Exploiting the self-similarity in ERP images by nonlocal means for single-trial denoising

Event related potentials (ERPs) represent a noninvasive and widely available means to analyze neural correlates of sensory and cognitive processing. Recent developments in neural and cognitive engineering proposed completely new application fields of this wellestablished measurement technique when using an advanced singletrial processing. We have recently shown that twodimensional diffusion filtering methods from image processing can be used for the denoising of ERP singletrials in matrix representations, also called ERP images. In contrast to conventional onedimensional transient ERP denoising techniques, the twodimensional restoration of ERP images allows for an integration of regularities over multiple stimulations into the denoising process. Advanced anisotropic image restoration methods may require directional information for the ERP denoising process. This is especially true if there is a lack of a priori knowledge about possible traces in ERP images. However due to the use of event related experimental paradigms, ERP images are characterized by a high degree of selfsimilarity over the individual trials. In this paper, we propose the simple and easy to apply nonlocal means method for ERP image denoising in order to exploit this selfsimilarity rather than focusing on the edgebased extraction of directional information. Using measured and simulated ERP data, we compare our method to conventional approaches in ERP denoising. It is concluded that the selfsimilarity in ERP images can be exploited for singletrial ERP denoising by the proposed approach. This method might be promising for a variety of evoked and eventrelated potential applications, including non stationary paradigms such as changing exogeneous stimulus characteristics or endogenous states during the experiment. As presented, the proposed approach is for the a posteriori denoising of singletrial sequences.