PaperlayerView exact revision on arXiv

Source-linked AI summary

Multi-Scale CLEAN deconvolution of radio synthesis images

T. J. Cornwell

arXiv:0806.2228v1astro-ph

TL;DR

CLEAN deconvolution is effective for point sources but converges slowly and can become unstable for extended emission. The paper introduces Multi-Scale CLEAN, which models emission with components across simultaneously evaluated scales and improves convergence and stability while remaining simple to implement.

  • Problem

    CLEAN works well for point sources but models extended emission pixel by pixel, causing slow convergence and instabilities without a guarantee of global optimality.

  • Method

    Multi-Scale CLEAN extends CLEAN by modeling sky brightness as scaled, centered extended components selected simultaneously across a specified range of scales.

  • Results

    Multi-Scale CLEAN improves convergence and stability, remains stable with spatially varying backgrounds when sufficiently large scales are included, and works well in practice.

  • Takeaways & Limitations

    For extended emission, Multi-Scale CLEAN is faster than Högbom or Clark CLEAN, shows less low-to-moderate signal-to-noise bias than MEM, and handles negative flux.

  • Takeaways & Limitations

    The scale set must be chosen arbitrarily, and the algorithm introduces an arbitrary bias toward small scales.

Abstract

from arXiv · show

Radio synthesis imaging is dependent upon deconvolution algorithms to counteract the sparse sampling of the Fourier plane. These deconvolution algorithms find an estimate of the true sky brightness from the necessarily incomplete sampled visibility data. The most widely used radio synthesis deconvolution method is the CLEAN algorithm of Hogbom. This algorithm works extremely well for collections of point sources and surprisingly well for extended objects. However, the performance for extended objects can be improved by adopting a multi-scale approach. We describe and demonstrate a conceptually simple and algorithmically straightforward extension to CLEAN that models the sky brightness by the summation of components of emission having different size scales. While previous multiscale algorithms work sequentially on decreasing scale sizes, our algorithm works simultaneously on a range of specified scales. Applications to both real and simulated data sets are given.

I. Introduction

CLEAN enables radio synthesis imaging from incomplete Fourier-plane coverage but converges slowly and can become unstable for extended emission. Multi-scale deconvolution addresses this by modeling emission across different size scales, with the proposed algorithm selecting among scales simultaneously.

  • CLEAN enables synthesis imaging of complex objects despite relatively poor Fourier-plane coverage.
  • Extended objects expose CLEAN's slow convergence and instability because the algorithm identifies emission pixel by pixel.
  • Multi-scale methods model the recovered object using components or channels at different scale sizes, reducing reconstruction degrees of freedom.
  • Existing approaches include Multi-Resolution CLEAN, Multi-scale Maximum Entropy, wavelets, pixons, and adaptive scale pixels, with some methods carrying computational or application limitations.Adaptive Scale Pixels are described as computationally expensive, while pixons lack published success on synthesis observations because their compact-PSF assumption does not hold for Fourier synthesis.
  • The paper presents a conceptually and algorithmically straightforward CLEAN extension that improves convergence and stability while selecting among scales simultaneously.It retains CLEAN's pattern-matching motivation and extends the model to extended emission as well as point sources.

II. Background

Radio synthesis reconstructs sky brightness from incomplete, noisy Fourier samples, producing a dirty image that is the true image convolved with a dirty beam. Högbom CLEAN iteratively fits point-source components by locating residual peaks and subtracting scaled point-spread functions, but its greedy choices are not guaranteed to be globally optimal.

  • Radio synthesis arrays measure Fourier components of sky brightness through antenna baselines and a virtual aperture.
  • Incomplete discrete visibility samples yield a dirty image, which equals the true image convolved with the dirty beam.
  • The deconvolution problem is to recover the true image from the dirty image and point spread function.
  • Högbom CLEAN models sky brightness as a collection of point sources and iteratively locates residual peaks before subtracting scaled, centered point-spread functions.
  • CLEAN is greedy: each iteration uses only currently available information, so global optimality is not guaranteed.

III. The Multi-Scale CLEAN algorithm

Multi-Scale CLEAN extends CLEAN by representing sky brightness with centered components across multiple scales and searching those scales simultaneously. The approach is straightforward to implement, supports extended emission, and has stated convergence and scale-selection boundaries.

  • Model: The sky model sums appropriately scaled, centered extended components, with component width represented by αq.A zero-width scale is included so the finest-scale emission can be fit.
  • Search and updates: The algorithm searches for component position, strength, and scale using a greedy strategy based on peaks in the residual-related image.Component magnitude is set by loop gain times the peak, while precomputed terms reduce updates to scaling, shifting, and subtraction.
  • Component design: The component shape is a tapered, truncated parabola chosen to support finite-window constraints, while a prolate spheroidal wave function closely approximates a scaled Gaussian.Gaussian tails can violate support constraints and mainly become problematic at high dynamic range or with image-plane support limits.
  • Properties and limits: Convergence is assured under the stated proof conditions when the PSFs are accurate and the component shape is appropriate, but the greedy method is not globally optimal.The convergence proof requires positive-semidefinite PSFs.
  • Scale search: Unlike Multi-Resolution CLEAN, Multi-Scale CLEAN searches all specified scales simultaneously, allowing scale-selection errors to be corrected immediately.Multi-Resolution CLEAN instead processes scales sequentially from broad to fine emission.
  • Properties and limits: Scale selection is ill-defined but often workable with a few geometric scales, while rotational symmetry reduces search dimensionality at the cost of representing sharp edges less well.The paper identifies both scale-grid choice and exhaustive parameter search as areas for improvement.

A. Performance on real data

On a low-signal-to-noise VLA observation of NGC1058, broad sidelobes and a negative bowl make integrated emission difficult to estimate. Multi-Scale CLEAN addresses the broad structure first and recovers substantially more integrated flux than Högbom CLEAN in the reported test.

  • Data and challenge: The NGC1058 test uses one HI channel observed with the VLA, where extended structure and low signal-to-noise make broad sidelobes more troublesome than fine-scale sidelobes.The dirty image contains a broad negative bowl that prevents accurate integrated-emission estimates.
  • Flux recovery: 3.08Jy versus 0.32Jy integrated flux: Multi-Scale CLEAN recovered more flux than Högbom CLEAN at convergence.The Multi-Scale value remained robust when the integration region was changed beyond the nominal source extent.
  • Deconvolution behavior: Multi-Scale CLEAN removes the largest-scale emission first and then proceeds toward finer detail, opposite to Högbom CLEAN's fine-detail-first behavior.This ordering targets the broad structure responsible for the negative bowl before estimating finer emission.
  • Convergence behavior: At many iterations, Multi-Scale CLEAN produces comparable peak residuals across scales and relatively stable scale-dependent flux values.Högbom CLEAN continues finding more flux as iteration progresses.
  • Convergence behavior: Recovering more flux with Högbom CLEAN requires cleaning below 4σ, which is slow and distorts the appearance of the noise.The Multi-Scale test stopped when the peak residual reached 4σ.

B. Performance on simulated data

Simulated VLA observations tested CLEAN variants, Maximum Entropy, and Multi-Scale CLEAN on the M31 and Hydra images. Multi-Scale CLEAN converged efficiently, recovered flux well, and showed less source-correlated residual structure, while Clark CLEAN exhibited divergence and flux loss in the M31 test.

  • Simulation setup: The simulations evaluated Högbom and Clark CLEAN, Maximum Entropy, Multi-Scale CLEAN, and related restored, residual, and error images on M31 and Hydra observations.The M31 simulation used a well-sampled VLA C-configuration observation with Multi-Scale CLEAN scales of 0, 1.5, 3, 6, 12, and 24 pixels.
  • Convergence: Clark CLEAN diverged above 300,000 CLEAN components, recovering 1391Jy of the full 1495Jy, whereas Multi-Scale CLEAN converged well in only 5000 iterations.Högbom CLEAN also converged well in this simulation.
  • Residual structure: Multi-Scale CLEAN produced residuals with little correlation to the source structure, unlike the more obvious residual structure in the other cases.The comparison used error patterns calculated against the appropriately smoothed model image.
  • Flux recovery: Entropy, Högbom CLEAN, and Multi-Scale CLEAN recovered the full object flux well, while Clark CLEAN was biased down by about 7%.The Clark CLEAN error image also showed a substantial negative bowl.
  • Limitations: All CLEAN algorithms remained liable to fringing on extended emission, but simulations showed this effect at a lower level for Multi-Scale CLEAN.Fringing arises from poor interpolation in holes and at the edge of the sampled Fourier plane and appears in error images rather than residual images.

C. Performance for varying source size and noise level

Simulations varied source size and thermal noise to compare deconvolution quality using recovered flux and dynamic range. Multi-Scale CLEAN approached MEM for very extended structure, retained accurate flux estimates at lower SNR, and benefited from appropriately chosen scales and masks.

  • Simulation design: The simulations varied source size and thermal noise, assessing image quality through recovered flux and dynamic range.Recovered-flux deviations indicate systematic imaging errors relevant to quantitative use of reconstructed images.
  • Varying source size: Multi-Scale CLEAN was nearly as good as MEM at recovering flux from very extended M51 model structure.Figure 6 compares Multi-Scale CLEAN, MEM, Maximum Emptiness, and Clark CLEAN as source structure becomes larger.
  • Varying source size: For very large sources near 0.8 λ/D and 1 λ/D, larger maximum scales appreciably increased recovered flux.These sources exceeded the size that could be accurately imaged in the VLA C array.
  • Noise response: MEM overestimated source flux from image-plane SNR as high as 100, whereas CLEAN, Maximum Emptiness, and Multi-Scale CLEAN remained accurate to about SNR 4.CLEAN also estimated total flux reasonably well below image-plane SNR 1.
  • Scale and mask interaction: For a 0.83 λ/D model image, image quality improved with a maximum scale of about one quarter of the object size and a very tight mask.The result reflects interaction between maximum scale size, mask tightness, and image quality as Multi-Scale CLEAN began to fail.
  • Figure 6: Figure 6 varies simulated source size on the horizontal axis while comparing algorithms in panel (a) and Multi-Scale CLEAN scale choices in panel (b).Each case uses four scales from a point source to a listed maximum ranging from 8 pixels (0.027 λ/D) to 64 pixels (0.21 λ/D).

V. Summary

Multi-Scale CLEAN works well in practice and is simple to understand and implement. Its practical profile combines reduced low- to moderate-SNR bias and support for negative flux with arbitrary scale selection and a speed trade-off relative to other methods.

  • Summary: Figure 7 presents image quality versus image signal-to-noise ratio using recovered flux and dynamic range.Dynamic range is defined as reconstructed-image peak divided by off-source RMS.
  • Summary: Multi-Scale CLEAN has worked well in practice over 5–6 years and is simple to understand and implement.The paper identifies simplicity alongside the algorithm’s practical efficacy as a major attraction.
  • Summary: Multi-Scale CLEAN shows less bias than MEM at low to moderate signal-to-noise and can be applied to images with negative flux.For extended emission, it is faster than Högbom or Clark CLEAN but usually slower than MEM.
  • Summary: The scales must be chosen arbitrarily, and Multi-Scale CLEAN requires an arbitrary bias toward small scales.These are identified as the algorithm’s most notable deficiencies.
  • Summary: Figure 8 examines how maximum scale size interacts with imaging-mask tightness for a 0.83 λ/D M31 HII-region model.The figure addresses image-quality effects of these two choices together.
Loading 0806.2228v1…