Restoration of spatially varying images using multiple model extended Kalman filters

Image restoration based upon unrealistic homogeneous image and blur models can result in highly inaccurate estimates with excessive ringing. Thus it is important at each pixel location to restore the image using the particular image and blur parameters characteristic of the immediate local neighborhood. One approach which has been used for control systems for both online parameter identification and state estimation is the extended or linearized Kalman filter (EKF). However this procedure was found to be unsuitable for blur parameter identification because of the presence of a significant process noise term that caused large deviations between the predicted pixel estimates and the true pixel intensities. Thus as an alternative, a multiple model EKF procedure was developed and tested for spatially varying parameterized blurs. Results show this procedure to be very useful for restoring representative images with significant simulated variations of the blur parameter