Due　to　the　significant　intra-class　differences　and　inter-class　similarities　inherent　in　facial　expressions,　distinguishing　certain　highly　similar　facial　expressions　such　as　disgust,　fear,　and　sadness　has　become　one　of　the　main　challenges　in　facial　expression　recognition　(FER).To　address　this　challenge,　we　propose　a　novel　Rejected　Domain　Attention　Network　(RDANet),　specifically　designed　to　distinguish　these　easily　confusing　facial　expressions.　We　aggregate　and　embed　the　correlations　of　regional　features　and　facial　expressions　generated　by　convolutional　neural　networks　into　a　compact　interval.　Adaptive　weight　adjustment　is　performed　on　the　features　in　the　interval,　and　a　strict　penalty　mechanism　is　added.　We　identify　anti　correlated　regions　by　correlation　scores　and　classify　them　as　rejected　domains　and　the　rest　as　supported　domains.　We　select　important　rejected　feature　elements　via　RDANet　and　give　them　higher　weight　than　supported　domain　features.　An　optimized　contrastive　loss　is　used　to　achieve　a　dynamic　balance　between　intra-class　clustering　and　inter-class　separation,　combined　with　regularization　constraints　to　make　the　model　converge　faster.　Through　experimental　verification,　our　method　has　superior　performance　in　recognizing　confused　facial　expressions.