Research On Image Annotation Based On Optimal Tag Sets

With the rapid rise and development of computer technology, network communication technology and multimedia, image database become more and more enormous. How to gain information that people really want efficiently becomes an important subject. The substance of automatic image annotation is to obtain high-level semantic keywords through mining and processing the underlying information features of image which is convenient to access image data efficiently. In this paper, we research automatic image annotation. It’s main contribution can be concluded in the following aspects:(1) In order to gain the region feature of image, an improved Normalized Cuts algorithm is used to segment images. Firstly, Mean shift algorithm is adopted to pre-segment image and form a region that is represented by a weighted region adjacency graph; then Normalized Cuts is adopted to fuse those small regions, it can obtain the final result of image segmentation.(2) In this paper, maximum density clustering algorithm (MDCA) is adopted to cluster region features and SIFT feathures of image, then discrete visual words is gained. Discrete visual words are expressed as represent image by using a "Bag of word" model.(3) In this paper, each image can be represented by a collection of visual words and text keywords respectively. Aiming to the textual modality and the visual modality,we adopt an modified automatic and semantic annotation model to capture latent semantic topics respectively from visual and textual modalities. Furthermore,an adaptive asymmetric learning approach is proposed to fuse these semantic topics.And in this way,the model share the same latent semantic topics.(4) In order to increased the accuracy of automatic image annotation, we introduce a optimization algorithm to automatic image annotation system. After annotation, in order to improve the accuracy,the correlation between words based on term frequency factor that improve metric accuracy of the correlation between the words and heuristic iterative algorithm are adopted to optimize label set efficiently which is obtained from the model.Experimental results show that the proposed method effectively improve the effect of automatic image annotation. Figure32, Table6, Reference66.