A camera neural model

Most of the techniques proposed for camera calibration consist of identifying a set of physical parameters (focal length, optical center position, coefficient of radial distortion). The determination and inversion of such a model in the case of strong or unstructured distortions is not straightforward. Instead we propose a new approach that uses artificial neural networks´ abilities of interpolation and extrapolation to perform the global identification of any optical system, provided a set of pairs of calibration points given in a camera-centered frame of reference and their corresponding image in the image plane are available. A methodological framework is given that deals with the architecture of the neural network the training algorithm and the choice of the learning database. Validity and robustness of the method are shown by comparing reconstruction errors obtained by the neural model, by a theoretical prediction and by classical estimations on both linear and nonlinear lenses