Blind deconvolution of blurred images
A.H. Tewfik · Annual Meeting Optical Society of America · 1989
We consider the problem of reconstructing a finite image x from a blurred observation y, where y = Akx + n. In this equation Ak is an unknown blurring operator from a known class {Ak; k = 1, K} of such operators and n is an observation noise. The operator Ak may be continuous or discrete and n is modeled as a 2-D white Gaussian noise process. The reconstruction procedure is based on the observation that Ak is singular (or highly ill-conditioned and thus singular for all practical purposes). It consists in projecting the received image y onto the null space of each of the possible blurring operators Ak. The projection onto the null space of the actual blurring operator will consist of noise only whereas the projections onto the other null spaces will have a signal component in addition to the noise component. Thus, we can estimate the actual blurring operator by picking the projection that has minimum energy. It is shown that this is equivalent to performing a simultaneous maximum likelihood estimation of Ak and x. To minimize the probability that the wrong projection is chosen a projection dependent constant is subtracted from the energy of each projection before the comparison is done. The constant depends on the variance of the noise process and the actual null space being considered. This modification allows us also to deal efficiently with the case where the null space of a blurring operator is entirely included in that of another operator. Once the blurring operator is identified, the image x is reconstructed using the pseudoinverse of the identified blurring operator.