Efficient 3D kernel estimation for non-uniform camera shake removal using perpendicular camera system

Non-uniform camera shake removal is a knotty problem which plagues the researchers due to the huge computational cost of high-dimensional blur kernel estimation. To address this problem, we propose an acceleration method to compute the 3D projection of 2D local blur kernels fast, and then derive the 3D kernel by interpolating from a minimal set of local blur kernels. Under this scheme, a perpendicular acquisition system is proposed to increase the projection variance for reducing the ill-posedness of 3D kernel estimation. Finally, based on the minimal 3D kernel solver, a RANSAC-based framework is developed to raise the robustness to estimation error of 2D local blur kernels. In experiments, we validate the effectiveness and efficiency of our approach on both synthetic and real captured data, and promising results are obtained.