| As image analysis becomes widely used in the police system, digital library,medical treatment and diagnosis, knowledge copyright and satellite remote sensingetc, the size of the image database is becoming larger continuously, so it has beenfocused on how to govern, search and inquire these images better.The choice of image feature is vital to the performance and effectiveness of theimage retrieval technology. As color, texture and shape are the basic features of theimage content, in this paper, we retrieve the images according to these three features.In the experiment, we choose five feature-extracting methods as the color histogram,color correlogram, texture co-occurrence matrix, Tamura texture and Hu momentinvariants, to perform the single feature retrieval to the10kinds of images in thedatabase, and then statistic the precision ratio of each feature to each kind of theimages by using the precision ratio as the evaluation criteria.As the performance of the image retrieval based on the single feature is finite, inorder to get better result in different kinds of images, we propose a retrieval methodbased on the weighted multi-features and another method based on the multi-featureswith the improved DS theory in this paper. The retrieval method based on theweighted multi-features is according to the precision ratio of single feature, in thispaper; we use the precision ratio of each type of features as the weight basis to eachkind of images, and then perform retrieval mixed by the weighted multi-features, andcompare the result with the retrieval mixed by equal ratio multi-features. And then,we introduce the DS theory on the basis of the above method, and research theprinciple and process of the DS theory multi-features retrieval. We choose theCorel-test image database, according to the DS merging principle, and remove thefeature with the lowest precision ratio of the single feature, then calculate thecombination trust degree, and finally calculate the feature similarity according to thetrust degree. As the results show that, on some degree, the image retrieval based on the multi-features DS theory can make up the shortage of image retrieval based onthe weighted multi-features on some kinds of images.After all above jobs, we research the current semantic retrieval technology afterthe analysis, and choose the sophisticated SVM classification theory, by choosing apart of images as the training set to do the training classification, considering thepriority of different features mentioned in previous chapter, we propose the imageretrieval technology using SVM and multi-features, to obtain better retrievaleffectiveness. |
