Study On The Techniques Of Image Retrieval Based On Global Feature And Feature Selection

Content Based Image Retrieval is a technology to describe image content through visual information, such as color, texture, shape, and index images by inputting query example. Content Based Image Retrieval is an efficient way to find and manage huge multimedia information of large database system, Content Based Image Retrieval has gradually become a research hotspot in the field of computer vision and image processing as a popular emerging technology, and there is still a vast room for development.On the basis of reading plenty of literature domestic and overseas, and summing up the works of predecessors, this paper introduces the research situation and development status of Content Based Image Retrieval, and elaborates key techniques applied to Content Based Image Retrieval. And on this basis we present two new methods for image representation for the issue of image expressing, and two new methods for feature selection for the question of semantic gap. Work and dedication of this paper is mainly on the following aspects:(1) The process of Content Based Image Retrieval has been analyzed, and key techniques applied in Content Based Image Retrieval are summarized in this paper. In the first place, characteristics of low-level visual features are analyzed, and several classical methods for feature extraction are introduced. Then, some kinds of typical methods for similarity measurement are introduced consequently. After that, the development status and application of feature selection technique is summarized.(2) As for the issue of image content representation, this paper presents two new feature extraction methods. For the one method, we compute the direction variation and intensity of pixel values in an image block first, which is divided into different texture patterns. Two image blocks in neighborhood make up a pattern pair, we obtain the texture direction feature by counting the number of the pattern pairs. For the other method, color and texture features of an image block are extracted, and image is divided into multiple regions utilizing K-means clustering algorithm and the extracted features. Then, calculate the mean value of color and spatial distribution of pixels in each region. The image is represented by the multi-features of all regions finally.(3) To narrow the "gap" between image low-level features and human high-level semantics, this paper presents two new feature selection methods. We test the discriminating ability of every feature in the feature set with the discrimination criterion, and then we can find out the feature subset, which expresses image content in an even better fashion, feature subset with the highest precision rate is considered to be the optimal feature subset.For the feature representation and feature selection methods mentioned above, this paper conducts a large number of experiments on different testing datasets, and contrasts with image retrieval algorithms in existence. The results of the computer simulation experiments show that the proposed texture direction feature and region partition feature can represent image content effectively, and the feature selection methods can narrow the semantic gap, and improve the retrieval precision visibly.