In　the　field　of　image　scene　classification,　with　the　continuous　innovation　of　DL　(Deep　Learning)　technology,　these　technologies　have　achieved　excellent　results　in　traditional　image　classification　tasks.　This　article　is　based　on　the　CNN　(Convolutional　Neural　Network)　method,　which　extracts　features　from　the　original　image　input　and　equally　processes　the　features　of　each　channel.　This　process　wastes　unnecessary　calculations　to　obtain　rich　low-frequency　features,　lacks　discriminative　learning　ability　across　feature　channels,　and　ultimately　hinders　the　representativeness　of　deep　networks.　Research　has　shown　that　in　the　test　results　of　three　methods　on　different　datasets,　it　can　be　found　that　the　classification　accuracy　of　our　method　is　91.59%,　while　the　accuracy　of　Full　and　Object　are　72.76%　and　81.46%,　respectively.　It　can　be　seen　that　the　classification　accuracy　of　our　method　is　the　highest　among　the　three　methods,　proving　the　effectiveness　of　our　method.　In　addition,　compared　with　other　better　methods,　this　method　is　also　superior　to　other　methods　because　it　extracts　useful　information,　which　is　more　discriminative　and　can　better　classify　image　scenes.　Determine　the　importance　of　features　on　the　channel　based　on　attention　weight,　in　order　to　focus　on　more　representative　features　and　improve　network　classification　performance.　©　2023　IEEE.