I have a shape fitting problem in image processing that I'm solving using EM but there are some constraints to the final shape. The way I'm currently solving this is by checking at each iteration if the shape is within those constraints, and forcing it in place if not (in those places where it is going wrong).
Does anyone know how I could integrate a prior probability into the EM process, in a way to make it mathematically sound? Any reference on how to do EM within a Bayesian framework?
You could add a regularization term that reflects the shape prior. Another approach would be to build a statistical model of the shape, ie via eigen-shape as in ASM, and optimize within the linear shape space. Check out my work on lip tracking using that idea (although I use gradient descent for optimization). I believe you could easily adapt the shape space idea to the EM setting if you want to use EM. The prior would reflects the density distribution of the principle modes in the eigen-shapes space.
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing