Image Annotation Using Maximum Probability Method And Nearest Neighbor Criteria

Recently, the image annotation has become a hot topic of computer vision. The main idea is to learn the latent relationship between the semantic and visual spaces from annotated images, and to predict annotation for test images. The development of the machine learning brings many learning models to the image annotation. The current algorithms can be classified into three categories:the classfier based, the statistic based, and the nearest neighbor based methods. The maximum of probability is based on statistical model, and the nearest-neighbor criterion is a simple yet effective method. They are two key points of this study.This research is dedicated to the reduction of the semantic gap and improvement of automatic image annotation. The main content includes:1) A topic model incorporated with saliency is proposed. The graphic model is presented with the analysis from the statistical and biological point of views. The saliency of an image region is determined by the weighting function of salient features. The high value of the weighting function indicates high level of saliency, and vice versa. Compared with the traditional ones, the new graphic model incorporates the function dependency, results in a more flexible generative model, which can better reveal the relationship among random variables. In addition, based on the theoretical analysis, experiments confirmed the feasibility of the model. The experiments showed that, for the annotation task, the proposed method performs better than the traditional CorrLDA (Correspondence Latent Dirichlet Allocation) model.2) A topic model incorporated with geometric context is proposed. Traditional topic model does not consider the influence of features’ geometric properties to the topics. We assume that the features have geometric properties. The distinguishable geometric properties help not only the sampling of the topics, but also a better context for annotation prediction. Compared with traditional approach, the proposed method includes a node which indicates the properties. The node is suitable for representing more features, and controlling the existence of the properties.3) An image annotation model based on pyramid histogram of edge contrast is proposed. The image representation is very compact, and has a low feature dimension. It is insensitive to several kinds of image degradations. It has a better classification or annotation performance than some state-of-the-arts methods. The pyramid histogram of contrast can encode the image so that both local and global information are incorparated.4) The image annotation is studied using the color-space and the differential operator. The influence of color-space, differential operator, and dimension reduction technique (e.g. PCA) on image annotation task is investigated. Experiments show that most color-spaces are better than RGB color-space, and differential operators can improve the performance of annotation. For nearest-neighbor based methods, the high resolution may not always lead to better annotation performance. The PCA operation may not suitable for image annotation.5) An image annotation algorithm using feature fusing and semantic similarity is proposed. The visual-semantic relationship is studied through feature-distance and semantic similarity functions. To find nearest neighbor image, the influence of the correlation and scale of feature distance are eliminated. The weak relation between semantic and visual similarity is revealed by the conditional feature-distance distribution on semantic similarity. Experiments show that the method has a wide range of adaptability. It has potential applications due to the good annotation performance.