Research On Image Feeback Technology Based On Log-database And Multi-classification SVM

Content-based image retrieval has been drawing more and more research attention in recent years. With the development of the related technology, the content-based image retrieval system appearing one after the other. However, as a result of the localization of current image understanding technology and human's cognitive subjectivity to visual content, there have a great gulf fixed between the low level feature and the actual high level semantic. For example, a blue sky picture and a sea picture, the extracted feature from the pictures by current computer vision technology is some low level feature such as color feature, texture feature and shape feature. The big difference between low level and high level semantic is one of the biggest obstructions to the CBIR.To resolve the problem in low level feature based image retrieval, improve the precision of the retrieval, provide more convenient retrieval mode, current research about CBIR has focus on interactive retrieval mechanism which takes human as a part of the retrieval process, such as interactive image region segmentation, interactive database mark, using learning mechanism in retrieval process and interactive combine high level semantic feature to improve the performance of the retrieval system. Introduce the interactive conception to the retrieval region just be the interactive feedback method of CBIR.My thesis was based on the IVDB system in my laboratory. Firstly the thesis introduce the key technology using in image retrieval system, which include image feature extract , similarity measure, relevance feedback and so on. All of this is the absolutely necessarily technology to realize the retrieval system. After that, we do some research on the support vector machine theory that will be used in my system. This is the part we must take more attention to if we want to obtain some improvement, and the thesis use rather big space to do research on this theory. Then discuss the realization of the relevance feedback schemes based on SVM in theory and in practice. After analyzing a great deal of books and articles about SVM and SVM based relevance feedback, we summarize the problems existed in current SVM base image retrieval system.Firstly, although the SVM is based on small-scale training set, but the classify is unstable on a small-scale training set,and over fitting happens because the number of the feature dimensions is much higher than the size of the size of the training set. It means that there still have some demands about the size of the training set. But in the process of retrieval, as the impatience of the user, the samples marked by user always are small, and this is a central problem puzzles researchers in image feedback research.Secondly, currently research which using SVM in image feedback schemes is dichotomy. user mark the positive image training set and negative image training set, then the dichotomy separate the images into two classes. But there are a large number of images in the image database, if the images are only separated into two classes; many imaged may be separated into wrong class, and affect the precision of retrieval. And the current arithmetic doesn't have the condition to use multi-classification SVM. The main reason is that the image marked by user especially the negative image set is too small.Besides, the number of the positive and negative images usually is different. Sometimes, this difference is large, and the SVM's optimal hyperplane may be biased when the positive feedback samples are much less than the negative feedback samples.For the problems above, the thesis find corresponding method to resolve it. We will do detailed presentation as below.Firstly,To resolve the problem that the SVM's feedback image samples is small, we use the log database that can noted the feedback information down, and make the best of the log information. Based on the cooperate-filtration mechanism, we find a way to make best of the log information when the log information in database is not too enough.Secondly, We analyze the cause why the currently SVM based image feedback schemes can use dichotomy only, we also point out all kinds of deficiency that the use of dichotomy in image feedback. After we designed the log database and enrich its information, we analyze the condition to use multi-classification SVM in the image retrieval system and the advantage of to use multi-classification SVM, construct a multi-classification arithmetic that suitable to the image feedback.Finally,we do some research on the feedback process of users, based on the psychology of user in the process of feedback, changing the illogicality condition of feedback, adjusting the feedback image weights dynamically, the thesis analyzes the effect of the imbalance between the positive feedback samples and the negative feedback samples, and resolves the problem in a smart method.We do experiments on a standard image database to test the effect of the size of the training set, to test the impact of our multi-classification SVM and the performance of the adjusting of punish parameter. The result validated that the proposed log database based and multi-classification SVM based image feedback method are much better than the traditional SVM based image feedback method.