With the development of Internet and the computer hardware, more and more images are generated, stored and transmitted. There is a great challenge to the traditional image retrieval and management system. The content-based image retrieval system cannot meet the user's demand due to the low-level features used. To retrieval the image effectively, the high-level information should be embedded into the image database. This thesis is mainly focusing on the image retrieval and semantic classification of the image database, and the main contributions are as follows: (1) Based on a brief review of the existing retrieval systems and image semantic classification systems, the low-level features used in the image retrieval system are analyzed. Their advantages and disadvantages of each feature and their application situation are compared. That wilt help users to realize the image feature selection and extraction of new feature. (2) We developed an Support Vector Machine(SVM)-based hierarchical image semantic classification system. A new image feature based on the objects of the image is presented according to the characteristic of the indoor and outdoor images. In addition, the new feature called parallel edge line feature is also suggested for the classification of building and landscape image. The standard two-class SVM, one-class SVM and multi-class SVM are used in different stage of the classification system and achieve a good experimented result. Experiment results show the feasibility and validity of the features and the classifiers. (3) We presented several approaches to improve the system accuracy, which include the dimension reduction of the image features and the local semantic features, rejection option based on the probabilistic output, incremental learning based on the SVM and the combination of different classifiers. Several experiment results confirm the effectiveness of the improved system. (4) In this thesis, image retrieval systems are presented based on the high-level semantic index and similarity learning of the user's feedback. Several machine learning methods are used in predicting the user's similarity habit and the experimental result show the applicability of the retrieval system. (5) An image filter system based on the Google image search engine is presented to overcome the low accuracy of the Internet image retrieval. This system is realized through combining the semantic extraction with the retrieval system, which is a combination of the keyword based and visual feature based image retrieval system. The experiment results show that the accuracy can be significantly improved through introducing the filter system. This model of image filter system provides a good solution to apply the existing searching engine but with improved retrieval accuracy.
修改评论