The　interference　of　natural　light　in　intelligent　manufacturing　will　affect　the　classification　and　positioning　accuracy　of　robot　vision　system.　In　order　to　improve　the　robustness　of　vision　system,　this　paper　first　uses　homomorphic　filtering　based　on　illumination　reflection　model　to　improve　the　problem　of　uneven　brightness　of　image,　and　then　uses　the　fast-adaptive　threshold　segmentation　method　based　on　integral　graph　to　binarize　the　image.　Then　a　variety　of　shape　feature　descriptors　are　used　as　features,　PCA　is　used　to　analyze　the　main　components,　and　then　input　them　into　a　BP　neural　network　for　shape　recognition　and　classification.　The　experimental　results　show　that　the　method　is　robust　and　real-time,　which　can　also　meet　the　actual　needs　of　the　factory.　©　2020　IEEE.