Abstract: The fundamental idea and the systematic design procedure of a novel internal model filtering (IMF) approach are proposed to deal with the filtering problem under extremely low SNR (signal-to-noise ratio) case, such as extracting the visual evoked potential (VEP) signal embedded in strong electro-encephalogram (EEG) background activity. The internal model parameters of a typical VEP signal are estimated via the standard RLS (recursie least squares) estimation method. Then the internal model filter is designed. finally, the filter is used for extracting the VEP. Both the simulations and real clinical applications have shown its superior signal-noise separation performance in the case of low SNR.