Double-bags Of Words Model Based Image Retrieval System

The main work of this paper is a double-bags of wordsmodel based image retrieval system.Double-bags of wordsmodel is the improvement of traditional bow(bag of words)feature. Traditional bow feature is a very effective image globalfeature.But it has some defects,the first one is that the clusteringalgorithm it used is unstable,the second one is that theinefficiency cased by oversize clustering dimensions,the thirddefect is that the representation of bow is weak.The model ourpaper proposes can solve these defects.Moreover we propose astrategy of the retrieval of image feature,which can improvethe efficiency of the feature matching greatly without reducingthe accuracy.Double-bags of words model means we can get two bagsabout image feature by clustering all of the image feature in theimage library twice.Then two image features are very similarwhen these two features drop in the same bags at two clustering.The building of Double-bags of words model uses the instabilityof clustering algorithm,meanwhile we can improve therepresentation of feature greatly because two bags willsupplement each other and increase retrieval efficiency throughdecreasing clustering dimensions.Accuracy of image retrieval will be improved because of more rigorous Requirements offeature matching.At the phase of retrieval,because of the huge number ofdata,it is time-consuming if uses traversal,even using kd-tree.Sowe suggest to cluster the image global features first,then find themost similar feature to target feature in the cluster that whosecenter is the most similar to target feature among the all clustercenters.With this method,efficiency will increase greatly.