Pole　sign,　recording　the　basic　information　of　transmission　lines,　is　the　main　reference　for　the　power　supply　department　to　manage　the　grid　lines.　The　traditional　ways　of　collecting　pole　signs　are　quite　manual.　With　the　development　of　computer　vision　and　deep　learning　technology,　it　raises　interest　on　automatic　collection　and　recognition　of　sign　information.　However,　there　is　much　work　to　make　such　method　applicable,　due　to　the　large　size　of　current　sign　detection　model,　low　speed　and　efficiency　of　signal　detection.　Therefore,　this　paper　proposes　a　lightweight　text　detection　model　named　as　Tiny-DBNet　based　on　a　lightweight　feature　extraction　unit　with　a　deep　separable　convolution　residual　block　and　attention　mechanism　and　then　uses　this　unit　to　build　a　feature　extraction　network.　Then　the　feature　map　is　input　into　the　DBNet-based　differentiable　binarization　segmentation　module,　and　the　text　detection　result　of　the　sign　is　output.　The　results　of　the　experiments　are　obtained　on　real　Pole　Sign　data　and　ICDAR2015　public　data　set,　and　show　that　the　detection　speed　is　3　times　faster　and　the　parameter　size　is　reduced　by　45.15%　under　the　small　loss　of　0.6%　precision,　compared　with　the　original　model　of　DBNet　based　on　Resnet_50.　According　to　the　results,　the　proposed　model　is　lightweight　and　efficient　and　is　a　good　solution　for　cylindrical　pole　identification.　©　2022,　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.