Block Wiener-based image registration for moving target indication

This paper describes two novel Wiener-based approaches for frame-to-frame image registration. The first approach, Local Wiener Registration, incorporates the crosscorrelation between the two frames to determine the Wiener filter at each pixel location in order to predict the second frame from the first frame. The second approach, Block Wiener Registration, divides the image in nonoverlapping blocks and computer the Wiener filter for each block. The output of the Block Wiener Registration is the a weighted sum of the bank of Wiener filter outputs. By computing the Wiener filter for a few blocks in the imagery instead of at each pixel location, the Block Wiener Registration is significantly less computationally complex than the Local Wiener filter. Furthermore, both Wienerbased methods are able to compensate for localized motions of the background, e.g., parallax. Experimental results indicate that the Block Wiener Registration is nearly as effective as Local Wiener Registration and is orders of magnitude faster. Finally, both Wiener-based approaches outperform parametric registration approaches that account for the global changes from one frame to the next.