To　solve　the　problem　that　when　the　dimension　of　variables　was　high,　canonical　correlation　analysis　couldn＇t　give　a　stable　model　of　the　problem,　a　facial　expression　recognition　method　based　on　adaptive　weights　sparse　canonical　correlation　analysis　was　proposed.　Sparse　canonical　correlation　analysis　attached　a　constraint　of　coefficient　convergence,　some　of　the　factors　in　the　basis　vectors　converged　to　zero,　therefore　it　would　be　able　to　remove　some　useless　variables　for　the　facial　expression　recognition.　In　the　process　of　solving　sparse　canonical　correlation　analysis,　sparse　weight　was　a　fixed　value,　therefore　the　method　of　adaptive　weights　was　used　to　reduce　the　error　when　solving　the　sparse　canonical　vector.　Results　on　Jaffe　and　Cohn-Kanade　tests　of　facial　expression　database　show　that　the　proposed　method　is　correctness　and　effectiveness.