Floating　garbage　on　sea　surface　has　always　been　a　key　issue　in　the　long-term　research　of　environmental　pollution.　In　order　to　effectively　solve　the　problem　of　marine　garbage　pollution,　this　paper　conducts　in-depth　research　on　the　existing　VGG16　network　model　and　proposes　an　improved　lightweight　VGG　network　model.　Instead　of　the　fully　connected　layer,　our　model　uses　the　global　average　pooling　layer　to　reduce　the　number　of　network　parameters,　and　adds　a　residual　module　to　the　convolution　module　to　improve　the　accuracy　of　the　model.　The　experimental　results　show　that　the　accuracy　of　the　improved　lightweight　VGG　network　model　is　as　high　as　97.8%　in　the　self-built　sea　surface　waste　data　set.　Compared　with　the　traditional　VGG16　network　model,　the　number　of　parameters　is　reduced　by　98.5%　and　the　calculation　amount　is　reduced　by　78.5%,　achieving　the　goal　of　rapid　and　accurate　classification.　©　2023　SPIE.