Traditional　canonical　correlation　analysis　as　an　effective　feature　extraction　method　has　been　widely　used　in　many　areas　of　pattern　recognition,　including　facial　expression　recognition.　The　canonical　correlation　analysis　makes　the　within-class　covariance　matrix　constituted　of　two　groups　of　feature　vectors　to　be　singular,　resulting　in　small　sample　size　problem,　meanwhile　the　canonical　correlation　analysis　as　a　global　linear　projection　method,　not　a　good　description　of　nonlinear　facial　expression　recognition　problem.　In　order　to　solve　the　above　problems,　we　present　a　facial　expression　recognition　algorithm　based　on　modular　two　dimensional　canonical　correlation　analysis.　The　algorithm　uses　a　modular　method,　and　better　takes　advantage　of　local　non-linear　characteristics　of　the　facial　expression　image,　meanwhile　the　direct　use　of　two-dimensional　facial　expression　feature　matrix　is　effective　to　relieve　the　small　sample　size　problem.　In　recognition　stage,　we　use　the　voting　method,　which　can　improve　the　robustness　of　local　expression　feature　changes.　Experiments　on　JAFFE　facial　expression　database　show　that　the　proposed　method　is　robust　and　effective.　©　2013　by　Binary　Information　Press.