Efficient image registration by decoupled parameter estimation using gradient-based techniques and mutual information

We present an efficient and accurate method for similarity (rigid+scale) registration of 3D images by decoupled estimation of transformation parameters. The estimated parameters are used to initialize the search for a more accurate alignment by optimizing a mutual information-based cost function. We use computationally inexpensive methods based on image gradients to obtain the original estimates. Our method improves the efficiency of the overall registration task by allowing the more expensive optimization step to converge quickly. It also makes the registration more robust and less sensitive to the shape of the cost function by reducing the chance of the optimization algorithm being caught by a local minimum. Our experiments demonstrate up to 10- fold increase in computational efficiency and significant reduction in misalignment.