Ontology-based Image Retrieval

With the rapid development of computer, multimedia and Internet techniques, amount of images has been produced .Therefore, it becomes an urgent problem that how to find needed image efficiently in large-scale image database. One effective ways has been proposed to solve the problem: image retrieval based on Ontology. Our research work focus on the following aspect:(1)Image retrieval model based on ontology is proposed in this paper. The ontology is described by a directed acyclic graph (DAG). Here, each node in the DAG represents a concept .each concept in the ontology contains a label name, low feature vector and semantic feature vector. The line between the node is described the inter-relationship. The retrieval model contains low visual feature and high semantic concept, so it can reduce the semantic gap between the low feature and the semantic visual. In order to extract image region accurate, annotate image region and image and measure the similarity between the images, we extract the color feature based on color histogram, Gabor wavelet is used to extract texture feature and Zernike moments is used to extract shape feature.(2)An improved genetic FCM method is proposed to segment image, and the region of interest is extract. The proposed method overcomes the bug of FCM which converged to local optimum, and it resolve the problem of FCM which sensitivity to the initialization of cluster center. And then the validity of clustering is discussed and the optimum function of validity is proposed and at last the improved image segmentation algorithm is proposed.(3) A method of ontology construction for sport image is proposed. The color, texture and shape feature is used as low feature, and K-means method based on semi-supervised and constrained clustering is used to cluster the image region, and then the statistics method is used to construct every cluster and semantic concept. And then the semantic vector is used as high feature, the improved K-means method is also used to cluster the image, in this way, we gain the ontology of sport field.(4)A relevance feedback method based on improved adaboost integrate SVM is proposed in the paper. The adaboost method adopts a new method to acquire parameters. The weighted parameters of weak classifiers are determined not only by the error rates, but also by their abilities to recognize the positive samples. We used the sign image to train the support vector machines (SVM), and attained the classification function and then feed back the retrieval image to users according to the classification function value.