Optical Aberration Correction by Divide-and-Learn for Accurate Camera Calibration

The accuracy of three dimensional vision depends heavily on the accuracy of camera calibration. A major source of calibration error is the system nonlinearity due mainly to optical aberration. Although there are various physical models that have been employed to correct the nonlinear image distortion due to the aberration, it is uncertain practically that which model best fits a given optical system. In this paper, an intelligent learning technique to correct errors from the nonlinear optics is proposed. Data errors are first divided into small groups using k-means clustering algorithm, and an error correction function is approximated by training a small neural network that is allocated to each divided group. Compared with conventional methods, the proposed method showed higher accuracy in our tests.