To　improve　the　efficiency　of　image　relevance　feedback　algorithm　rapidly,　an　algorithm　of　auto-adapted　weight　revision　combining　with　support　vector　machine　is　proposed.　In　early　retrieval　stage,　the　weight　coefficients　of　different　features　are　adjusted　quickly　by　auto-adapted　weight　revision　algorithm,　using　quick　deletion　strategy　of　negative　samples　to　improve　the　accuracy　of　early　retrieval　stage,　which　providing　more　positive　samples　for　the　SVM　models　in　later　retrieval　stage;　In　later　retrieval　period,　retrieval　models　are　designed　by　SVM　models,　and　they　are　optimized　by　the　algorithm　of　active　learning　and　semi-supervision　relevance　feedback.　Experiment　results　on　5000　Corel　images　database　indicate　that　this　algorithm　can　obviously　improve　the　efficiency　and　performance　of　learning　machine　and　accelerate　the　convergence　to　user＇s　inquiry　concept.　©　2013　IEEE.