Non-dominated　sorting　genetic　algorithm　II　(NSGA-II)　works　effectively　in　selecting　bi-objective　features,　but　may　result　in　local　convergence　and　prematurity　in　the　process　of　optimization.　In　order　to　solve　this　problem,　an　improved　NSGA-II　feature　selection　algorithm　is　proposed.　In　this　algorithm,　firstly,　the　first　elite　strategy　is　operated　to　select　the　elite　population　from　parent　population.　Secondly,　the　selected　parent　elite　population　is　combined　with　the　offspring　population　to　form　a　combined　population.　Finally,　the　second　elite　strategy　is　executed　to　obtain　the　next　parent　population.　After　the　selection　of　3D　face　expression　candidate　features,　the　selected　features　are　classified　by　means　of　probabilistic　neural　network.　Experimental　results　show　that　the　proposed　algorithm　improves　the　performance　of　NSGA-II　with　local　convergence　and　prematurity　problems　greatly　and　increases　the　accuracy　of　facial　expression　recognition　effectively.　©,　2015,　South　China　University　of　Technology.　All　right　reserved.