Facial　expression　recognition　has　a　wide　range　of　applications　in　various　human-computer　interaction　fields.　In　view　of　the　problems　of　a　large　number　of　network　model　parameters,　insufficient　generalization　ability　in　a　complex　environment,　high　requirements　on　hardware　equipment　and　difficulty　in　the　practical　deployment　of　the　current　deep　learning　models　applied　to　facial　expression　recognition,　this　paper　proposes　a　lightweight　face　expression　recognition　model　based　on　multi-scale　dense　connection　convolutional　neural　network.　First　of　all,　the　model　uses　Densenet＇s　idea　for　reference　in　feature　learning,　and　the　dense　convolutional　layer　realizes　the　reuse　of　feature　maps　so　that　the　network　can　learn　more　original　features.　Secondly,　based　on　the　deep　separable　convolution,　a　multi-scale　deep　separable　convolution　is　proposed　to　expand　the　receptive　field　of　the　convolution　and　enhance　the　richness　of　the　characteristic　channels　of　the　convolution.　Different　scales　of　the　convolution　kernel　can　get　different　scales　of　receptive　fields,　and　eventually　fuse　the　features　of　different　scales,　which　is　conducive　to　model　recognition　of　facial　expression　images　of　different　scales　and　improve　the　classification　accuracy.　At　the　same　time,　deep　separable　convolution　can　further　reduce　the　number　of　parameters　and　the　amount　of　calculation.　Finally,　a　channel　attention　mechanism　is　added　to　improve　the　accuracy　of　the　model.　Experimental　results　show　that,　compared　with　other　classical　convolutional　neural　networks,　the　lightweight　model　proposed　in　this　paper　can　ensure　a　higher　accuracy　while　reducing　the　number　of　network　parameters　and　computation　in　the　facial　expression　recognition　task,　which　is　of　great　significance　for　the　practical　deployment　of　the　model.　©　2021　ACM.