Image registration has many real life applications. Affine image registration is one of the commonly-used parametric models. Iterative solution methods for the underlying least squares problem suffer from convergence problems whenever good initial guesses are not available. Variational models are non-parametric deformable models that have been proposed based on least squares fitting and regularization. The fast iterative solution methods often require a reliable parametric (affine) method in a pre-registration step. In this paper we first survey and study a class of methods suitable for providing the good initial guesses for the affine model and a diffusion based variational model. It appears that these initialization methods, while useful for many cases, are not always reliable. Then we propose a regularized affine least squares approach that can overcome the convergence problems associated with existing methods. Combined with a cooling idea in a multiresolution setting, it can ensure robustness and selection of the optimal coupling parameter efficiently. Numerical examples are given to demonstrate the effectiveness of our proposed approach.