An　improved　common　special　pattern　(CSP)　algorithm　was　proposed　to　improve　classification　accuracy　of　motor　imagery　electroencephalograph　(EEG)　in　a　brain-computer　interface　(BCI)　system.　Three　channels　EEGs　were　filtered　through　the　time　window　and　band-pass　filter　to　obtain　the　most　obvious　features　of　event-related　desynchronization　and　event-related　synchronization.　Improved　CSP　algorithm　combined　with　support　vector　machine　(SVM)　as　adopted　for　the　classification　of　motor　imagery　EEGs.　The　experiment　results　show　that　the　improved　CSP　algorithm　can　avoid　the　repetitive　eigenvector　selection　and　discriminate　the　left　hand　and　right　hand　mental　task　more　accurately　than　the　traditional　CSP.　©　2009　IEEE.