In　order　to　eliminate　security　risks　and　reduce　the　probability　of　terrorist　incidents,　X-ray　security　scanners　have　been　installed　in　places　such　as　airports,　customs　and　railway　stations　to　check　whether　luggage　packages　contain　prohibited　items.　At　present,　the　security　inspection　is　manually　checked　by　professional　security　personnel.　In　the　face　of　a　large　number　of　objects　to　be　detected,　the　efficiency　of　manual　detection　is　low,　and　the　detection　results　are　susceptible　to　human　experience　and　psychological　interference.　The　design　of　automatic　identification　and　detection　of　prohibited　items　in　X-ray　security　inspection　scenes　can　effectively　improve　the　efficiency　of　security　inspection　and　reduce　the　difficulty　of　potential　risks　and　human　operation.　With　the　development　of　deep　learning　in　recent　years,　computer　vision　technology　can　realize　fast　and　reliable　automatic　detection　of　prohibited　items.　However,　due　to　the　penetrating　imaging　nature　of　X-ray,　the　X-ray　security　images　contain　a　lot　of　occlusion　and　overlap,　which　increases　the　difficulty　of　identifying　prohibited　items.　Based　on　the　SCNet　[1]　algorithm,　this　paper　proposes　an　instance　segmentation　algorithm　multi-level　semantic　feature　fusion　network(MSFNet).　MSFNet　enhances　the　upsampling　ability　of　the　model　to　recover　detailed　information,　and　effectively　fuses　low-order　shallow　detail　information　and　high-order　abstract　semantic　information.　Firstly,　a　dynamic　upsampling　feature　fusion　module(DUFM)　was　proposed　to　capture　a　longer　range　of　context　information　and　reduce　the　interference　of　occlusion　noise.　Secondly,　a　multi-stage　global　semantic　extraction　module(MGSEM)　is　proposed　to　extract　multi-scale　global　semantic　information　from　semantic　features　of　different　granularities.　Experiments　on　the　Pidray　dataset　[2]　show　that　MSFNet　has　further　improvement　in　accuracy　compared　with　common　instance　segmentation　algorithms,　and　the　accuracy　index　in　detection　and　segmentation　tasks　reaches　72.3%and　62.4%　mAP　respectively.　©　2023　IEEE.