Efficient　detection　of　smoke　plays　a　critical　role　in　preventing　and　suppressing　fires.　However,　smoke　is　generally　of　variable　shapes　and　colors,　blurred　borders,　and　irregular　textures,　which　makes　smoke　detection　based　on　deep　learning　a　challenging　task.　Aiming　at　this　problem,　a　smoke　detection　adaptive　deep　model　named　DB-YOLO　is　proposed.　In　the　model,　Spatial　attention-based　Dynamic　Convolution　kernel　(SDConv)　is　designed　and　embedded　as　a　feature　extraction　block　to　improve　the　ability　of　extracting　representative　features　from　images　of　diverse　textures.　Besides,　an　improved　Bi-directional　Feature　Pyramid　Network　(BiFPN)　is　integrated　as　a　feature　fusion　block　to　fuse　multi-scale　features.　Results　show　that　mAP@0.5　of　the　DB-YOLO　can　increase　by　6.08%　to　14.40%　in　smoke　dataset　compared　to　currently　popular　object　detection　models.　©　2023　IEEE.