Research On Image Semantic Annotation Based On Sequential Prediction Learning

In the world of digital life, multimedia data is growing. Multimedia data play a crucial role in people’s life in the era of big data. In multimedia data, the image and video account for the most. Image is the basis of the videos. Naturally, it plays an important role in the management of multimedia data. Through fast and accurate image retrieval, people can work quickly and efficiently on network activities, largely improve people’s quality of life. Automatic image annotation is an important and challenging role in the content-based image retrieval. To some extent, it can solve the problem of semantic gap. If automatic image annotation can be realized, the existing image retrieval problem can be converted into quite mature text retrieval problem. Therefore, the automatic image annotation is a very meaningful research.(1)The underlying image feature extraction is the foundation of image annotation. Paper states the image feature extraction method, by experiments on different image features, compares and analyzes the characteristics of all kinds of image features, sums up and summarizes the advantages and disadvantages of the methods and scope of application.(2)In order to solve the problem of semantic gap between underlying image features and high-level semantic, based on the figure labels transmission principle and basis, we research a semi-supervised learning method based on Sequential prediction, to realize the effective transmission of labels. The method is effective in image classification. Combining with the characteristics of multi-labels, we research the method of image annotation based on Sequential prediction semi-supervised learning, analyze the time complexity and space complexity of the method. The experiments show that image annotation accuracy can be improved evidently through this method.(3) Inevitably, there is noise in image annotation, the phenomenon of marking words meaningless and the inconsistency between annotation words, In order to solve these problems, this paper makes use of a symbiotic relationship model to optimize image annotation, uses the symbiotic relationship model to mine the semantic relationships, and to optimize the initial semantic. The experiments show that image annotation accuracy has been improved evidently.This thesis extracts the features that can express accurately the image visual information on the basis of extracting image semantic feature. On this basis, it realizes image labeling with the help of algorithm. This thesis improves the performance of image annotation from three aspects:extraction of image feature、 building valid annotation model, refining the obtained image labels. And at last, this thesis achieves a good result.