Characterizing the surface of microlenses by optical profilers has the important advantages of measurement speed, flexibility and automation. Nevertheless, the accuracy of such characterization is limited by error occurring in non-flat measurements. Here, we propose a method that uses multiple measurements of a single reference ball combined with a machine learning algorithm that fits the experimental data to correct the measurements. The success of the method is demonstrated by showing that the residual error after correction reaches 20nm RMS. Such results extend greatly the quality of microlens characterization by optical profilers.