A　image　annotation　algorithm　based　on　Cross　Media　Relevance　Model　was　proposed　to　bridge　the　semantic　gap　of　content-based　image　retrieval.　The　algorithm　reduced　the　word　bias　observed　in　probabilistic　models　by　converting　the　word-image　joint　probability　to　image　probability　conditioned　on　annotation　words　and　estimated　the　probability　of　a　set　of　annotation　words　by　measuring　the　semantic　similarities　of　each　annotation　word　to　all　other　word.　It　used　a　contextual　term　vector　to　represent　a　annotation　word,　and　implemented　image　annotation　by　estimating　maximum　correlation　between　an　image　and　a　set　of　annotation　words.　Compared　with　image　annotation　algorithm　based　on　Cross　Media　Relevance　Model,　the　proposed　algorithm　stopped　making　the　assumption　that　each　keyword　was　independent　to　each　other,　instead,　took　contextual　relevance　information　of　annotation　words　into　account.　Experimental　results　on　the　typical　corel　dataset　demonstrated　the　effectiveness　and　the　increasing　annotation　precision　of　the　proposed　algorithm.