U-Net　is　widely　lightened　to　achieve　fast　semantic　segmentation　of　medical　images　and　remote　sensing　images.　The　cheap　operation　of　the　GhostNet　series　provides　new　lightening　ideas.　Currently,　the　GhostNet　series　has　three　versions,　GhostNet,　G-GhostNet,　and　GhostNetV2,　each　of　which　has　its　own　advantages.　In　this　paper,　we　realize　the　lightweight　improvement　of　U-Net　with　the　help　of　the　GhostNet　series　to　provide　a　reference　for　the　lightweight　research　of　U-Net.　First,　the　backbone　parts　of　GhostNet,　G-GhostNet,　and　GhostNetV2　were　used　as　the　encoder　embedding　model　of　U-Net　in　turn,　and　the　convolutional　operation　of　U-Net　is　replaced　by　their　unitary　module.　Second,　the　three　models　were　trained　and　tested　with　medical　public　dataset　and　remote　sensing　image　dataset　to　obtain　various　performance　scores　of　the　models.　Finally,　the　scores　were　compared.　The　experimental　results　show　that　the　G-GhostNet-based　lightweight　model　is　the　most　efficient　in　the　remote　sensing　image　segmentation　task　and　achieves　real-time　segmentation.　While　in　the　medical　imaging　segmentation　task,　the　GhostNet-based　lightweight　model　is　relatively　faster　and　the　segmentation　accuracy　is　much　higher　than　that　of　the　G-GhostNet-based　model,　and　the　model　is　more　efficient.　GhostNetV2　is　not　dominant　in　both　segmentation　tasks.　　©　2023　IEEE.