## Deconvolution and Restoration

Miriad contains two tasks to deconvolve the mosaiced dirty images produced by invert. In terms of theory, practical use and indeed internal implementation, these tasks are quite similar to the deconvolution tasks described in Chapter 14. The major difference is that the convolution' operation (which turns a prospective model into a dirty image) is somewhat more involved. Also account must be made of the changing noise level across the dirty image.

The two mosaic deconvolution tasks are mosmem, which implements a maximum-entropy-based deconvolution algorithm, and task mossdi, which uses a Steer, Dewdney & Ito (SDI) CLEAN algorithm. Generally mosmem is superior, although mossdi can be better for images containing point sources. Note that, although you can make mosaiced, multi-frequency synthesis images with invert (and, indeed, produce a mosaiced, spectral dirty beam), there is no mosaic equivalent to mfclean. In deconvolving a mosaiced, multi-frequency image you will have to tactically assume that the spectral index is 0. This should not be a problem - primary beam model errors are probably more significant than spectral errors in these deconvolutions.

If you are deconvolving, note the recommendations for invert's imsize parameter, and the use of options=double.

If you are familiar with the inputs to the conventional deconvolvers, the inputs to mosmem and mossdi should be fairly straightforward. In the case of the inputs to mosmem and maxen, apart from differences in the options, the meaning of the flux keyword and the default region, the only significant difference is in specifying the expected RMS noise level in the dirty image. Because the noise level varies across the dirty image, mosmem uses the theoretically expected noise level (which it computes) times a user-specified fudge factor, rmsfac. That is, if rmsfac is set at 1 (the default), then mosmem uses the theoretical noise level when calculating its statistic.

Typical inputs to mosmem are:

 MOSMEM map=lmc.map Dirty image produced by invert. beam=lmc.beam Beam dataset. model An initial model estimate - generally unset. default The image that the solution should tend towards - generally unset. out=lmc.model The output dataset. niters=30 Maximum number of iterations - default is 30. region= Region to deconvolve. The default is the entire image. measure Leave unset gives you the Gull measure. flux= Estimate of the total flux - its best to give a value. rmsfac=1 RMS noise fudge factor. Default is 1. q An initial estimate of the beams volume. Generally you can leave this unset. options Generally leave unset, or options=doflux use doflux` to enforce the flux constraint.

The inputs and use of mossdi should be equally simple for someone familiar with clean.

Having produced a model, we generally want to convolve this with a Gaussian CLEAN beam and add in the deconvolution residuals. This is done by restor. The inputs and use of restor is identical to a conventional observation (restor is the only general task which is smart enough to recognise a mosaiced experiment directly). Task restor uses a constant CLEAN beam - it is not a function of position. The only caveat is that, when determining a default CLEAN beam, restor fits a Gaussian to the synthesised beam which corresponds to the first pointing. Provided the first pointing is a fairly typical pointing, this will probably be adequate. Otherwise you may wish to use task mospsf (see Section 21.6 below) to generate an actual point-spread function (at some position) and then use imfit to determine Gaussian parameters for it.

Typical inputs to restor are:

 RESTOR map=lmc.map Dirty image produced by invert. beam=lmc.beam Beam dataset. model=lmc.model Model produced by mosmem. mode Leave unset to get restored image. fwhm Beam size - leave unset to let restor fit it, but to the first pointing! pa Again leave unset to let restor fit it.