The electroencephalography (EEG) consists in measuring brain electrical activity thanks to electrodes located on the scalp surface. This technique is mainly used for the diagnostic of epilepsy. Some grapho-elements like slow waves and spike waves can appear on the EEG, enabling the neurologist to detect an epilepsy pain.
      Unfortunately, this activity can be highly contaminated by parasitical signals called artifacts. These artifacts have for main origins, the ocular activity, the muscular activity, the cardiac rhythm and tight electrode displacements. The frequencies of pathological grapho-elements recover those of artifacts, and it is then required to use spatial filter which rests on source separation. The principle is to determine a set of cerebral sources and a set of artifacts sources. Artifact sources are then cancelled and the cerebral ones are used to rebuild the signal.
      This thesis presents several methods using both spatial and frequential filters, making the EEG filtering automated. A quantitative approach of filtering validation is defined, which enables the author to choose the most efficient called Adaptive Filtering by Optimal Projection (AFOP). According to the neurologist, tests on clinical recordings of epileptic patients prove AFOP efficiency on cancelling most of artifact types as well as on respecting cerebral rhythms.