Segmenting　the　tongue　body　is　an　essential　step　for　automated　tongue　diagnosis,　which　is　a　challenge　task　due　to　the　tongue　body’s　specificity　and　heterogeneity.　The　current　deep-learning　based　tongue　image　segmentation　networks　are　bloated　with　high　computational　complexity.　In　this　study,　a　light-weight　segmentation　network　for　tongue　images　is　proposed　under　the　basic　encoder-decoder　framework,　in　which　MobileNet　v2　is　adopted　as　the　backbone　network,　due　to　its　few　parameters　and　low　computational　complexity.　The　high-level　semantic　information　and　low-level　positional　information　are　combined　together　to　detect　the　tongue　body’s　boundary.　And　the　dilated　convolution　operations　are　performed　on　the　final　feature　maps　of　the　network　to　enlarge　the　receptive　field,　so　as　to　capture　rich　global　semantic　information.　An　attention　mechanism　is　embedded　to　re-calibrate　the　feature　maps　spatially　and　channel-wise　to　enhance　important　features　for　the　segmentation　task,　while　suppressing　the　irrelevant　ones.　Moreover,　a　supervision　output　is　added　to　each　level　of　the　decoder　to　guide　the　network　to　capture　both　the　local　and　global　image　features　for　accurate　tongue　image　segmentation.　All　supervision　outputs　are　fused　to　produce　good　segmented　results.　The　quantitative　and　qualitative　results　on　two　tongue　datasets　indicate　that　the　proposed　network　can　achieve　a　competitive　performance　with　smaller　model　size　and　lower　computational　cost.　The　proposed　method　could　accurately　extract　the　tongue　body,　which　can　fully　meet　the　requirements　of　practical　applications.　©　2022,　The　Author(s),　under　exclusive　licence　to　Springer　Science+Business　Media,　LLC,　part　of　Springer　Nature.