To　address　the　challenges　of　leakage,　misdetection,　and　low　accuracy　caused　by　environmental　complexity,　object　occlusion,　and　insufficient　lighting　in　smoke　and　fire　detection　within　warehouse　scenarios,　this　paper　proposes　a　cross-scale　multi-branch　feature　fusion　network　for　smoke　and　fire　detection,　named　SF-YOLO.　Firstly,　a　multi-branch　feature　enhancement　module,　Res2Net-C,　is　proposed　to　introduce　an　attention　mechanism　using　a　multi-branch　feature　fusion　network,　which　not　only　extends　the　sensory　field　of　the　network,　but　also　enhances　the　ability　of　extracting　effective　information,　while　effectively　suppressing　the　interference　of　irrelevant　information.　It　effectively　improves　detection　performance　under　complex　environmental　conditions　and　in　the　presence　of　object　occlusion.　Secondly,　a　cross-scale　feature　fusion　network　structure,　MSFPN,　is　proposed,　which　achieves　effective　complementation　and　fusion　of　high-level　and　low-level　information　by　establishing　hopping　connections　between　features　at　different　scales,　and　significantly　improves　the　recognition　accuracy　of　the　type　and　location　of　the　smoke　and　fire　in　low-light　conditions.　In　addition,　the　added　small　target　detection　layer　further　improves　the　network＇s　ability　to　detect　small　targets.　Finally,　in　order　to　better　adapt　to　pyrotechnic　targets　with　variable　morphology,　EIoU　is　adopted　as　the　bounding　box　loss　function.　Test　results　on　a　self-built　dataset　show　that　SF-YOLO　achieves　a　mean　accuracy　rate　(mAP)　of　87.5%　in　the　task　of　detecting　smoke　and　fire　under　multiple　environmental　factors　in　a　warehouse　scenario,　proving　its　effectiveness　and　usefulness.　©　2024　Copyright　held　by　the　owner/author(s).