The　prototype　selection　plays　critical　roles　in　synergetic　pattern　recognition　(SPR).　K-means　clustering　is　widely　adopted　to　determine　appropriate　prototypes　in　SPR.　However,　the　selection　of　initial　cluster　centers　significantly　affects　clustering　results.　We　propose　an　improved　k-means　clustering　to　handle　this　challenge.　According　to　inner-class　distances　among　samples　within　the　same　cluster,　we　will　dynamically　adjust　interclass　distances　among　clusters.　Initial　cluster　centers　will　then　be　highly　representative　in　that　they　are　distributed　among　as　many　samples　as　possible.　Consequently,　local　optima　that　are　common　in　k-means　clustering　can　be　effectively　reduced.　After　we　obtain　final　cluster　centers　output　from　the　improved　k-means　clustering,　we　then　use　these　centers　as　the　prototype　vector　to　train　a　synergetic　neural　network　(SNN),　which　will　be　utilized　to　recognize　human　face　expressions.　Experimental　results　demonstrate　that　our　algorithm　greatly　improves　the　accuracy　in　recognizing　face　expressions　and,　in　a　more　efficient　manner.