Signal Processing and Communications Laboratory

Department of Engineering

Image Enhancement

A number of members are active in the Image Enhancement group within Signal Processing and Communications Laboratory. Prof. Nick Kingsbury is the head of the group.

Faculty

Post-doc

  • Rich Wareham

Research Students

Recent past members of the group

  • Yingsong Zhang (2012)
  • Henry Gomersall (2011)
  • James Ng (2007)
  • Mark Miller (2006)

Background

Our work on Image Enhancement is based on wavelet-domain image processing, and arises from the relatively unique property of wavelets to 'sparsify' images and video signals. For maximum performance we usually choose to use the Dual-Tree Complex Wavelet Transform (DT-CWT), because of its unique combination of attributes: strong directional selectivity, shift invariance of basis functions, tight-frame energy preservation, and computational efficiency. We also employ the approximately linear phase-shift properties of complex wavelet coefficients to estimate motion when the input image is subject to displacements due to object motion or atmospheric disturbance.

Our work on enhancement divides into two main areas: (a) deconvolution / super-resolution of images or 3D datasets; and (b) stabilisation / fusion of video imagery.

For (a), we regard the problems as being large inverse problems in the presence of measurement noise, which are often under-sampled or lossy (the compressive sensing scenario) and which require strong regularisation in order to produce good solutions. We are therefore developing new methods to encourage sparse solutions in the complex wavelet domain and which contain tree-structured model constraints too. The methods are usually iterative but with a strong emphasis on rapidly convergent techniques that require only 10 to 50 iterations to achieve good solutions, even with very large dataset sizes (107 or more). Currently we favour non-convex reweighted-L2 (Gaussian scale mixture) models for inducing sparsity and link these with variational Bayes continuation techniques in order to achieve rapid convergence to locally optimal solutions that appear experimentally to be close to global optimality. We have found that non-convex reweighted-L2 methods significantly outperform more conventional convex L1 techniques for encouraging sparsity.

For (b), we employ complex-wavelet-based motion estimation methods, which use inter-frame phase shifts of coefficients in a coarse-to-fine manner to converge rapidly to smooth and realistic displacement fields. After frame-to-frame registration with these fields, blocks of frames are non-linearly fused so that any remaining motion jitter is removed and high-frequency edges are recovered from regions that are locally in-focus. These methods are highly effective at combatting the effects of atmospheric turbulence and image speckle on video sequences.

You can find more details by clicking on the individual web-pages of the group's members.