Research Of Relevance Feedback Technology Of Image Retrieval Based On SVM

With the continually development of Multimedia database,traditional image retrieval method which is based on keywords can not satisfy most of the requirement of image retrieval.In recent years, more and more researchers have transferred their focus of scientific researches on content based image retrieval.In this paper,the author firstly introduces and analyzes the recent researches on CBIR,the color histograms and Gabor transform based texture feature are choosed after analyzing and comparing several color and texture feature.An image retrieval algorithm based on region-of-interest is proposed in this paper.This algorithm firstly decomposes all the images in image database into sub-images by multiresolution tree, and then the feature of the ROI selected by the user is extracted to construct the feature database; While retrieving ,the user chooses ROI of any scale from sample image ,then extracts the feature of ROI for similarity measure with feature database.This algorithm is practicable,and covers those two shortages of the other two retrieval method-image retrieval using the feature of the entire image ,and segmented regions-based retrieval.The former neglects the objects in which the user is interested, while the result of the latter excessively depends on complicated segmentation algorithm.To make up for the semantic gap ,the main research work of this paper is to study the application of SVM in retrieval based ROI. As a result of the study,a semantic model of image retrieval which is based on SVM relevance feedback is built so that the retrieval can be improved.The author makes a prototype with MATLAB 6.5 in the course of study.Retrieval by color histogram, texture feature based Gabor transform or dual feature can be done on this prototype, also three relevance feedback algorithm-MARS,MARS+ and SVM ,the parameters of which can be set by the user, can be done. The conclusion that the SVM based feedback algorithm can achieve higher average rate of recall than MARS or MARS+ ,is found by comparison.Finally,the best parameters of kernel function is figured out by analyzing the plot of average recall rate.