Correcting distortion of image by image registration with the implicit function theorem

摘要

We propose a method for correcting image distortion due to camera lenses by calibrating intrinsic camera parameters. The proposed method is based on image registration and doesn't require point-to-point correspondence. Parameters of three successive transformations view change, radial distortion and illumination change are estimated using the Gauss-Newton method. Estimating all 19 unknowns simultaneously, we introduce the implicit function theorem for calculating the Jacobian. To avoid local minima, we first estimate parameters for view change and employ coarse-to-fine minimization. Experimental results using real images demonstrate the robustness and the usefulness of the proposed method.