Web-based Collaborative Image Retrieval And Annotation

Image retrieval has been always hot in the academic and industrial area, with the rapid development of information technology and big data technology, people’s activity in the Internet is becoming more and more focused on image, video and other multimedia, image retrieval on the Internet is playing a more and more important role,and it is widely used in e-commerce, Internet social and other fields, how to find a picture that you want in a large scale image datasets become hot area in the field of image retrieval research.This thesis will combine two low-level image features, automatic annotation algorithms, and user’s relevant feedback, in order to solve the two problems of image retrieval: semantic gap problem and the curse of dimensionality, at last complete the study on Collaborative image annotation and retrieval on Web, the main work are as follows:Firstly, image features plays an important role in image retrieval, the outstanding feature expression will have a huge impact on the performance of image retrieval. Based on the summarization of the visual features of the image, This thesis proposes a new feature Scatter-SIFT that combine scattering coefficient features and scale invariant feature transform(SIFT) to complete the image low-level feature matching, it can improve image classification effectively.Secondly, image annotation manually is not realistic in large scale image dataset. Through study on Fast Tag algorithm and Negative Bootstrap algorithm,This thesis proposes FT-NB algorithm that combine these two algorithms to tag images. The new algorithm take the advantages of Fast Tag and Negative Bootstrap, and the experiments are implemented to verify the effectiveness of FT-NB algorithm.Generally, image retrieval system does not make full use of the users’ relevant feedback, while the web is extensive naturally, it can connect anyone in anywhere. Therefore, this thesis proposes the method that combine relevant feedback logs and FT-NB algorithm, after the image annotation results are iteratively updated, the result of image retrieval will be further improved.Finally, after the comprehensive understanding of the above excellent algorithm and relevant feedback, this thesis implements Just Pic1.0 which is a collaborative image retrieval prototype system on Web, It can collect user’s relevant feedback conveniently. Besides, FT-NB method, relevant feedback’s learning, Fast Tag, Negative Bootstrap and BOW methods are all tested on public data sets such as corel5 k, Caltech101 and ESP, through the experiment results, the effectiveness of relevant feedback is verified.