The　ecological　imbalance　caused　by　environmental　pollution　has　gradually　become　a　thought-provoking　topic.　In　addition,　the　development　of　global　smart　cities　urgently　requires　the　emergence　of　a　powerful　smart　waste　management　system　which　can　realize　trash　segregation.　Therefore　we　focus　on　the　this　issue　and　proposed　a　novel　method　for　trash　classification　based　on　deep　learning　strategy.　In　order　to　realize　the　efficiency　of　real-time　outdoor　trash　detection　by　robot,　we　adopt　a　lightweight　GhostNet　as　backbone　of　detection　network.　The　network　was　trained　in　our　self-made　dataset　which　includes　4　kinds　of　object(namely　:　plastic　bottle,　glass　bottle,　metal　can,　carton).　Moreover,　for　comparison　and　evaluation,　we　set　up　a　comparative　experiment.　The　experimental　results　show　that　the　improved　version　of　the　YOLOv4　method　proposed　has　improved　detection　performance　compared　to　the　original　YOLOV4　algorithm,　and　has　produced　satisfactory　Generalization　performance　in　different　types　of　trash.　©　2021　IEEE.