Semantic-based　image　retrieval　is　the　desired　target　of　Content-based　image　retrieval　(CBIR).　In　this　paper,　we　proposed　a　new　method　to　extract　semantic　information　for　CBIR　using　the　relevance　feedback　results.　Firstly　it　is　assumed　that　positive　and　negative　examples　in　relevant　feedback　are　containing　semantic　content　added　by　users.　Then　image　internal　semantic　model　(IISM)　is　proposed　to　represent　comprehensive　pair-wise　correlation　information　for　images　through　analyzing　the　feedback　results.　Finally,　correlation　learning　method　is　proposed　to　represent　the　images＇　pair-wise　relationship　based　on　statistical　value　of　access　path,　access　frequency,　similarity　factor　and　correlation　factor.　Experimental　results　on　Corel　datasets　show　the　effectiveness　of　the　proposed　model　and　method.