. We analyze vorticity fields obtained from direct numerical simulations (DNS) of statistically stationary two-dimensional turbulence where the forcing is done in wavelet space. We introduce a new eduction method for extracting coherent structures from two-dimensional turbulent flows. Using a nonlinear wavelet technique based on an objective universal threshold we separate the vorticity field into coherent structures and background flow. Both components are multi-scale with di#erent scaling laws, and therefore cannot be separated by Fourier filtering. We find that the coherent structures have non-Gaussian statistics while the background flow is Gaussian, and we discuss the implications of this result for turbulence modelling. 1. Introduction Our aim is the numerical simulation of large Reynolds number flows using present computers. In this context the key question is: which quantities should be computed and which quantities should be discarded or modelled ? The usual choice, based on ...