Smoke　detection　is　the　key　to　industrial　safety　warnings　and　fire　prevention,　such　as　flare　smoke　detection　in　chemical　plants　and　forest　fire　warning.　Due　to　the　complex　changes　in　smoke　color,　texture　and　shape,　it　is　difficult　to　identify　the　smoke　in　the　image.　Recently,　more　and　more　scholars　have　paid　attention　to　the　research　of　smoke　detection.　In　order　to　solve　the　above　problems,　we　propose　a　convolutional　neural　network　structure　designed　for　smoke　characteristics.　The　characteristics　of　smoke　are　only　complicated　in　simple　features,　and　no　deep　semantic　structure　information　needs　to　be　extracted.　Therefore,　there　is　no　performance　improvement　in　deepening　the　depth　of　the　network.　We　use　a　10-layer　convolutional　neural　network　to　hop　the　features　of　the　first　layer　of　convolution　extraction　to　the　back　layer　to　increase　the　network＇s　ability　to　extract　simple　features.　The　experimental　results　show　that　our　convolutional　neural　network　model　has　fewer　parameters　than　the　existing　deep　learning　method,　and　the　accuracy　rate　in　the　smoke　database　is　optimal.