (D, R)

D = number of dimensions
S1 = radius of the base of the hemisphere (units; pixels)

hanning

Hann filter

hanning = Norm*(0.5-0.5*cos(2*pi*sqrt(x^2+y^2+...)/(S1-1) ))

The parameters are (D, N, S1)

D = number of dimensions
N = Normalization
S1 = total length of smoothing function (units; pixels)

racos

Generic raised cosine filter. With an alpha=0.5 it is identical to the Hanning filter. With alpha=0.54 it is a Hamming filter.

racos = Norm*(alpha-(1-alpha)*cos(2*pi*sqrt(x^2+y^2+...)/(S1-1)))

The parameters are (D, N, A, S1)

D = number of dimensions
N = Normalization
A = alpha parameter controling how quickly cos() is dampened.
S1 = total length of smoothing function (units; pixels)

The (optional) origin specification is used to locate the origin of the kernel. Typically 2D kernels have the origin in the "center" of the image. Many 1D kernels have the origin as the first element in the array.

The user can explicitly specify the origin. This alleviates common restrictions in some other tools (e.g. length of data axes must be an odd number).

If no origin is specified, the default is to assign the origin to the center of the array (rounded down). Thus a (5,4) array will have its origin set at (2,2).

Parameter=writekernel (boolean default=no)

Output kernel to an image?

Parameter=kernelfile (file filetype=output default=./. autoname=yes)

File name for kernel image. Autoname suffix is "_kernel".

Parameter=writefft (boolean default=no)

Output FFT of data and kernel?

Parameter=fftroot (file filetype=output default=./. autoname=yes)

Root file name for FFT output files. The autoname suffix for the real image is "_kernel_real". For the imaginary image, "_kernel_imag". The data file suffixes are "_data_real" and "_data_imag". ALWAYS use autonaming in this parameter.

Parameter=method (string default=slide)

Convolution method: (slide|fft).

Two convolution methods are implemented. Under some simple constraints they will provide the same results.

method=slide

Sliding cell convolution is a convolution from first principles.

O(a1,a2,...aN) = S_x1 S_x2... S_xN

I(x1,x2,..xN)*K(a1-x1,a2-x2,...aN-xN).

[replace S_x by summation signs]

method=fft

The FFT of the data and the kernel is computed. The arrays are multiplied and the inverse FFT is taken. The FFT route used is FFTW.

Internally the kernel and data MUST be the same size. The program will pad either/or both the data and/or kernel to the maximum length along either axis, defined by the larger of the two datasets.

If the edges of the data are "wrapped" (see below) then the FFT and slide methods will yield the same results.

For moderately-large to large kernels, the FFT convolution is much faster ( O(N log N) as opposed to O(N**2) ), however a) it requires much more memory and b) it restricts the edge treatment to "wrap".

Parameter=edges (string default=wrap)

How should the code handle edges? (wrap/nearest/mirror/constant/renorm).

As the kernel moves across the data, the edge of the kernel may fall off the data space. When this happens the program needs to know how the user wants to handle the edges. Five methods for handling the edges are implemented: For the 5 cases described below, consider the following example data space:

  data = { 1, 2, 3}

edges = wrap : When the kernel runs off the data space, it is wrapped around the data. Thus the data goes {...1,2,3,1,2,3,1,2,3...}
edges = nearest : When the kernel runs off the data space, extrapolate the data nearest to the edge, e.g. data goes {...1,1,1,2,3,3,3,3,3...}
edges = mirror : When the kernel runs off the data space, reflect the data nearest the edge, e.g. data goes {...3,2,1,1,2,3,3,2,1...}
edges = constant : When the kernel runs off the data space, use the constant supplied by the const parameter, e.g. const=0, data goes {...0,0,1,2,3,0,0,..}
edges = renorm : When the kernel runs off the data space, use a constant = 0 at the edge; however, re-normalize the kernel by the amount of area that remains on the data space.

If you set method="fft" then you are implicitly using the edges="wrap" option.

Parameter=const (real default=0)

Constant value to use at edges.

Parameter=pad (boolean default=no)

Pad the data to the next integer power of 2.

The user can specify that the data be padded such that the length of each data axes is promoted to the next integer power of 2. The data are always padded on the "right" hand side, thus the array 1,2,3,4,5 would be padded to 1,2,3,4,5,0,0,0. This way of padding results in no change in the origin.

Parameter=center (boolean default=no)

Center FFT output?

If center = yes, zero frequency will appear at the center of the FFT images. This option is currently inoperative.

Parameter=normkernel (string default=area)

Normalize the kernel?

Allows aconvolve to automatically normalize the kernel. There are three options.

area - divide the kernel values by the sum of the values. NB: negative values in kernel may produce undesired result.
max - divide the kernel values by the maximum kernel value (thus the max value is normalized to 1).
none - do not normalize the kernel.

Parameter=clobber (boolean default=no)

Clobber existing output?

Parameter=verbose (integer default=0 min=0 max=5)

Verbosity level: 0 - no output, 5 - max display.

Changes in CIAO 4.17

Internally, the tool now always performs FFT convolutions at double precision. The increased dynamic range improves numerical accuracy in particular for small pixel values in data sets that have a large range of pixel values. Users can expect to see some differences compared to previous versions.

FFT convolutions are now performed at double precision. The increased dynamic range provides better numeric accuracy especially for small pixel values in the presence of a large range of pixel values.
Bug fix when the input image was an unsigned byte (8-bit) image. Previously the data was read in as a signed byte.

Bugs

The 2D raised cosine and hanning convolution kernels are not correct.: The 1D versions are correct.