Image　automatic　annotation　is　a　significant　and　challenging　problem　in　pattern　recognition　and　computer　vision.　Aiming　at　the　problems　that　the　existing　models　have　low　utilization　and　they　are　affected　by　unbalanced　positive　and　negative　samples,　a　hierarchical　image　annotation　model　is　proposed.　In　the　first　layer,　discriminative　model　is　used　to　assign　topic　annotations　to　unlabeled　images,　and　then　the　corresponding　relevant　image　sets　are　obtained.　In　the　second　layer,　a　keywords-oriented　method　is　proposed　to　establish　links　between　images　and　keywords,　and　then　the　proposed　iterative　algorithm　is　used　to　expand　semantic　words　and　relevant　image　sets.　Finally,　a　generative　model　is　used　to　assign　detailed　annotations　to　unlabeled　images　on　expanded　relevant　image　sets.　Hierarchical　model　uses　less　relevant　training　images　to　obtain　better　annotation　results.　Experimental　results　on　Corel　5K　datasets　verify　the　effectiveness　of　proposed　hierarchical　image　annotation　model.