My algorithm uses a basic SSD as the metric, with an image pyramid to optimize. Before computing the SSD, the algorithms crops out the outer quarter of both images, so it only compares the middle segment with half the width and height of the entire image. This elimates all the noise from the borders from the computation.
The only issue that was sometimes encountered was that the pyramid actually severely reduced the quality of the algorithm. If the iamge was shrunk and blurred too much, the initial guess would be wildly inaccurate which would cause the image to be severely misaligned. This was corrected by using a smaller pyramid depth.
