In　the　traditional　image　annotation　methods,　the　manual　selection　of　features　is　time-consuming　and　laborious.　In　the　traditional　label　propagation　algorithm,　semantic　neighbors　are　ignored.　Consequently　visual　similarity　and　semantic　dissimilarity　are　caused　and　annotation　results　are　affected.　To　solve　these　problems,　an　automatic　image　annotation　method　combining　semantic　neighbors　and　deep　features　is　proposed.　Firstly,　a　unified　and　adaptive　depth　feature　extraction　framework　is　constructed　based　on　deep　convolutional　neural　network.　Then,　the　training　set　is　divided　into　semantic　groups　and　the　neighborhood　image　sets　of　the　unannotated　images　are　set　up.　Finally,　according　to　the　visual　distance,　the　contribution　value　of　each　label　of　the　neighborhood　images　is　calculated　and　the　keywords　are　obtained　by　sorting　their　contribution　values.　Experiments　on　benchmark　datasets　show　that　compared　with　the　traditional　synthetic　features,　the　proposed　deep　feature　possesses　lower　dimension　and　better　effect.　The　problem　of　visual　similarity　and　semantic　dissimilarity　in　visual　nearest　neighbor　annotation　method　is　improved,　and　the　algorithm　effectively　enhances　the　accuracy　and　the　number　of　accurate　predicted　tags.　©　2017,　Science　Press.　All　right　reserved.