Research On Image Annotation Based On Scene Semantic

With the development of computer technology and the popularity of network applications, it has been an important topic in the field of the information retrieval to effectively manage and retrieval images. Owing to the unresolved problem of the semantic gap between the low-level visual features and the high-level semantic concepts for the content based image retrieval, the semantic-based image retrieval is gradually becoming a hot research area. The key point of the semantic-based image retrieval is the semantic-based image annotation. This thesis firstly reviews the developing processes of the image annotation, the corresponding theories and their advantages and disadvantages. The main contributions of this paper are as follows:Proposed an semantic-based image annotation algorithm based on combine Probabilistic Latent Semantic Analysis and Gaussian Mixture Model. This algorithm included the following process: the low-level visual features extraction, use Probabilistic Latent Semantic Analysis for training images latent semantics extraction, the Gaussian Mixtures Model representation of images. Firstly, each image is seen as a document, each keywords for used annotation image is seen as a word in the document, use the"document-words"matrix be construction through the above way as input values for probabilistic latent semantic analysis model, use the expected maximum algorithm for calculation model parameters, use this parameters as a basis for image classify and the latent semantics extraction. Secondly, Each image is divided into several homogeneous regions by use image segmentation, each regions as a point in the feature space, extracted feature vectors of each region, clustering by use expected maximum algorithm, get an Gaussian Mixture Model for each image category. Finally , in each model, calculation the posterior probability for each test images, and as a basis for image semantic annotation. The semantic-based image annotation method we proposed in this paper, does not require prior knowledge of classified information in training images. Experiment conducted on the relevant images dataset demonstrates the effectiveness of the method.