Neural network decision directed edge-adaptive Kalman filter for image estimation

A neural network-based scheme for decision directed edge-adaptive Kalman filtering is introduced in this work. A backpropagation neural network makes the decisions about the orientation of the edges based on the information in a window centered at the current pixel being processed. Then based upon the neural network output an appropriate image model which closely matches the local statistics of the image is chosen for the Kalman filter. This prevents the oversmoothing of the edges, which would have otherwise been caused by the standard Kalman filter. Simulation results are presented which show the effectiveness of the proposed scheme.