A　fire　detection　method　based　on　improved　PP-YOLO　is　proposed　to　promote　the　performance　of　flame　detection.　Specifically,　based　on　PP-YOLO,　a　feature　fusion　network　is　introduced　to　fuse　the　two　adjacent　output　feature　maps　of　backbone　so　that　the　high-level　features　can　better　fuse　the　details　of　low-level　features.　Then,　an　attention　module　is　employed　in　the　intermediate　fusion　feature　map　of　two　adjacent　outputs,　which　lets　the　network　selectively　fuse　the　valuable　information　in　the　feature　maps　in　a　self-learning　manner　to　alleviate　dilution　and　aliasing　effect　of　information　during　feature　fusion.　Finally,　combined　with　different　training　tricks,　such　as　data　augments　and　learning　rate　adjustment　strategy,　the　model　is　trained　and　tested　on　two　testing　sets.　The　experiment　results　demonstrate　that　the　improved　model　can　achieve　86.87%　and　85.66%　mean　average　precision　(mAP)　on　two　testing　sets.　Additionally,　the　precision　of　improved　model　is　97.23%　and　93.39%,　and　the　false　alarm　rate　can　achieve　0.83%　and　1.68%,　respectively.　And　the　average　detection　time　is　25.62　ms.　In　conclusion,　the　model　is　suitable　for　various　fire　scenarios　and　can　be　well　utilized　in　actual　conditions.