It　is　necessary　to　improve　the　effect　of　rehabilitation　for　people　with　dyskinesia　by　realizing　the　consistency　between　motor　intention　and　practical　rehabilitation　exercise　based　on　brain-computer　interface　technology.　It　is　difficult　to　recognize　the　electroencephalogram　(EEG)　which　is　performed　by　the　same　body　part　under　different　motor　imagery　tasks　and　rewritten　as　EEGs.　Meanwhile,　it　is　important　to　research　about　feature　extraction　method.　In　this　paper,　the　feature　extraction　method　is　presented　about　EEGs　under　flexion　and　extension　imagery　tasks　of　index　finger.　An　adaptive　optimal　frequency　band　extraction　method　based　on　wavelet　packet　decomposition　and　entropy　criterion　is　proposed　because　of　the　characteristics　of　EEGs　including　its　weaker　phenomenon　of　event-related　synchronization　(ERS)　and　larger　individual　differences　of　time　and　frequency　band　where　ERS　appears.　Firstly,　EEGs　is　decomposed　by　wavelet　packet　analysis.　Then,　entropy　criterion　is　selected　to　measure　the　separability　of　the　characteristic　frequency　bands.　Furthermore,　the　feature　vector　is　constructed　with　the　wavelet　packet　coefficients　corresponding　to　some　obvious　wavelet　packets.　Lastly,　the　optimal　band　is　obtained　by　combining　with　support　vector　machine.　The　EEGs　of　is　analysed　for　10　subjects,　and　experimental　results　show　that　the　proposed　method　can　choose　the　feature　bands　with　larger　difference　in　ERS　phenomenon　of　EEGs　adaptively,　and　the　highest　classification　accuracy　is　81.75%.　So,　the　correctness　and　validity　of　the　proposed　method　is　proved.