Post date: Jun 28, 2009 6:53:16 PM
I am currently implementing a multi-scale registration framework. This project was not explicitly stated in my goals, but in attempting to find the best settings for the optimizer and metric, it is becoming a critical subtask.
Here is an outline of the method:
1. use median filter to remove detail from 2D fixed image and 3D DRR image. The kernel size is determined by the scale of the registration process (larger kernel for coarser scales).
2. Step size of the optimizer will be directly proportional to the registration scale (larger steps for coarser scales).
3. After each scale, when the optimizer has converged to a solution, that solution will be used as the starting point for the next scale of the optimizer.
4. At each scale, the kernel used in the median filter will be reduced until it eventually gets to 1 (full resolution image).