Flame　detection　is　a　key　technical　link　to　realize　intelligent　forest　fire　prevention　and　control.　However,　the　current　fire　detection　methods　generally　have　the　problems　of　low　detection　rate,　high　false　alarm　rate,　and　poor　real-time　performance.　In　order　to　achieve　rapid　and　accurate　recognition　of　forest　fires　in　natural　environments,　this　paper　proposes　a　lightweight　convolutional　neural　network　flame　detection　algorithm　Yolo-Edge.　MobileNetv3　has　deep　separable　convolutional　structure　features,　which　can　replace　Yolov4＇s　original　CSPDarknet53　feature　extraction　backbone　network,　and　can　reduce　the　number　of　network　layers　and　model　size,　so　that　it　can　adapt　to　the　working　environment　of　edge　devices　and　multi-scale　prediction.　Feature　fusion　is　carried　out　through　the　feature　pyramid　to　improve　the　detection　accuracy　of　small　targets.　Use　2059　flame　images　in　different　occlusion　environments　as　a　data　set　for　training　and　testing,　and　use　F1　value　and　AP　value　to　evaluate　the　difference　of　each　model.　The　test　results　show　that　the　lightweight　improved　neural　network　model　proposed　in　this　paper　has　good　recognition　accuracy　and　speed,　which　significantly　reduces　the　memory　usage　of　the　model　and　achieves　a　good　lightweight　effect.