Recursive learning of image parameters for totally blind image restoration

In this work, totally blind image restoration problem is solved using an Expectation Maximization (EM) based learning technique. A new formulation on simultaneous recursive image and blur parameter identification and image restoration is given in a dynamic Bayesian network (DBN) framework. This technique incorporates optimal Kalman smoothing equations for Maximum Likelihood (ML) parameter identification and state estimation. Because of the computationally heavy processing of smoothing, we use filtering approximation of the Kalman instead of Kalman Smoothing. The experimental results are given using a 64 x 64 "lena" image which is both noisy and artificially blurred.