In　this　paper,　we　propose　a　novel　image　annotation　algorithm　based　on　Markov　model.　This　algorithm　treats　each　candidate　keyword　as　a　state　in　Markov　chain,　and　implements　image　annotation　by　estimating　the　probability　of　Markov　transition.　On　one　aspect,　compared　with　classical　algorithms,　the　proposed　algorithm　stops　making　the　assumption　that　each　keyword　is　independent　to　each　other;　instead,　they　are　related　to　the　existed　keywords;　On　the　other　hand,　not　only　considering　correlation　between　keywords　that　improves　results　of　image　annotation,　but　our　proposed　approach　also　takes　image　visual　content　into　account.　Experimental　results　on　the　typical　Corel　dataset　demonstrate　the　effectiveness　and　the　increasing　annotation　precision　of　our　proposed　algorithm.　©　2010　IEEE.