This　paper　proposes　an　approach　of　face　recognition　using　composite　features.　This　approach　first　aligned　3D　faces　based　on　non-uniform　mesh　re-sampling　which　result　the　uniform　number　of　vertex　and　topology　structure　and　kept　in　a　vector.　Second,　we　changed　the　vector　to　the　matrix　by　division　and　overlap,　and　then　2DLDA　suggests　a　feature　selection　strategy　to　select　the　most　discriminative　features　from　the　corner.　This　approach　not　only　avoided　face　information　missing,　increased　the　number　of　composite　feature,　but　also　avoided　the　small　sample　size　problem　in　theory.　Experimental　results　for　3D　face　data　set　from　BJUT-3D　face　database　have　demonstrated　the　performance　of　our　algorithm.