-based Multi-label Learning Image Semantics Automatically Marked

With the rapid development and widespread of multimedia digital techniques, the reduction in storage cost and the transmission bandwidth growth of the network, multimedia information such as image and video become ever more available, and play big role in people's life and social development. "Explicit semantics" is an important prerequisite for large scale multimedia information management. So automatically obtaining the semantics of the multimedia data by using information techniques has important meaning in theory and practice, and attracts great attentions in academic and industrial fields.Image is the basis of the video. It occupies an important position in multimedia data management. So automatic image annotation (AIA) is the hot research issue in the related fields. The nature of AIA is a process of "learning", that is, associating images with semantic keywords according to their visual contents. Thus, machine learning methods and statistical inference techniques have both been applied to solving the problem of AIA, which is continuously deepening and promoting. However, due to the problem of "semantic gap" and multi-labeling, the annotation performance of existing methods is not satisfactory, and needs to be further improved.This paper studies multi-label characteristic of AIA. By concentrating on the correlation between multiple keywords, we addresses the problems of data imbalance and overlapping brought about by multi-labeling and Web image annotation problem. Based on this, several image annotation methods with good performance are proposed, which are mainly about generative model based multiple class label propagation, the combined generative model and discriminative techniques and noisy training set based web image annotation.The main work of this paper is as follows:1. A new image annotation method via extended generative model is proposed: inorder to exploit the correlation between keywords, we propose a new image annotation method by extending the tradition generative model to estimating the probability of a set of keywords being the caption of an image, and present a heuristic iterative algorithm to solve the problem. In this method, we propose a topic-image-region multi-granular hierarchical feature estimation model, and analyze the correlation between keywords. Both of their contributions to image annotation are extensively exerted according to our heuristic iterative algorithm. The experimental results on a real world benchmark show that our method outperforms the traditional generative model based annotation method.2. The discriminative hyperplane tree based image annotation method is proposed: this method leverages the benefits of the generative and discriminative models by building the local latent topic hierarchy and the corresponding hyperplane tree based on the high generative probability neighborhood of the unlabel image. The experimental results on a real world benchmark show that our method outperforms the state-of-the-art generative model based annotation method and discriminative model based method.3. The local multi-label classification based image annotation method is proposed:this method provides another solution to combine generative model and discriminative techniques to improve AIA. We further explore the underlying semantics of visual similarities, and try to find the optimal margin in both visual and semantic spaces when generating the latent topic to obtain large separation in both spaces. The experimental results on a real world benchmark show that our method outperforms the state-of-the-art generative model based annotation method and discriminative model based method.3. A new Web image annotation method based on noisy training set is proposed:we present a novel web annotation framework. We introduce a "light weight" method to obtain the training set automatically. Then, we propose a novel annotation method based on mixture component based local fisher discriminant analysis to deal with the bad influence of the noisy training data. The experimental results on a real world Web image data set show that our method outperforms the traditional annotation approaches with noisy training data.