Image Retrieval Research Based On Interest Region And Svm Relevance Feedback

Since the popular ity o f Inter net and the high speed develop ment technolo gy o f digital image processing in t he infor mat ion age, lots of ima ges have come into var ious fie lds of societ y. There comes the problem: how to retrie va l the desired ima ge quickly and accurately as the image database is becoming huge r and huger? Therefore, content based ima ge retrieva l(C BIR) comes into the world and become one of the hot spots in the domest ic and foreign scho lars’ research. Tradit iona l C BIR system extracts features in the whole ima ge, whic h conta ins lots redundant infor mat ion and a large amount of ca lculat ion and undoubtedly bring inaccurate results to the C BIR system. As the above proble ms are taken into consideratio n, this thesis narrows down retr ieva l region fro m the who le image to t he local regio n and focuses on t he ima ge retr ieva l based on interest region and SVM relevance feedback. The specific study works are as follows:Aiming at this weakness of the two methods which based on interest points convex hulls and interest points concentr ic rings of equal interva l both can be easily affected by no ise corner points in the ima ge background when extract ing image region of interest and w ill reduce t he accuracy of ima ge retrie va l, an image retr ieva l met hod based on t he interest regio n is proposed in this paper. F irst, ima ge interest points is detected; Then, densit y ratio of interest points in each distr ib ute regio n is calculated, the regio n w ith lower interest points densit y ratio is abandoned and the circular doma in that t he rema ining interest points are clustered in is regarded as the image regio n of interest. Fina lly, mult i- features of color, text ure and shape whic h all located in the regio n of interest are fused for the init ia l quer y, Simulat io n results show that, the proposed met hod can achieve more accurate image region of interest extract ion and higher image retrieval evaluating index.To ascertain the user ’s fur ther retrie val intent io n and narrow down the gap between t he high- le ve l sema nt ic features and the under lying vis ua l features of ima ge, releva nce feedback based on support vector machine(SVM) class ificat ion is introduced into the content based ima ge retrieva l in t his thes is. As the drawback when there are lack of a certain quant it y of posit ive samp les, the opt ima l hyperplane of SVM class ificat ion w ill ha ve a deviat ion, in light of t he thought of bagging, we bag the negat ive samp les and then chose the best SVM class ificatio n model to train the sets. Fist, we use the proposed method image retr ieva l based on interest region in t his thesis for the init ia l query. The n we use bagging relevance feedback to retrieva l unt il the user is satis fied wit h the outputs. The experimenta l results ind icate that, wit h the introduct ion of bagging SVM releva nce feedback proposed of this thes is, the C BIR system can obtains a higher average precis io n in eva luat ion indexes tha n that w it h the traditional SVM relevance feedback.