For　the　network　environment　with　the　limited　transmission　capacity,　a　multi-nodes　image　retrieval　method　based　on　visual　words　is　proposed.　Firstly,　the　visual　words　of　query　image　are　built　by　using　the　K-means　clustering　method　after　the　color　features　and　SIFT　features　of　query　image　are　extracted.　Then　the　visual-words　histogram　of　the　query　image　is　carried　by　the　mobile　Agent.　The　image　similarity　in　each　node　is　measured　with　the　Euclidean　distance.　The　results　of　image　retrieval　are　returned　to　the　user.　Finally　the　weight-based　relevance　feedback　method　is　utilized　to　optimize　the　retrieved　results.　Experimental　results　show　that　the　proposed　method　can　improve　the　speed　and　the　accuracy　of　the　image　retrieval.