Next: The Boundary and Up: Adaptive Least Squares Image Previous: Restoration Based on

Optimization of the Regularization Function

In order to complete the design of the restoration algorithm we have to select the regularization function . In Eq. (8) diagonal weighting matrices were introduced which means that correlations are neglected. The result of the convolution operation should be uncorrelated, or in other words, ``white''. Therefore is also called a ``whitening filter'' (Yaroslavsky 1985). Approximations to this type of filter are differential operators (Twomey 1963), the Laplace operator (Pratt 1978), and the linear predictor (Bundschuh 1993):

The application of the optimal whitening filter leads to a constant power density spectrum. Obviously the modulus of the transfer function of the optimal whitening filter must be reciprocal to the modulus of the transfer function of the original signal. Its phase still remains to be determined by some other optimality criterion. The resulting whitening operator is obtained by means of an inverse Fourier transform:

The size of the whitening operator which is used in practice should be as small as possible in order to save computer time. In this paper an operator with 55 coefficients is used for the processing of simulated as well as measured data. Basically this truncation reduces the decorrelation properties. The optimality criterion for the phase function is minimal spatial dispersion which minimizes the truncation error. Partial differentiation with respect to the unknown phase function and equating the result to zero leads to a system of non-linear equations:

A general solution cannot be found analytically, but some special solutions can easily be obtained. The restriction to real operators enables the use of symmetry properties of the phase function. A special solution which leads to minimal dispersion is

From Eq. (19) we see that basically the power density spectrum and the autocorrelation function of the original signal should be known. If available, some a priori knowledge of the statistics of the original signal can be used. In many cases images can be modeled quite well as two-dimensional first order Markov processes (Pratt 1978):

Since the coarse structure of the underlying signal can be seen from the measured data, the power density spectrum of the measured data was used for the processing of simulated as well as measured data instead of the unknown power spectrum of the original signal:

Eventually the self-updating iterative procedure which is used for the weighting function using the intermediate results of the iteration can be used for the power density spectrum:

Generally a variation of the regularization function within the image should enable the adaptation to local statistical image properties. Fig. 2 shows an example for whitening operator . Obviously it is a high pass filter which is reasonable for regularization functions. Qualitatively it shows some similarity to the Laplace operator.



Next: The Boundary and Up: Adaptive Least Squares Image Previous: Restoration Based on


rlw@
Thu Jun 2 15:47:14 EDT 1994