Multimodal　emotion　recognition　effectively　uses　cross-modal　information　to　enhance　model　performance.　However,　in　practical　applications,　the　missing　modalities　issue　often　degrades　emotion　recognition　accuracy　due　to　the　modality　gap　arising　from　differences　in　information　representation　and　semantic　inconsistencies　across　modalities.　To　address　this　challenge,　this　paper　introduces　a　specific　and　invariant　feature　learning　approach(SIFL).　Specifically,　we　employ　feature　extraction　techniques　using　a　self-attention　mechanism　for　modality-specific　features　and　leverage　a　denoising　autoencoder　for　invariant　feature　extraction　to　enhance　semantic　richness　and　expressiveness.　Additionally,　we　develop　a　reconstruction　network　to　generate　high-quality　modality　features.　To　further　optimize　the　process,　we　design　and　implement　multiple　optimization　objectives,　effectively　bridging　the　semantic　gap　between　modalities.　Experimental　results　on　the　CMU-MOSI　dataset　demonstrate　that　the　proposed　method　surpasses　current　mainstream　baselines　and　exhibits　robust　performance,　particularly　under　conditions　with　high　missing　rates,　validating　its　efficacy　and　versatility　in　handling　missing　modalities.　©　2025　IEEE.