Computer　vision-based　fire　detection　methods　have　recently　gained　popularity　as　compared　to　traditional　fire　detection　methods　based　on　sensors.　According　to　whether　or　not　use　hand-crafted　features　for　fire　detection,　computer　vision-based　fire　detection　methods　can　be　divided　into　two　categories:　hands-crafted　based　methods　and　deep　learning　based　methods.　However,　because　of　the　limited　representation　of　hand-crafted　features,　the　performance　of　hand-crafted　based　methods　are　limited　by　the　illumination,　quality　and　background　scenes　of　fire　images.　Thus,　in　this　study,　we　propose　an　improved　deep　convolution　network,　which　uses　the　global　average　pooling　layer　instead　of　the　full　connected　layer　to　fuse　the　acquired　depth　features　and　detect　fire.　Besides,　to　further　improve　the　accuracy　of　fire　detection,　we　construct　multi-features　input　data　to　compensate　for　the　insufficiency　of　experimental　data.　Because　there　is　no　common　dataset　for　fire　detection,　we　verify　the　effect　of　our　proposed　method　on　our　collected　dataset　and　get　89.9%　accuracy　for　fire　detection.　©　2019　Association　for　Computing　Machinery.