Image registration in shifted focus

Image registration is a key requirement in many multi-frame imaging solutions like focus stacking, stereo imaging, 2D to 3D video reconstruction etc. Several techniques are proposed in literature based on key point or feature detection. Most of these techniques involve correlation, feature or key point detection, projection on multiple coordinates etc. But such methods work fine for images with uniform focus as the features or key points are consistent across the frames. But if the focus regions are different in different frames, these features do not show consistency across the frames and hence the alignment of frames becomes difficult. In this paper we propose a method of aligning images with different focus points using multilevel down-sampling and correlation maximization. The proposed method is highly accurate and reasonably fast. The method may be useful in focus stacking, HDR and many other applications.