Research On Relevance Feedback Of Content-Based Image Retrieval

In the Content-based image retrieval, it's hard to accurately describe the users' retrieval needs for the system by physical characteristics of the underlying image. That means image features extracted by system from the sample that the user submitting usually can't reflect the user's retrieval purpose very well. Therefore, it's necessary to introduce the relevant feedback mechanisms to continuously improve the satisfaction of search results. Against this background, in the mid-90s,20th century, relevance feedback (RF). which first proposed in text retrieval field, was introduced into the field of content-based image retrieval.The typical method of relevance feedback includes Query point movement method and weight adjustment method. The former method is by modifying the query vector. Make the query vector move toward the center of the image-related, And the latter is by using of the feedback message. modify the formula from the weight of each component, reduce the weight of the query vector that is not important and highlight the more important component. With the RF technology' development and maturation, since the advantages of (1) Small sample; (2) Without prior knowledge; (3) High-dimensional problems may be easily resolved; (4) Nonlinear feature. SVM technology began to be applied in the field of content based image retrieval by the researchers.The thesis make feedback to the content-based image retrieval results using SVM method, then compare with the weight adjustment based method, which is a classic feedback method. First, the user uses color, texture or shape retrieval algorithm to retrieve, and allowed the user to mark the initial retrieval results. Use the user's marking message as the training samples of the SVM and developed to the classifier. Finally, predict the remaining images by this classifier and return the predict results. The paper use recall and precision as evaluation indicators of experimental results, SVM algorithm'advantage in the precision rate is not obvious, but the recall rate can be significantly improved than the weight adjustment based algorithm.Since SVM does not require prior knowledge, it is widely used in various fields, how to choose the appropriate kernel function and parameters, and how to improve existing kernel function will be the direction of future research.