Ontology-based Image Retrieval Technology Research

In recent years, with the rapid development of computer performance and multimedia information, image information is used widely in life. People's demands on the images and other multimedia data are increasingly intensive. Traditional search techniques completely cannot satisfy the people's need, and it is necessary to enhance the efficiency of image retrieval system to manage and retrieve image information from the large database of image. We focus on ontology-based image retrieval and the main contributions are following:(1) We propose a method of semi-automatic ontology construction. First, we use the ontology construction tools to artificially build ontology basic architecture. Then we use the parallel fuzzy inference mechanism based on ontology learning approach to expand and improve the concept of ontology. This method can effectively solve the ontology created uncertainty and incomplete.(2) We extract the image blobs using the method combines the CRF and BOF. We regard image segmentation as process of optimal annotation of pixel based on MAP-CRF framework. Then the problem of optimal annotation is considered as energy minimization problem by Gibbs distribution function. The Graph Cuts is adopted to solve energy minimization problems. Considering the complex boundaries of images, this potential function of the CRF model is improved. Namely Jensen-Shannon dispersion is used as a condition of the edge of the random energy term in that this method. Taking into account the complexity of computation, we use pre-watershed algorithm for image segmentation. We extract the image blobs from the image segmented regions using BOF.(3) A new image annotation framework using ontology is presented in this paper. We use two steps to automatically annotate semantic for the image. First of all, a probability table of the blobs and ontology concept is constructed using statistics learning, and then the Bayes is employed to annotate the posterior probability between annotation image, and ontology concept. In the second step, combining the semantic relationship between concepts of Domain Ontology, we annotate image to gain the higher semantic concept for annotation of image semantics.(4) We propose a space-based pyramid coding method for the query. That PCA points feature of low dimension extracted. Feature space is divided into multi-resolution by the pyramid matching idea and every feature is encoded according to certain standards. Images are calculated the similarity of coding sequences of feature points. We annotate the image of the candidate to obtain semantic vector and then calculate the similarity of feature code, semantic features and ontology concepts between the images to obtain the best most relevant images.