Research Of Key Technology Of Content-based Image Search Engine

Content-based image search engine is an important and challenging field of academic research. Developing practical content-based image search engine and finding out relationships between images have the important practical significance in the field of research of key technology of content-based image search engine. This thesis aims to explore new ways of image database index technique of content-based image search engine after regional weighted entropy is applied to image feature extraction. With careful study and comparison of machine learning of several business search engines, this thesis seeks to explore the machine learning methods of content-based image search engine, and develop the corresponding software. The main contributions are included as follows:A new description method of image entropy named regional weighted entropy is proposed, which combines the concept of image entropy and image segmentation algorithm after analyzing the current color-space image feature extraction algorithms. Some properties of regional weighted entropy are proved. The distribution change of image entropy, which is caused by weight's change, is described by using entropy performance evaluation index from the point of view of probability, considering the interested regions and weights precision applied by users, then the reasonable weight is determined. Experimental results show that the accuracy of image content described by regional weighted entropy method is 50% higher than that of traditional entropy methods.Multidimensional indexing concept is applied to content-based image search engine. Index structure of text-based search engine could not be used due to features of content-based image search engine. R~* tree index is improved adaptively and applied to content-based image search engine in this thesis. Multiple eigenvalues of images are normalized in the preprocessing of image multi-feature, which is convenient to build trees and query. Concept of similarity distance is defined in multi-featured image matching by R~* tree circle-field query, then leaf nodes which contain similar images is found. Experiment results show that, searching time used by R~* tree reduced greatly, and timing performance is superior to that of simple indexing structure.On the basis of analysis for machine learning of current business search engines, this thesis combines features of content-based image search engine, designs and implements machine learning function of three aspects of content-based image search engine. Efficiency and accuracy of image search engine which adds machine learning function improved greatly.On the basis of above research results, this thesis designs and implements V2.0 system of content-based web image search engine. The system adopts Gordian technique of several content-based image search engine, such as regional weighted entropy which is applied to image feature extraction, R* tree multidimensional indexing structure, and so on. The system V2.0 of content-based image search engine is developed, and system's accuracy and user's response speed meet expected object.