This　paper　proposes　an　improved　scheme　of　using　the　inter-cluster　distance　in　the　feature　space　to　choose　the　kernel　parameters.　First,　the　candidate　vectors　of　the　training　set　are　selected.　Then　calculate　the　inter-cluster　distance　between　classes　to　choose　the　proper　kernel　parameters.　Finally　the　selected　kernel　parameters　are　used　to　train　the　Support　Vector　Machine　(SVM)　models.　The　basic　principle　is　that　the　Support　Vector　(SV)　set　contains　all　information　necessary　to　solve　a　given　classification　task.　Experiment　results　show　that　our　scheme　costs　much　less　computation　time.　Moreover,　suitable　kernel　parameters　can　also　be　selected　at　the　same　time.