With the rapid development of computer network and imaging technology, the number of image is growing quickly. However, the increasing number of images brings some trouble to people:they cannot find what they really need from huge amount of available data. How to organize and manage the image data effectively has emerged as a hot topic in the multimedia information field. Automatic image annotation technology is being attached more attention in recent years due to its potentially application on image understanding and image retrieval. This paper mainly focuses on how to establish an effective learning model to solve the annotation problem, as well as how to supplement semantic image descriptions by associating color words with each existing tag.By analyzing the fact that the annotation issue exist ambiguity both in the input space and out put space, this paper proposes a novel supervised multi-instance learning method to solve the issue of automatic image annotation. This method regards each image as a bag, which contains a number of instances. For each keyword, first, we calculate the semantic similarity between the keyword and the instances of positive bags through a reinforced diverse density algorithm. Then, we select instance prototypes of each keyword by assigning a specific threshold. Last, the selected instance prototypes are modeled using the Gaussian mixture model (GMM) in order to build the semantic mapping relation between image semantics and the corresponding visual features.Besides, in view of the state that the content of current image annotation is noun form, this paper proposes a new annotation style named keyword with color description. First, we build a color feature information store through PLSA-bg model. Then, this paper uses an effective strategy to estimate the color name of a particular region. Finally, we combine color words with keyword to form the image annotation with color describing words.The proposed method demonstrates a promising performance based on Corel database and Microsoft Research Cambridge database compared with current state-of-the-art algorithms. |