Deep learning for rectification of radial distortion image

Abstract

Fisheye lens and telephoto lens for long distance are becoming more widely used because of their larger field of view or much longer focal length than normal pinhole camera. The larger field of view is obtained by the image radial distortion - the effect which should be rectified not only for visualization but for use by manifold applications using image based measurements. This research article applies deep learning to address the computer vision problem of image rectification using the single image acquired by the unknown camera. A large scale of radial distortion image dataset is synthesized following the general optical ray projection model in angular form to train the proposed network. The experimental results verify that the constructed network learns distortion information from input images with high accuracy and performs the image rectification without explicit use of image semantic information. It was possible due to generating of good training distortion models and their inversion by Newton optimization and pseudo-inverse of resulting Vandermonde matrix.