With　the　rapid　development　of　handheld　photographic　devices,　a　large　number　of　unlabeled　images　have　been　uploaded　to　the　Internet.　In　order　to　retrieve　these　images,　image　recognition　techniques　have　become　particularly　important.　As　there　is　often　more　than　one　object　in　a　picture,　multi-label　image　annotation　techniques　are　of　practical　interest.　To　enhance　its　performance　by　fully　exploiting　the　interrelationships　between　labels,　we　propose　a　unified　modular　framework　with　deep　graph　convolutional　networks　(MDGCN).　It　consists　of　two　modules　for　extracting　image　features　and　label　semantic　respectively,　after　which　the　features　are　fused　to　obtain　the　final　recognition　results.　With　classical　multi-label　soft-margin　loss,　our　model　can　be　trained　in　an　end-to-end　schema.　It　is　important　to　note　that　a　deep　graph　convolutional　network　is　used　in　our　framework　to　learn　semantic　associations.　Moreover,　a　special　normalization　method　is　employed　to　strengthen　its　own　connection　and　avoid　features　from　disappearing　in　the　deep　graph　network　propagation.　The　results　of　experiments　on　two　multi-label　image　classification　benchmark　datasets　show　that　our　framework　has　advanced　performance　compared　to　the　state-of-the-art　methods.　©　2021,　Springer　Nature　Switzerland　AG.