ImageDeconvolve

ImageDeconvolve[image,ker]
gives a deconvolution of image using kernel ker.

Details and OptionsDetails and Options

  • The deconvolution kernel is given as a numerical matrix or as an image and cannot be larger than image in any dimension.
  • The kernel represents the point spread function, which is assumed to model the blur present in the image.
  • ImageDeconvolve[image,ker] always gives an image of a real type of the same dimensions as image.
  • ImageDeconvolve operates separately on each channel in image.
  • The deconvolution kernel must have a single channel or the same number of color channels as the image.
  • The following options can be specified:
  • MaxIterations10maximum number of iterations to use
    Method"DampedLS"method to use
    Padding"Reversed"padding method
  • Possible settings specifying spectral deconvolution methods are:
  • "DampedLS"damped least squares, generalized Tikhonov regularization
    "Tikhonov"Tikhonov regularization method
    "TSVD"truncated singular value decomposition
    "Wiener"Wiener deconvolution
  • For spectral deconvolution methods, a regularization parameter p can be given with a setting Method->{"method",p}. For non-negative deconvolution kernels in which the sum of all elements is equal to 1, the regularization parameter is typically in the range from 0 to 1.
  • With a setting Method->{"method",{p1,p2,}}, separate regularization parameters can be given for each color channel.
  • The following settings for the Method option specify iterative deconvolution methods:
  • "Hybrid"TikhonovGolubKahan bidiagonalization regularization
    "RichardsonLucy"RichardsonLucy iterative deconvolution
    "SteepestDescent"modified residual norm steepest descent
  • The and methods always return non-negative pixel values.
  • Iterative deconvolution methods typically provide better results than spectral methods but are computationally more expensive. By default, preconditioned versions of the iterative methods are used. Preconditioning can be disabled by setting Method->{"method","Preconditioned"-> False}, resulting in slower convergence.
  • The classical RichardsonLucy deconvolution method does not use preconditioning.
  • ImageDeconvolve also supports an iterative method that effectively implements an iterative total variation regularization algorithm.
  • The following suboptions can be specified with a setting Method->{"TotalVariation",subopt}:
  • "NoiseModel""Gaussian"noise model
    "Regularization"Automaticregularization parameter
  • Possible settings for are , , or .
  • Note that ImageDeconvolve and ImageConvolve use different default settings for Padding.
  • ImageDeconvolve works with Image3D objects.

Background
Background

  • ImageDeconvolve performs deconvolution, which is typically used to enhance local details and edges in an image. Deconvolution is the converse operation to convolution, but unlike convolution, it is nonlinear, ill-posed, and non-unique.
  • Images may be blurred due to camera motion, object motion, lens defects, orin the case of astronomical imagesatmospheric turbulence. Deconvolution can sometimes be used to partially reverse the effects of these and other undesirable convolution-type processes in an image.
  • A kernel used for deconvolution is often referred to as a "point spread function" (commonly abbreviated "psf") and is assumed to model the blur whose removal from an image is being attempted. If the point spread function does not match the blur actually present in an image, deconvolution will fail to recover details and may even add spurious artifacts. Furthermore, even if a representative point spread function is chosen, deconvolution may still not recover all details in the original image.
  • The converse operation to ImageDeconvolve is performed by ImageConvolve. Sharpen is another simple way of enhancing edges.
Introduced in 2010
(8.0)
| Updated in 2012
(9.0)