To　quickly　grasp　what　interesting　topics　are　happening　on　web,　it　is　challenge　to　discover　and　describe　topics　from　User-Generated　Content　(UGC)　data.　Describing　topics　by　probable　keywords　and　prototype　images　is　an　efficient　human-machine　interaction　to　help　person　quickly　grasp　a　topic.　However,　except　for　the　challenges　from　web　topic　detection,　mining　the　multi-media　description　is　a　challenge　task　that　the　conventional　approaches　can　barely　handle:　(1)　noises　from　non-informative　short　texts　or　images　due　to　less-constrained　UGC;　and　(2)　even　for　these　informative　images,　the　gaps　between　visual　concepts　and　social　ones.　This　paper　addresses　above　challenges　from　the　perspective　of　background　similarity　remove,　and　proposes　a　two-step　approach　to　mining　the　multi-media　description　from　noisy　data.　First,　we　utilize　a　devcovolution　model　to　strip　the　similarities　among　non-informative　words/images　during　web　topic　detection.　Second,　the　background-removed　similarities　are　reconstructed　to　identify　the　probable　keywords　and　prototype　images　during　topic　description.　By　removing　background　similarities,　we　can　generate　coherent　and　informative　multi-media　description　for　a　topic.　Experiments　show　that　the　proposed　method　produces　a　high　quality　description　on　two　public　datasets.　(C)　2017　Elsevier　B.V.　All　rights　reserved.