A　high　resolution　time-frequency　analysis　method　to　discriminate　among　mental　tasks　for　the　brain-computer　interface　(BCI)　was　presented.　It　estimated　the　time-frequency　energy　of　electroencephalogram　(EEG)　with　Matching　pursuit　and　Wigner-Ville　distribution.　Matching　pursuit　decomposed　Event-Related　EEG　into　a　linear　expansion　of　waveforms.　A　time-frequency　energy　distribution　was　derived　by　adding　the　Wigner-Ville　distribution　of　each　selected　atoms.　With　the　prior　knowledge,　the　approach　obtained　the　ratios　of　energy　value　over　the　motor-cortex　as　extracted　features.　The　classifying　time　segment　was　determined　according　to　Mahalanobis　distance,　and　the　feature　vectors　in　this　segment　were　used　to　train　an　artificial　neural　network　(ANN)　as　a　classifier.　The　results　showed　that　the　minimum　error　rates　for　4　subjects　were　all　lower　than　15%,　and　the　maximum　mutual　information　were　more　than　0.6.　The　proposed　method　gives　generally　applicable,　robust　and　high-resolution　estimates.