The expectation-maximization (EM) algorithm applied to image identification and restoration

Prior to restoring a noisy blurred image, the degradations the image has suffered need to be determined. We present a maximum likelihood blur identification method which estimates the required parameters from the observed noisy blurred image itself, and propose to employ the expectation-maximization algorithm to solve the resulting complicated problem of optimizing the likelihood function. A priori information in the form of [i] initial conditions about the unknown parameters, and [ii] parametric image and blur models are incorporated to make the algorithm applicable to realistic blurs and to improve the identification results. The proposed approach results in a flexible iterative algorithm which is computationally far more efficient than directly optimizing the likelihood function.