Research On Image Automatic Semantic Annotation

With the development of computer technology and network communication technology,different kinds of images have been emerged everyday. How to retrieve the required imageinformation from such a large number of image information has become an importantresearch topic in the field of information retrieval. Content-based image retrieval technologybrought to the attention of the researcher, since it has the advantage of effective analysis,organization and retrieval of images. Visual features and semantic features are two commonused important characteristics which are content-based image retrieval. Visual features arethe underlying characteristics, reflects the people’s intuitive feel of the image content;semantic features are high-level characteristics, reflect people’s subjective understanding ofthe image content. However there is a significant difference between them, which isso-called "semantic gap". At present, how to overcome the "semantic gap", it is still aproblem to be solved. So, In-depth study of the intrinsic link between the visual features andsemantic features, and keep people’s visual experience with subjective understanding, whichhas become an important research topic in content-based image retrieval technology.Semantic-based image automatic annotation is one of the key technologies to solve the"semantic gap" in image content-based retrieval.This thesis focused on the describe problem between the concepts of visual featuresand image semantic concepts mappings in the semantic annotation, which based on theresearch of the image automatic semantic annotation related technology, discussedparticularly a gaussian mixture model and its application in image automatic semanticannotation, proposed an image annotation method with GMM based on semantic concept,confirmed the effectiveness of the method by experiment. The main research included thefollowing sections in this thesis:(1)Proposed an improved image annotation method with GMM based on semanticconcept. The main idea was that merged semantic concept class with Ncut imagesegmentation, and established GMM for semantic concept classes based on color andtexture features, obtained the semantic concent by using the EM algorithm, eventually,syncretized probability calculated semantic concept in two GMM, taked the excellent toannotate image area block, until completed the unknown image annotation.(2)Proposed an EM algorithm initialization method with PSO-Kmeans algorithm. Itmaked PSO-Kmeans algorithm applied to the initialization of the EM algorithm in GMM image annotation training with semantic concept, searched the optimal solution of semanticconcept classes training data set as the initial parameters of the EM algorithm. It simplifiedGMM with complex parameter estimation problem to iteration for EM algorithm withexpectations E-step and maximize the M-step by maximum likelihood estimate to estimatethe optimal parameters of GMM image annotation.(3)Experimented research by the Core15K image data set.The experimental resultsshowed that image annotation method with GMM based on semantic concepts canaccurately predict the number of text keywords for the image to be annotated, and improvedthe image retrieval precision and recall rate. The initialization of the EM algorithm withPSO-Kmeans algorithm can effectively improved the speed of image annotation.