Facial　expression　recognition　plays　an　important　role　in　humancomputer　interaction.　At　present,　the　traditional　algorithm　applied　to　facial　expression　recognition　is　not　accurate　enough,　and　deep　learning　methods　have　problems　such　as　a　large　number　of　network　model　parameters,　insufficient　generalization　ability,　and　high　requirements　on　hardware　devices　in　actual　deployment.　To　solve　these　problems,　a　lightweight　facial　expression　recognition　model　based　on　multiscale　feature　fusion　and　attention　mechanism　is　proposed　in　this　paper.　Firstly,　a　multi-scale　deeply　separable　densely　connected　convolutional　neural　network　is　constructed,　which　can　greatly　reduce　the　number　of　parameters　and　computation.　The　convolution　kernel　of　different　scales　can　obtain　the　receptive　fields　of　different　scales,　and　finally　integrate　the　features　of　multiple　scales,　which　is　beneficial　to　the　learning　of　the　features　of　different　scales.　A　dense　connection　reduces　gradient　disappearance　and　improves　model　generalization　performance.　Secondly,　group　convolution　is　introduced　to　further　reduce　the　number　of　parameters.　Finally,　the　attentionmechanism　and　multi-scale　were　integrated　to　improve　the　classification　accuracy　of　the　model.　The　network　model　proposed　in　this　paper　has　been　trained　and　verified　on　RAF-DB,　FER2013,　CK+　and　FERPlus　datasets.　Experimental　results　showthat　comparedwith　other　classical　convolutional　neural　networks,　the　model　proposed　in　this　paper　can　significantly　reduce　the　number　of　parameters　and　calculation　amount　while　ensuring　high　accuracy.　In　the　future,　it　is　of　great　significance　for　the　practical　application　of　facial　expression　recognition　in　medical　monitoring,　education　assistance,　traffic　warning　and　other　fields.